This action might not be possible to undo. Are you sure you want to continue?
VLSI INTERCONNECTS
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF ELECTRICAL ENGINEERING
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
TingYen Chiang
June 2004
ii
©
Copyright by TingYen Chiang 2004
All Rights Reserved
iii
I certify that I have read this dissertation and that, in my opinion, it is fully adequate
in scope and quality as a dissertation for the degree of Doctor of Philosophy
___________________________________________
Professor Krishna C. Saraswat, Principal Adviser
I certify that I have read this dissertation and that, in my opinion, it is fully adequate
in scope and quality as a dissertation for the degree of Doctor of Philosophy
__________________________________________
Professor Robert W. Dutton
I certify that I have read this dissertation and that, in my opinion, it is fully adequate
in scope and quality as a dissertation for the degree of Doctor of Philosophy
_________________________________________
Dr. James P. McVittie
Approved for the University Committee on Graduate Studies:
_________________________________________
iv
v
Abstract
The scaling of VLSI structures leads to continuous increase in current density that
results in ever greater interconnect Joule heating. In addition, a variety of lowk materi
als have been introduced to reduce the RC delay, dynamic power consumption and
crosstalk noise in advanced technology. Together with the poor thermal conductivity of
such materials and more metal levels added, the increasing thermal impedance further
exacerbates temperature rise in interconnects. As a result, not only will thermal effects be
a major reliability concern, but also the increase of resistivity with temperature can
degrade the expected speed performance. On the other hand, overly pessimistic estima
tion of the interconnect temperature will lead to overly conservative approach. Hence,
performing a more realistic thermal modeling and analysis of interconnects is critical.
This research proposes both compact analytical models and fast SPICE based 3D
electrothermal simulation methodology to characterize thermal effects due to Joule
heating in high performance Cu/lowk interconnects under steadystate and transient
stress conditions. The results demonstrate excellent agreement with experimental data
and those using Finite Element (FE) thermal simulations (ANSYS). The effect of vias, as
efficient heat transfer paths to alleviate the temperature rise in the metal wires, is included
in our analysis for the first time to provide more accurate and realistic thermal diagnosis.
It shows that the effectiveness of vias in reducing the temperature rise in interconnects is
highly dependent on the via separation and the dielectric materials used. The simulation
vi
methodology has also been applied to quantify the use of dummy thermal vias as addi
tional heat sinking paths and possible solution to hot wires.
The impact of Joule heating on the scaling trends of advanced VLSI interconnects
has been evaluated in detail. It shows the interconnect Joule heating can strongly affect
the maximum operating temperature of the global wires which will, in turn, constrain the
scaling of current density to mitigate electromigration and, thus, greatly degrade the
expected speed improvement from the use of lowk dielectrics. Coupled analysis of
reliability and delay, under the influence of thermal effects, is performed to optimize
interconnect structures such as wire aspect ratio and ILD thickness. Finally, potential
bottlenecks and opportunities of future heterogeneous three dimensional (3D) ICs with
various integration scenarios are identified from the thermal point of view. It is shown
that under certain scenarios, 3D ICs can actually lead to better thermal performance than
planar (2D) ICs. Tradeoffs among power, performance, chip real estate and thermal
impact for 3D ICs are evaluated.
vii
Acknowledgements
I am deeply indebted to many people who made my Stanford years an enlighten
ing, rewarding and memorable experience. It is a great pleasure to acknowledge the
people who have supported me throughout these years.
First and foremost, I would like to thank my adviser Professor Krishna C.
Saraswat for his guidance, insight and patience. I am forever grateful his unwavering
confidence in me throughout my time at Stanford, both in my research and in my startup
dream. I feel very fortune to have worked with Professor Saraswat, who provides a
creative academic atmosphere for the students and always shows his unconditional
support to the students in all aspects. Being one of the most friendly, downtoearth
professors, Professor Saraswat is not only an admirable educator, but also a lifetime
mentor. The experience I have gained under his guidance is invaluable and has left an
indelible mark on my self and persona.
I would like to thank members of my Oral and Reading Committees, Professor
Robert W. Dutton, Professor Piero A. Pianetta, and Dr. Jim McVittie for their enthusias
tic discussion and valuable feedback. I also thank Professor Michael A. Kelly for his
continuous help and encouragement. My gratitude also goes out to Professor Kenneth E.
Goodson for his time and effort acquainting me with thermal analysis and, to Dr. Pawan
Kapur for numerous entertaining discussions and invaluable brain storming sessions, and
to Dr. Shukri Souri, whose sense of outofbox thinking and professionalism I thoroughly
enjoyed, and who made possible the thermal analysis of 3D ICs through his 3D ICs
performance study. I also thank Dr. Kaustav Banerjee and Sungjun Im whose generosity
viii
of discussions and challenges elevated my intellectual awareness of my own work. In
addition, I would like to thank MARCO Interconnect Focus Center for funding my Ph.D.
work.
My colleagues from various research groups have been terrific and my work has
benefited tremendously from interactions with the past and present members. I am very
thankful for the assistance and friendship of Dr. Shahram Alibeik, Dr. David Bang, Dr.
Richard Chang, Ray Chen, TsuJu Chin, Hoon Cho, Hoyeol Cho, ChiOn Chui, Dr. Mar
Hershenson, YuLi Hsueh, Dr. Nabeel Ibrahim, Dr. Amol Joshi, Dr. Mayur Joshi, Dr,
Steve Jurichich, Dr. Pranav Kalavade, Dr. Kambiz Kaviani, Rohan D. Kekatpure,
Donghyun Kim, Tejas Krishnamohan, Dr. Marci Liao, Dr. Alvin Loke, YiChang Lu,
Ammar Nayfeh, Ali Kemal Okyay, Abhijit Jayant Pethe, Dr. Xiao Ning Qi, Pinkesh
Sachdev, Rohit Shenoy, Dr. Ben Shieh and Junfeng Xu.
Although it was short, the wonderful experience of writing business plan, present
ing our ideas in front of venture capitalists and winning the $10,000 prize in the Stanford
Entrepreneur’s challenge business plan competition was thrilling. I would like to thank
Ashish Chopra and Alper Tekin for all the dreams of riches and early retirement and for
our TAAV team which aimed to be the leader in the physical analysis segment of elec
tronic design automation with its breakthrough full chip thermal simulation technology.
I thank Irene Sweeney for great administrative support. I would also like to thank
Dr. Richard Dasher for coordinating and coming along for a SPIE trip to Philips, Nether
lands. I especially thank Professor Tom Lee, Carmen Miraflor and Maureen Rochford
for working hard to get me back to U.S. when my green card got stolen in Amsterdam,
Netherlands, during the SPIE trip.
ix
Without question, none of this would have been possible without the support,
moral and otherwise, of family and loved ones. To Aunt Chuling and her family, for all
your support in a time of need, my deepest appreciation. To my brother and sister, Ting
Chun and SherFeng, I am forever grateful. For your friendship and taking care of our
parents when I am far away from home. To my grandparents, I thank you for your
endless love and encouragements. Most of all, my eternal gratitude goes to my mother
and father, MinChuan and HoChen, whose constant inspiration and unreserved support
have made it all possible. This dissertation, its accomplishment through all the obstacles,
is a manifestation of their dedication, sacrifices, commitment and hard work towards each
one of us, their children, for all their years. To them, I am eternally indebted and dedicate
this work.
Finally, for my beloved wife, ChingJung, the greatest blessing in my heart, for
being always there for me. Without all her unparalleled love and support through the
years, I would not go back to pursue Ph.D. degree and this thesis would never have been
finished. Thanks to my precious son, Eric, for always greeting me with a smiling, loving
face to go home to after a tough day. There will be another book. ChingJung and Eric,
are the starting actors, and I am grateful to be part of the cast.
x
xi
Contents
Chapter 1
Introduction 1
1.1 Background 1
1.2 Thermal Implication of Interconnect Scaling 4
1.3 Via Effect 8
1.4 Dissertation Organization 10
Chapter 2
Analytical Interconnect Thermal Model 13
2.1 Motivation 13
2.2 Preliminary Observations 16
2.2.1 First glance at via effect 17
2.3 Interconnect Thermal Model and Assumptions 22
2.3.1 Energy Conservation Law 22
2.3.2 Conduction Dominant Heat Transfer 24
2.3.3 Steady State Thermal Modeling 27
2.4 Impact of Via Effect on Effective k
ILD
35
2.4.1 Via Correction Factor 35
2.4.2 Hot Spot Location 36
2.5 Temperature in Multilevel Metal Layers 39
2.5.1 Multilevel interconnect Formula 39
2.5.2 Interconnect Temperature Rise Trend 41
2.6 Summary 43
xii
Chapter 3
Thermal Impact on Interconnect Design 45
3.1 Introduction 45
3.2 Thermal Effect on Interconnect Metrics 47
3.2.1 RC Delay 48
3.2.2 Dynamic Power Consumption 50
3.3.3 Cross Talk Noise
3.3.4 Electromigration Reliability 50
3.3 Scaling Trend of Joule Heating 52
3.4 Temperature Effect on Cu/lowk Interconnects 54
3.4.1 Definition of Worst Case Condition 54
3.4.2 Thermal Impact on Cu/lowk Interconnects 56
3.5 Delay and Reliability Optimization 58
3.6 Impact of Joule Heating on Scaling Trend 60
3.7 Summary 62
Chapter 4
SPICEBased ElectroThermal Simulation Methodology 63
4.1 Motivation 63
4.2 SPICEBased Thermal Modeling 64
4.3 Steady State Analysis 68
4.3.1 Impact of Via Separation on Effective k
ILD
68
4.3.2 Thermal Coupling Effect 72
4.4 Transient Stress Analysis 74
4.4.1 Analytical Model vs. SPICEBased Simulation
Methodology 75
4.4.2 Al vs. Cu Interconnects 78
4.4.3 Impact of Via Separation and Lowk Dielectrics 79
4.4.4 Impact of Dummy Thermal Vias 84
4.4.5 Impact of Interconnect Aspect Ratio 86
xiii
4.5 Conclusions 88
Chapter 5
Thermal nalysis of 3D IC 89
5.1 3D Integration: Background and Motivation 89
5.2 3D Thermal Modeling 94
5.3 Power Analysis of 3D ICs 99
5.4 Thermal Impact on Heterogeneous 3D Integration 106
5.5 Conclusions 108
Chapter 6
Conclusions 109
6.1 Summary 109
6.2 Future work 111
Bibliogrphy 113
xiv
xv
List of Tables
Table 1.1: Evolution of characteristics for Intel® commercial processors. ......................3
Table 3.1: Interconnect parameters for 130 nm to 22nm technology nodes based on
ITRS [5] ..........................................................................................................52
Table 3.2: Materials properties of various conductors and dielectrics used in this study.
k [W/mK] is thermal conductivity, ? [µOcm] is electrical resistivity near
100ºC and er is the relative dielectric constant..............................................56
Table 3.3: Coupled evaluation of electromigration reliability and performance for global
interconnects for 22130 nm technology nodes..............................................61
Table 4.1: Materials properties used in this work. ..........................................................69
Table 5.1: Comparison of power dissipation due to logic, interconnect, clock distribu
tion and repeaters for 2D and 3D ICs with 2 active layers for ITRS 50nm
technology node. 3D IC cases are presented for comparison by varying the
chip area, Ac, and operating frequency, fc, and represent the same 2D IC
(conserving feature size, number of transistors and functionality) converted to
3D with 2 active layers................................................................................101
xvi
xvii
List of Figures
Fig. 1.1: Predictions for device and wire delays from ITRS 2001 [5]............................2
Fig. 1.2: Evolution of power density for Intel processor [13] (Courtesy of Intel). ........4
Fig. 1.3: Schematic showing the hierarchy of metal levels for distribution of intercon
nects in modern ICs. Figure taken from ITRS 2001 [5]. ..................................6
Fig. 1.4: Both dielectric constant and thermal conductivity of ILD materials decrease
with advanced technology nodes [5]. ...............................................................7
Fig. 1.5: SEM photo of local level interconnect. ............................................................9
Fig. 2.1: Typical multilevel interconnect structure with heat sink attached to substrate.14
Fig. 2.2: Heat generated by Joule–heating is conducted downwards through both ILD
and vias. L represents the separation between vias. ......................................18
Fig. 2.3: Two extreme simulation conditions: (a) no current flows through via, so there
is no heat generation in the via; (b) all the current flowing through the wire
will also flow through the via and cause heat generation. It should be noted
that, in both cases, the via helps transfer heat to underlying layer and reduce
the temperature rise on the wire. ....................................................................19
Fig. 2.4: Temperature distribution along the metal wire with via separation of 250 µm
(top), 100 µm (middle) and 5 µm (bottom) for case (a) and (b) of Fig. 2.3. It
can be observed that there is no significant difference between the two simu
lation conditions and temperature rise in the via is minimum........................21
Fig. 2.5: A differential control volume, from a conducting metal wire embedded in a
dielectric medium, shows energy balance. .....................................................23
xviii
Fig. 2.6: With the help of heat sink, particularly with fan, the majority of heat will
conduct to outside ambient through substrate rather than to printed circuit
board (PCB) [35]. ...........................................................................................26
Fig. 2.7: Geometry used to derive interconnect thermal model, representing parallel
metal wire array. .............................................................................................27
Fig. 2.8: Geometry used for calculating Rspr, Rrect and the spreading factor S’m.
Crosssectional view is shown on the left hand side. Quasi 2D heat conduc
tion is used to correct 1D heat conduction model. Space between any two
wires is shared for heat dissipation. ................................................................32
Fig. 2.9: Temperature profile along the Cu wires with 100 µm via separation. H =
tILD = 0.8 µm, w = d = 0.3 µm, and jrms,m = 3.7E6 A/cm2. kpolymer = 0.3
W/mK and koxide =1.2 W/mK are assumed here. .........................................34
Fig. 2.10: Via correction factor, η, with two different D/W ratios, are plotted for three
types of dielectrics. Dielectrics with lower nominal kILD experience lower η
even with longer via separation. Wires with shorter via separation feel
stronger via effect. ..........................................................................................37
Fig. 2.11: Effective ILD thermal conductivity increases with decreasing via separation.
The lower the nominal kILD, the longer the LH,m, and hence, the stronger is
the via effect. ..................................................................................................37
Fig. 2.12: The figure shows the temperature profile along the wire from the middle of
the interconnect to the via at the right end. The hot spots usually locate in the
wire except for extremely small via dimensions. ...........................................39
Fig. 2.13: Temperature rise distribution along metal layers from substrate to top metal
level. In the case of via effect included, the via separations assigned to the
metal layers, from 1st to 10th levels, are 5, 10, 30, 50, 100, 150, 200, 300,
500 and 1000 metal pitches,respectively, based on 65nm technology node. .41
xix
Fig. 2.14: Trend of interconnect temperature rise along technology nodes for different
scenarios. (a) both j and kILD scale according to ITRS. (b) j scales with
ITRS, but kILD stops scaling at 65nm node. (c) j stops scaling at 65nm node,
while kILD continues scaling. (d) both j and kILD stops scaling at 65 nm
node. ...............................................................................................................42
Fig. 3.1: Schematic showing the intermetal and the interlevel components of capaci
tance. Also showing aspect ratio (AR). .........................................................48
Fig. 3.2: Trends of chip power density and interconnect Jmax along technology nodes
suggested by ITRS [5]. Chip power density is calculated by total power of
the chip divided by chip size. .........................................................................53
Fig. 3.3: Configurations of the three thermally worst case scenarios. Current flows
through all wires at all metal layers in cases (1) and (3). ...............................55
Fig. 3.4: Temperature on the toplevel interconnects rises rapidly with increasing
current density. ...............................................................................................55
Fig. 3.5: Temperature of top global interconnects rises sharply for Lowk dielectrics.
εr: relative dielectric constant, k: thermal conductivity [W/Km]. ..................56
Fig. 3.6: RC delay is strong function of current density on the wires because of Joule
heating. The lower the dielectric constant, the stronger the Joule heating and
greater RC degradation. ..................................................................................57
Fig. 3.7: Maximum allowed current density is limited by the maximum allowed ∆T on
the metal wires. The constraint is more stringent for lowk dielectrics. .......58
Fig. 3.8: Constant RmetCtotal contour plots of global level wiring as functions of
wire and ILD thickness. Solid curves represent the case thermal effect ne
glected. Dashline curves include the influence of thermal effect.
(H/S=0.32/0.15 µm). ......................................................................................59
xx
Fig. 3.9: Constant normalized electromigration MTF, MTF(Tm)/MTF(105°C), con
tour plots of global level wiring as functions of wire aspect ratio (AR) and
ILD thickness. (H/S=0.32/0.15 µm). .............................................................60
Fig. 3.10: The solid curve shows the MTF (%) that can be achieved under the current
density specified in ITRS. The brokenline curves show the MTF under vari
ous wire temperature (Tm) criterions and the values refer to the right axis...62
Fig. 4.1: ThermalElectrical analogous parameters. .....................................................65
Fig. 4.2: 3D thermal circuit RC transmission line model for transient thermal analysis
of interconnect structures. ..............................................................................66
Fig. 4.3: (a) Interconnect configuration. (b) Correspondent distributed thermal cir
cuitry. ..............................................................................................................68
Fig. 4.4: Validation of HSPICE thermal simulation with Finite Element Analysis
(ANSYS) data.................................................................................................69
Fig. 4.5: Effective ILD thermal conductivity increases with decreasing via separation.
The lower the nominal kILD, the longer the LH, and hence, the stronger is
the via effect. ..................................................................................................71
Fig. 4.6: (a) T is defined as the maximum temperature in the middle wire. Reference
temperature T0 is the maximum temperature in the middle wire with no cur
rent flowing in the nearest neighboring wires and with equal line width and
spacing (S). The current density in the middle wire is J0, 1.4x106 A/cm2.
Line width is kept at 0.3µm in this simulation. (b) The current density in the
middle wire is one third of J0. Line width is 0.3µm. The temperature of
wires carrying lower current density is strongly affected by the heat coupled
from the neighboring wires. (c) Top view of interconnect structures, with
xxi
different spacing, used in this simulation and current flow patterns showing
current flowing on alternate............................................................................73
Fig. 4.7: (a) Schematic cross section of Al and Cu interconnects with cladding layers
Fig. 4.7(b) and for simulations in Fig. 4.8. (b) Validation of HSPICE thermal
simulation with experimental data from [60]. Al interconnect test structure
shown in Fig. 4.7(a) is used. Two different wire thickness , t, are under ex
amination........................................................................................................77
Fig. 4.8: For the same wire cross sections, shown in Fig. 4.7(a), Cu wire shows much
lower temperature rise compared to Al wire under a 200 ns pulse stress.
Symbols represent simulation data. ................................................................78
Fig. 4.9: Simulated normalized temperature (∆T/∆Tmax) profile along global Cu/low
k (polymer) interconnect for two pulse durations with current density
J=4x107A/cm2. ..............................................................................................80
Fig. 4.10: The temperature decay after a 200 ns current pulse with J=8x107 A/cm2, for
Cu/lowk (polymer) global wires. The decay is facilitated by the presence of
vias..................................................................................................................80
Fig. 4.11: (a) ∆Tmax of global interconnects for a 200 ns current pulse with J=6x107
A/cm2. ∆Tmax increases with via separation and saturates when via effect
has diminished. (b) For a short pulse of 100 ns, with the same current density
as in Fig. 4.11(a), ∆Tmax of the global Cu interconnect is much less depend
ent on the dielectric materials. ........................................................................82
Fig. 4.12: (a) ∆Tmax of Cu global interconnect with 100 µm via separation under a 200
ns current pulse. Temperature rises sharply with current density. (b) ∆Tmax
of Cu local interconnect with 1 µm via separation under a 200 ns current
pulse. Temperature rises are much lower and nearly independent of the sur
rounding dielectric materials. .........................................................................83
xxii
Fig. 4.13: Temperature profiles along local Cu interconnect under a 200 ns pulse are
gradual for all the dielectric materials and ∆Tmax are about the same with
J=1x108 A/cm2 ..............................................................................................84
Fig. 4.14: (a) ∆Tmax of Cu/air and Cu/polymer can match ∆Tmax of Cu/oxide global
interconnect if dummy thermal vias are added every 20 µm and 30 µm,
respectively, with tpulse = 200 ns and J=6x107 A/cm2. (b) Cu/air with
thermal vias every 20 um shows nearly the same ∆Tmax as that of Cu/oxide
global interconnect with 100 µm via separation under a 200 ns and J=6x107
A/cm2 current pulse .......................................................................................85
Fig. 4.15: (a) For the same metal thickness (2.5 µm), global interconnects with lower
aspect ratio (AR), shows higher ∆Tmax due to lower surface areatovolume
ratio. tpulse = 200 ns and J=6x107 A/cm2. (b) For the same metal wire cross
section area (5 µm2), the impact of aspect ratio (AR) and dielectric structure
strategy is shown. ∆Tmax peaks at aspect ratio=1 due to the smallest perime
ters. tpulse = 200 ns and J=6x107 A/cm2 .....................................................87
Fig. 5.1: Schematic of a 3D chip showing integrated heterogeneous technologies ....90
Fig. 5.2: Schematic representation of 3D integration with multilevel wiring network
and VILICs. T1: first active layer device, T2: second active layer device, Op
tical I/O device: third active layer I/O device. M’1 and M’2 are for T1, M1
and M2 are for T2. M3 and M4 are shared by T1, T2, and the I/O device. ..91
Fig. 5.3: Horizontal interconnects are replaced with VILICs, reducing wiring require
ment, chip area and interconnect delays .........................................................92
Fig. 5.4: Schematic of multi level 3D IC with a heat sink attached to Si substrate. ...94
Fig. 5.5: The required package thermal resistance, Rpkg, to achieve the maximum
junction temperature specified in ITRS and the ratio of RILD and Rpkg vs.
technology nodes ............................................................................................95
xxiii
Fig. 5.6: ElmoreDelay model – electrical analogy for the thermal mode employed
[45]. ................................................................................................................97
Fig. 5.7: Schematic of 3D structure fabricated by wafer bonding using Cu pad ther
mocompression. ..............................................................................................98
Fig. 5.8: Temperature distribution along the vertical layers from the Si substrate
surface (Si_1) to top metal level of the second stratum. The 3D structure is
shown in Fig. 6. M11 represents the first metal level in stratum 1, etc.........99
Fig. 5.9: Comparison of temperature performance among 2D ICs and five different
twoactivelayer 3D ICs scenarios. .............................................................103
Fig. 5.10: Thermal capability of high performance 3D ICs (Case 4) for four different 3
D logicmemory integration schemes...........................................................105
Fig. 5.11: The noise performance of a typical low noise amplifier (LNA) under
the temperature effect is evaluated. NF is assumed to be 2dB @ 290K. ....105
Fig. 5.12: Temperature effect on optical receiver performance is evaluated. A typical
example of TiSi metalsemiconductormetal (MSM) PD dark current vs.
ambience temperature is shown here............................................................107
xxiv
`
1
Chapter 1
Introduction
1.1 Background
The unprecedented prosperity of information technology has driven semicon
ductor industry to evolve at an incredible rate. The growing demand for higher
performance in integrated circuits (ICs) results in faster device switching speed,
greater number of transistors, increased functional density and larger chip size [1, 2].
Consequently, the communication, supported by onchip interconnects, between de
vices and between circuit blocks is becoming more complex and a challenging
problem. The connection of miniaturized and closely packed transistors requires re
duced wire crosssections in the local levels, while the rapid increase in functional
density and chip size leads to longerdistance communication in the global levels.
With the aggressive scaling of VLSI technology, the interconnect delay, due to longer
wires and smaller wire pitch, now plays a key role and becomes bottleneck in the con
Chapter 1: Introduction
2
tinued improvements in ICs density and speed [3, 4]. As illustrated by the Interna
tional Technology Roadmap for Semiconductors (ITRS) [5], interconnect RC delay is
dominating the chip performance in advanced technology nodes (Fig. 1.1).
Fig. 1.1: Predictions for device and wire delays from ITRS 2001 [5].
To mitigate this adversescaling trend of signal propagation (RC) delay, consider
able work has gone into overcoming the interconnect limitations. For instance,
hierarchical interconnect structure has been adopted to incorporate several metal levels,
so that long global interconnects can be routed to higher tier and maintain larger cross
sections to minimize wire resistance [68]. This interconnect architecture along with the
Technology Node [nm]
R
e
l
a
t
i
v
e
D
e
l
a
y
Technology Node [nm]
R
e
l
a
t
i
v
e
D
e
l
a
y
Section: 1.1 Background
3
demand for more wiring has resulted in an increasing number of interconnect metal
levels.
Table 1.1: Evolution of characteristics for Intel
®
commercial processors
(Table 1.1). In addition, Cu has replaced Al as the interconnect metal for its lower
resistivity [9, 10] and low dielectric constant (lowk) materials has been pursued as an
alternative to silicon dioxide to reduce interconnect capacitance [11, 12]. However,
although all these tremendous efforts have salvaged the attentiongetting interconnect
delay problem, the interconnect reliability due to electromigration and thermal effects is
quickly emerging as a serious integration issue. In addition, the fast increasing power
density in ICs (Fig. 1.2) [13], together with the scaling trend toward lower power supply
(Vdd) to reduce short channel effect, hot electrons, gate stress and leakage [14, 15], has
significantly increased the current density in interconnects. Furthermore, interconnects
are farther away from the substrate, where heat sink is usually attached. As a result,
interconnects generally experience higher average temperature than transistors do.
Year 1972 1982 1993 2002
Processor 8008 286 Pentium
®
Itanium 2
®
Technology Node 10 µm 1.5 µm 0.8 µm 0.18 µm
Frequency (MHz) 0.2 6 60 1000
Transistors 3500 120,000 3,100,000 221,000,000
Chip Area (mm
2
) 15.2 68.7 217 421
Metal Layers
1
2 3 6
Chapter 1: Introduction
4
Fig. 1.2: Evolution of power density for Intel processor [13] (Courtesy of Intel).
It is not difficult to imagine that not only will interconnect delay be a major show
stopper for continuous ICs performance advancement, but also the concern of reliability
and thermal effect of interconnect can potentially become another serious system design
constraint. In short, as semiconductor technology strives to keep up with Moore’s Law
[1], it is imperative to understand and analyze the emerging VLSI “hot wire” phenomena
to handle the impact of interconnect thermal effect in early design phase.
1.2 Thermal Implication of Interconnect Scaling
Thermal effects are an inseparable aspect of electrical power distribution and signal
transmission through interconnect nets due to Joule heating (or selfheating) caused by
Nuclear Reactor
Rocket Nozzle
Sun’s Surface
Nuclear Reactor Nuclear Reactor Nuclear Reactor
Rocket Nozzle
Sun’s Surface
Rocket Nozzle
Sun’s Surface
Section: 1.2 Thermal Implication of Interconnect Scaling
5
current flow in the wire. The temperature rise, on top of the substrate temperature, in the
interconnects is determined by the product of Joule heating power dissipation and the
thermal impedance from the wires to the substrate. Before proceeding to further discus
sion, it is instructive to be familiar with basic onchip interconnect structure and
terminology. The interconnects (or wires) in a modern IC chip can be broadly character
ized into three groups according to the functions they perform. These are the signaling
interconnects, the clock distribution interconnects and the power and ground supply
distribution interconnects. Interconnects can be further subdivided into three categories
according to the range of their lengths and their cross section dimension. These are the
local, semiglobal/intermediate and the global interconnects. The global interconnects are
responsible for long distance communication on a chip and have a larger cross sectional
area to minimize resistance. Whereas, the local interconnects have the shortest range and
the tightest cross sectional dimensions. Modern ICs have multiple levels of interconnects
to accommodate their large numbers, starting from the local at the lowest level to the
global at the top most levels. This is depicted in Fig. 1.3 [5]. The lower level intercon
nects are used in local routing and they connect transistors that are a few microns apart.
The higher level interconnects are used for global routing and they can span across an
entire chip and may be as long as 23 cm in modern highperformance chips. Vias are the
metal fillings enabling interlevel wire connections. The wires are separated by inter
layer dielectrics (ILD) from level to level and isolated by intermetal dielectrics (IMD)
within the same level. Traditionally, silicon dioxide (SiO
2
) has been
Chapter 1: Introduction
6
the choice of the dielectrics in ICs for its superior properties. However, a variety of low
dielectric constant materials, including organic and inorganic films, as well as the use of
porosity and airgaps [1618], have been fiercely pursued since a decade ago for the
benefit of reducing wire capacitance, crosstalk noise and dynamic power consumption.
Unfortunately, these lowk dielectrics inevitably have very poor thermal properties (Fig.
1.4 [5]). Compounded with the poor thermal conductivity of such materials and more
metal levels added with more advanced technology nodes, the increasing thermal imped
ance further exacerbates temperature rise in interconnects.
Fig. 1.3: Schematic showing the hierarchy of metal levels for distribution of intercon
nects in modern ICs. Figure taken from ITRS 2001 [5].
Section: 1.2 Thermal Implication of Interconnect Scaling
7
On the other hand, the power consumption in the chips is rapidly becoming unman
ageable as clock frequency continues to climb and chip temperature has been always a
major concern to further chip advancement. The reduction in supply voltage slows down
the power consumption at the transistor level; however, the resultant increased current
requirement leads to greater interconnect Joule heating. The typical supply current is in
the 100A range in modern chips and is expected to increase for future processors.
Although the power dissipated by Joule heating is not a significant portion of the total
chip power consumption yet, the interconnects are separated from the substrate by
dielectrics with low thermal conductivity, and heat cannot be removed efficiently
Therefore the temperature rise in interconnects can be nonnegligible. The interconnect
35 50 70 130 180 100
0
1
2
3
4
0
0.2
0.4
0.6
0.8
1
1.2
Technology Node [nm]
D
i
e
l
e
c
t
r
i
c
C
o
n
s
t
a
n
t
T
h
e
r
m
a
l
C
o
n
d
u
c
t
i
v
i
t
y
[
W
/
m
K
]
]
35 50 70 130 180 100
0
1
2
3
4
0
0.2
0.4
0.6
0.8
1
1.2
Technology Node [nm]
D
i
e
l
e
c
t
r
i
c
C
o
n
s
t
a
n
t
T
h
e
r
m
a
l
C
o
n
d
u
c
t
i
v
i
t
y
[
W
/
m
K
]
]
Fig. 1.4: Both dielectric constant and thermal conductivity of ILD materials
decrease with advanced technology nodes [5].
Chapter 1: Introduction
8
structures with decreasing critical dimension will inevitably experience everhigher
current density and, the resultant greater temperature rise.
1.3 Via effect
Since the lifetime of interconnects has an exponential dependence on the inverse
metal temperature owing to electromigration [19, 20], thermal effect will limit the
maximum allowable current density in the wires to limit the temperature rise in intercon
nects. It can be envisioned that, to some extent, the thermal effect may not only be a
major reliability concern, but also constrain the speed performance due to limited current
drive capability. With the aggressive scaling trend, it is essential not to underestimate
the impact of interconnect temperature rise. On the contrary, overly pessimistic estima
tion of the interconnect temperature will lead to overly conservative design. Therefore,
accurate temperature estimation is extremely important to regulate the design rules and,
hence, performing a more realistic thermal analysis of interconnects is critical.
Section: 1.3 Via effect
9
Several publications have addressed the issue of lowk dielectrics and their impact
on interconnect temperature, reliability and performance [21, 22]. Poor thermal conduc
tivity of lowk insulators has been a major concern to cause substantial rise in
interconnect temperature. However, the effect of vias (Fig. 1.5), which have much higher
thermal conductivity than the dielectrics and therefore can serve as efficient heat dissipa
tion paths, has not been adequately addressed in those simplified thermal model [23, 24].
Consequently, if via effect is ignored, the predicted temperature rise of interconnects can
be much higher than the reality. Neglecting via effect on interconnect temperature
estimation can lead to intolerable prediction errors. This research proposes both compact
analytical models and fast SPICE based 3D electrothermal simulation methodology to
Metal
wire
Metal
via
Transistors
Metal
wire
Metal
via
Transistors
Fig. 1.5: SEM photo of local level interconnect
Chapter 1: Introduction
10
characterize thermal effects due to Joule heating in high performance Cu/lowk intercon
nects under steadystate and transient stress conditions. The results demonstrate excellent
agreement with experimental data and those using Finite Element (FE) thermal simula
tions (ANSYS). The effect of vias, as efficient heat transfer paths to alleviate the
temperature rise in the metal wires, is included in our analysis for the first time to provide
more accurate and realistic thermal diagnosis.
1.4 Dissertation Organization
With the pursuit of realistic evaluation in mind, via effect is included in all the
thermal models and simulations. The impact of Joule heating on the scaling trends of
advanced VLSI interconnects has been evaluated in detail.
Chapter 2 derives the analytical interconnect thermal models to provide building
blocks, which will enable us to quantify the impact of via effect and various interconnect
parameters on the interconnect temperature rise. The model is then applied to estimate
the temperature rise of densely packed multilevel interconnects.
Chapter 3 discusses the impact of Joule heating on the scaling trends of advanced
VLSI interconnects. Through a combination of extensive electrothermal simulation and
2D field solver for capacitance calculation, the thermal characteristics of various Cu/low
k schemes are quantified and their effects on electromigration reliability and interconnect
delay is determined.
Section: 1.4 Dissertation Organization
11
Chapter 4 describes a compact 3D electrothermal simulation methodology to
evaluate interconnect design options from a thermal point of view. Use of dummy
thermal vias as additional heat sinking paths is evaluated to alleviate the thermal impact
on interconnect reliability. Furthermore, this simulation methodology provides an
efficient diagnosis tool for transient thermal stress analysis, which is usually extremely
cumbersome, mostly unsolvable, in analytical forms.
Chapter 5 presents detailed thermal analysis of high performance three dimensional
(3D) ICs under various integration schemes. A complete thermal model including
power consumption due to both transistors and interconnect joule heating from multiple
strata is presented. Furthermore, tradeoffs among power, performance, chip real estate
and thermal impact for 3D ICs are evaluated.
Chapter 6 summarizes the main findings and concludes this dissertation with sug
gestion on possible future work.
Chapter 1: Introduction
12
13
Chapter 2
Analytical Interconnect Thermal Model
2.1 Motivation
The conventional aluminum(Al)/silicon dioxide (SiO
2
) interconnect scheme
reached its performance limitation a decade ago [3]. Hence, lower wire resistance and
higher current density requirements have prompted an industrywide shift from alumi
num(Al) to copper(Cu). Similarly, the minimization of interconnect capacitance is
driving the introduction of a variety of lowpermittivity dielectric materials. Since these
lowk dielectrics also have lower thermal conductivity than silicon dioxide, heat dissipa
tion in the interconnect level has become more difficult. Furthermore, VLSI scaling
trends, such as increasing number of interconnect metal levels and increasing current
density, have caused everprofound thermal effects in the interconnects. In a stateof –art
chip, the interconnect temperature can rise more than 100 degrees above the ambient
Chapter 2: Analytical Interconnect Thermal Model
14
temperature due to Joule heating and the heat dissipation from transistors in the substrate
level.
Although knowing the interconnect temperature is the first step toward any kind of
further thermal effect analysis, the temperature profile of a single interconnect is difficult
enough to be obtained experimentally for its miniature scale [25], not to mention multi
level measurements. In general, interconnect thermometry based on temperature
dependent electrical resistivity of the wire is used to calculate a spatially averaged
Heat sink
via
M8
M7
M6
Substrate
I am Hot !
Heat sink
via
M8
M7
M6
Substrate
Heat sink
via
M8
M7
M6
via
M8
M7
M6
Substrate
I am Hot ! I am Hot !
Fig. 2.1: Typical multilevel interconnect structure with heat sink attached to substrate.
Section: 2.1 Motivation
15
temperature rise of the interconnects [26]. However, it can provide neither the local
temperature rise profile on the wire nor the temperature distribution along the multilevel
interconnects. On the other hand, various papers have been devoted to deriving analyti
cal expressions for the temperature distribution in IC chips [2729]. However, most of
them only consider a single heat source without considering the effect of vias, which is
not the realistic case for most interconnects. In addition, the complicated nature of these
expressions makes it very difficult to easily apply them to multilevel metal layers.
Recently, Im [24] and Hunter [30, 31] have proposed simplified formule, while Chen [23]
and Rzepka [32] applied 3D finite element analysis to estimate temperature rise on
multilevel interconnects. However, the effect of vias, as an efficient path for conduction
of heat has not been included in those simplified thermal model and simulation. Conse
quently, the predicted temperature rise of interconnects can be much higher than actually
observed.
Analytical thermal models are desirable to facilitate quick estimation of tempera
ture rise in order to provide better insight and thermal design guidelines for advanced
interconnect structures. This chapter presents compact analytical thermal models for
estimating the temperature rise of multilevel VLSI interconnect lines incorporating the
effect of vias as efficient paths for conduction of heat. The impact of vias has been
modeled using i) a characteristic thermal length representing the distance along a metal
line through which vias effectively remove the heat, and ii) an effective thermal conduc
tivity of interlayer dielectrics (ILD), k
ILD,eff
, with k
ILD,eff
= k
ILD
/η
,
where k
ILD
is the
Chapter 2: Analytical Interconnect Thermal Model
16
nominal thermal conductivity of the ILD and η is a physical correction factor represent
ing the heat removal by vias.. Both the spatial temperature profile along the metal lines
and their average temperature rise can be easily obtained using these models. The
predicted temperature profiles are shown to be in excellent agreement with the 3D finite
element thermal simulation results. The model is then applied to estimate the tempera
ture rise of densely packed multilevel interconnects. It is shown that in multilevel
interconnect arrays, via density along the lines can significantly affect the temperature
rise of such interconnect structures.. The influence of via self heating and thermal
impedance of vias is considered here, providing more comprehensive and accurate
results.
The remainder of this chapter is organized as follows. To provide a baseline, in
Section 2.2, we validate that it is a legitimate approach to consider via effect. Sections
2.3 and 2.4 describe the derivation of our analytical model and discuss the impact of via
effect. In Section 2.5, we apply the model to multilevel interconnects and discuss the
temperature distribution along metal layers. The main results are summarized in Section
2.6.
2.2 Preliminary Observations
The fact that the interconnect temperature does not increase inversely proportional
to the nominal thermal conductivity of dielectrics suggests that the vias, which have
Section: 2.2 Preliminary Observations
17
much higher thermal conductivity than ILD, serve as efficient heat dissipation paths. If
the via effect is not properly considered, the predicted temperature will be significantly
higher than the temperature in practical situations. As a result, there is danger of signifi
cant errors in using overly simplistic assumptions to assess thermal problems in advanced
interconnect structures with lowk materials. In this work, the via effect is evaluated in
details and is incorporated into our multilevel temperature distribution model. Further
more, due to the interconnect architecture that, in general, lower metal levels have much
higher via density than the higher levels, the temperature rise among metal layers is
expected to be quite different from what is predicted by the simplistic model.
2.2.1 First glance at via effect
Since the cross sectional area of vias is generally smaller than that of the wires, it
may raise a concern that the resultant higher current density in the vias may generate
significant heat, deteriorating their effectiveness in heat conduction. However, as shown
in [33], the temperature rise in the via is not as high as that in the wire. A simulation
validation was obtained by using ANSYS, a three dimensional finite element thermal
simulation package. The configuration simulated, as shown in Fig. 2.2, is Cu wire/via
embedded in a lowk dielectric with via separation, L, varied as 250 µm, 100 µm and 5
µm. The metal width is taken to be 0.3 µm and both of the metal height and via height
are 0.6 µm. Polymer is chosen as the lowk insulator with nominal thermal conductivity
of 0.3 W/mK. The electrical resistivity and thermal conductivity of Cu are assumed to
be 2.2 µΩcm and 400 W/mK. The RMS current density of 2.1e6 A/cm
2
is flowing in
Chapter 2: Analytical Interconnect Thermal Model
18
the wire.
A uniform temperature, T
0
, is assigned to the underlying layer of the block.
The temperature rise of wires on top of substrate temperature is represented by θ
m
(x) ≡
T
m
(x)T
0
, where T
m
(x) is the wire temperature. Adiabatic boundary conditions are applied
to the top and four sides of the block where symmetry boundary conditions are satisfied.
Two extreme simulation conditions have been carried out in Fig. 2.3, with no cur
rent flowing through the via, condition (a); and with all the current flowing through via,
condition (b). Therefore, only the thermal impedance caused by vias, R
th, v,
is included in
the analysis in condition (a), while the extra heat generated in the via due to current
Via
M
N
ILD
M
N1
Via
M
N
ILD
M
N 1
Heat conduction
L= via separation
via
wire
T
0
T
m
(x)
via via
Via
M
N
ILD
M
N1
Via
M
N
ILD
M
N 1
Heat conduction
L= via separation
via
wire
T
0
T
m
(x)
via via
Fig. 2.2: Heat generated by Joule–heating is conducted downwards through both ILD
and vias. L represents the separation between vias.
Section: 2.2 Preliminary Observations
19
current
x=0
x=L/2
current
x=0
x=L/2
(a)
(b)
current
x=0
x=L/2
x=0
x=L/2
current
x=0
x=L/2
x=0
x=L/2
(a)
(b)
Fig. 2.3: Two extreme simulation conditions: (a) no current flows through via, so there
is no heat generation in the via; (b) all the current flowing through the wire
will also flow through the via and cause heat generation. It should be noted
that, in both cases, the via helps transfer heat to underlying layer and reduce
the temperature rise on the wire.
Chapter 2: Analytical Interconnect Thermal Model
20
density flowing in the via, J
v
, is also taken into account in condition (b). In practice, all
the vias do not carry the same amount of current as in the wire.
The comparison of temperature distributions along the metal wires are shown in
Fig. 2.4 (a), (b) and (c) for via separation of 250 µm, 100 µm and 5 µm, respectively.
The temperature profiles of the simplified 1D thermal model, constant temperature
value, with via effect completely ignored are also plotted in Fig. 2.4 (a) and (b). It has
the same value as well for the case of 5 µm via separation, although it is not plotted in
Fig. 2.4 (c) to keep other curves more readable. It should be noted that, with shorter via
separation, the average temperature as well as the peak temperature in the wire is re
duced. It appears that within a certain range from via, the temperature reduction with the
help of vias is particularly consumptious. In addition, the simulations show that the
temperature rise in the vias is minimal. Even though percentage wise, the via contact
exhibits a major contribution to the temperature rise in the wire for short via separation,
the absolute value is quite small. The vias are simply too short to produce much heat.
Furthermore, under normal operating conditions, all the vias are not conducting current at
the same time but they always help dissipate heat. Consequently, it is legitimate to
consider vias as efficient thermal sink paths. Finally, as long as the vias are connected to
lowertemperature underlying layers, they will help dissipate heat no matter how far they
are away from substrate. With the above preliminary observations in mind, in the follow
ing sections, we will quantify via effect and evaluate if vias are always helpful.
Section: 2.2 Preliminary Observations
21
0
1
2
3
4
5
6
7
0 25 50 75 100 125
via effect ignored
both Jv and Rth,v included
only Rth,v included
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
[
°
C
]
(a)
0
1
2
3
4
5
6
7
0 10 20 30 40 50
via effect ignored
both Jv and Rth,v included
only Rth,v included
0
0.05
0.1
0.15
0.2
0 0.5 1 1.5 2 2.5
both Jv and Rth,v included
only Rth,v included
Location along Interconnect [um]
Location along Interconnect [um]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
[
°
C
]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
[
°
C
]
(b)
(c)
Location along Interconnect [um]
0
1
2
3
4
5
6
7
0 25 50 75 100 125
via effect ignored
both Jv and Rth,v included
only Rth,v included
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
[
°
C
]
(a)
0
1
2
3
4
5
6
7
0 10 20 30 40 50
via effect ignored
both Jv and Rth,v included
only Rth,v included
0
0.05
0.1
0.15
0.2
0 0.5 1 1.5 2 2.5
both Jv and Rth,v included
only Rth,v included
Location along Interconnect [um]
Location along Interconnect [um]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
[
°
C
]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
[
°
C
]
(b)
(c)
Location along Interconnect [um]
Fig. 2.4: Temperature distribution along the metal wire with via separation of 250
µm (top), 100 µm (middle) and 5 µm (bottom) for case (a) and (b) of Fig.
2.3. It can be observed that there is no significant difference between the
two simulation conditions and temperature rise in the via is minimum.
Chapter 2: Analytical Interconnect Thermal Model
22
2.3 Interconnect Thermal Model and Assumptions
2.3.1 Energy conservation law
In either steady state or transient condition, the first principle of energy conserva
tion law must be satisfied at any instance,
dt
dE
E E E
st
out gen in = − +
. . .
(2.1)
where the thermal energy entering, leaving, and generated in the control volume are Ė
in
,
Ė
out
,
and Ė
gen
, respectively. The rate of change of energy stored within the control volume
is designated as dE
st
/dt. Consider a rectangular metal wire with crosssectional area A
c
,
and periphery p embedded in a dielectric medium as shown in Fig. 2.5. The temperature
variation can be modeled as a onedimensional situation which refers to the fact that only
one coordinate (x) is needed to describe the spatial temperature variation. The tempera
ture gradient along y and z directions is negligible for a long metal wire with relatively
small cross section and high thermal conductivity. Garden has proven that the error due
to this simplification is less than 1% even for larger cross sections [34]. The conduction
heat rate perpendicular to the control surface at x is indicated by the term q
x
.
c m x
A
x
T
k q
∂
∂
= (2.2)
where k
m
is the thermal conductivity of the metal wire and A
c
is the crosssectional area
of the wire. The conduction heat rate at the opposite surface can then be expressed as a
Taylor series expansion where, neglecting higher order terms,
Section: 2.3 Interconnect Thermal Model and Assumptions
23
dx
x
q
q q
x
x dx x
∂
∂
+ =
+ (2.3)
dq
con
represents the heat conducted to the surrounding medium. It is proportional to the
difference between the surface and the surrounding temperatures and the contiguous
surface area.
) ) ( )( ' (
∞
− = T x T dA S h dq
m s m con
(2.4)
Ac
q
x
q
x+dx
dx
dA
s
=pdx
q”’
dq
con
dx
x
y
z
t
ILD
w
Ac
q
x
q
x+dx
dx
dA
s
=pdx
q”’
dq
con
dx
x
y
z
t
ILD
w
Fig. 2.5: A differential control volume, from a conducting metal wire embedded in a
dielectric medium, shows energy balance.
Chapter 2: Analytical Interconnect Thermal Model
24
where the proportionality constant, h, is termed as heat transfer coefficient. dA
s
is the
surface area where heat in the control volume can be transferred into the medium. S’
m
is
the spreading factor to model the multidimensional heat spreading phenomena. T
m
(x) is
the metal temperature at x and T
∞
is the temperature of the medium far from the metal
surface.
As required by energy conservation law Eq. 2.1, the heat entering the differential
control volume (q
x
q
x+dx
) plus the volumetric rate of thermal energy generation in the
control volume (q
’’’
) minus the heat dissipated to the medium (dq
con
) must be equal to the
rate of change of thermal energy stored within the differential control volume. Therefore,
the governing heat diffusion equation with temperature variation dominating in x direc
tion can be concluded as
t
T
A c A q T T p hS
x
T
A k
x
c p c m c m
∂
∂
= + − −
∂
∂
∂
∂
∞
ρ ' ' ' ) ( ' ] [ (2.5)
where ρ is the metal density and c
p
is the metal specific heat. The product, ρ c
p
[J/m
3
K],
commonly termed the volumetric heat capacity, measures the ability of a material to store
thermal energy.
2.3.2 Conduction dominant heat transfer
Since, in general, thermally insulated package materials encapsulate the IC chip,
heat convection to the ambient air is ignored by the application of the adiabatic boundary
condition on the four side walls and top of the chip. In addition, heat radiation is simply
too small to be taken into account. Therefore, only heat conduction will be considered as
Section: 2.3 Interconnect Thermal Model and Assumptions
25
the prevailing heat transfer mechanism. Therefore, the net heat transfer processes can be
quantified by the rate equation known as Fourier’s law,
T k q ∇ − =
"
(2.6)
where q” [W/m
2
] is the heat flux and k [W/mK] is the thermal conductivity. Further
more, the substrate, to which a heat sink is usually attached, is assumed to be the sole
heat dissipation path to the outside ambient, hence only heat conduction toward the
substrate will be considered in our interconnect thermal modeling. This assumption is
validated in Fig. 2.6, as Zhou has shown that, with heat sink attached, the majority of heat
will transfer through the substrate [35].
Eq. 2.4 can then be further simplified if we consider heat is only transferred down
wards and only heat conduction mechanism takes place, i.e.
) ) ( )( ' )( (
0
T x T wdx S
t
k
dq
m m
ILD
d
con
− = (2.7)
where k
d
is the thermal conductivity of the dielectric medium and t
ILD
is the underlying
dielectric thickness. k
d
/ t
ILD
acting as the proportionality constant and can be treated as
an equivalent conduction heat transfer coefficient. w is the width of the metal wire and
T
0
is the temperature of underlying layer. Accordingly, Eq. 2.5 results in
t
T
A c A q T T w S
t
k
x
T
A k
x
c p c m
ILD
D
c m
∂
∂
= + − −
∂
∂
∂
∂
∞
ρ ' ' ' ) )( ' )( ( ] [ (2.8)
Chapter 2: Analytical Interconnect Thermal Model
26
0
20
40
60
80
100
no heat
sink
fin heat
sink
fan heat
sink
%
h
e
a
t
d
i
s
s
i
p
a
t
e
d
cond. to PCB
thru. substrate
0
20
40
60
80
100
no heat
sink
fin heat
sink
fan heat
sink
%
h
e
a
t
d
i
s
s
i
p
a
t
e
d
cond. to PCB
thru. substrate
Heat sink
PCB
Heat sink
PCB
Flip Chip Assembly
Fig. 2.6: With the help of heat sink, particularly with fan, the majority of heat will
conduct to outside ambient through substrate rather than to printed circuit
board (PCB) [35].
Section: 2.3 Interconnect Thermal Model and Assumptions
27
2.3.3 Steady state thermal modeling
In this section, we will develop the interconnect thermal modeling under normal
chip operation circumstances. The thermal environment on the interconnects is consid
ered to be under steadystate condition. Transient stress analysis, such as under
electrostatic discharge (ESD), will be discussed in chapter 4. To begin with, consider a
rectangular metal wire with thickness H, width w
,
length L, electrical resistivity ρ
m
and
thermal conductivity k
m
, separated from the underlying layer by interlayer dielectrics
(ILD) of thickness t
ILD
and thermal conductivity k
ILD
. The two ends, at x=±L/2, of the
wire are connected to the underlying layer through vias (Fig. 2.7). The vias are modeled
x=0
y=0
L
t
ILD
H
W
T
0
D
T
v
(y)
T
m
(x)
d
x=0
y=0
L
t
ILD
t
ILD
H
W
T
0
DD
T
v
(y)
T
m
(x)
d
Fig. 2.7: Geometry used to derive interconnect thermal model, representing parallel
metal wire array.
Chapter 2: Analytical Interconnect Thermal Model
28
as round pillars, with diameter D and height t
ILD
. The electrical resistivity and thermal
conductivity of vias are ρ
v
and
k
v
. The underlying layer temperature set to be T
0
, and
therefore temperature at the via bottom T
v
(0)=T
0.
The temperature rise of wires and vias
is represented by θ
m
(x) ≡ T
m
(x)T
0
and θ
v
(y) ≡ T
v
(y)T
0
, respectively. It is assumed that
heat only flows downwards toward silicon substrate which is usually attached to a heat
sink. Following Eq. 2.8, under steady state conditions, with uniform rootmeansquare
current, j
rms
, flowing in the conductors, the governing heat equations for the wire and the
via are
m
m
m H
m m
k
q
L
x
dx
x d
' ' '
2
,
2
2
) ( ) (
− = −
θ θ
, (2.9)
and
v
v
v H
v v
k
q
L
y
dy
y d
' ' '
2
,
2
2
) ( ) (
− = −
θ θ
, (2.10)
where q
m
’’’
(=j
2
rms,m
ρ
m
) and q
v
’’’
(=j
2
rms,v
ρ
v
) are the volumetric heat generation in the wire
and the via due to Joule heating. L
H,m
and L
H,v
are the thermal healing length of the wire
and the via, defined as
2
1
2
1
, (
¸
(
¸
=
(
¸
(
¸
≡
m ILD
m
m ILD
m m
m H
S k
wH k
S k
A k
L , (2.11)
and
2
1
2 2
1
,
) 4 (
(
¸
(
¸
=
(
¸
(
¸
≡
v ILD
v
v ILD
v v
v H
S k
D k
S k
A k
L
π
, (2.12)
where S
m
and S
v
are the shape factor per unit length of the wire and the via, respectively,
to represent heat spreading and, thus, to correct the deviation from onedimensional heat
Section: 2.3 Interconnect Thermal Model and Assumptions
29
flow. It should be noted that the shape factor is only a function of geometric parameters.
As it will be seen later, shape factor S
m
and spreading factor S’
m
are interchangeable and
both describe the multidimension heat conduction flow from metal surface to surround
ing medium. The physical meaning of L
H,m
is that within the range of thermal healing
length, L
H,m
, from vias, heat generated will flow through vias to the underlying layer.
Beyond L
H,m
, heat will flow through ILD and the via effect is diminished.
The two boundary conditions used to solve Eq. 2.9 are the adiabatic condition at
the middle of the wire (x=0) due to symmetry, and the junction temperature θ
J
at the two
ends of the wire (x=±L/2), with θ
J
=θ
m
(x=±L/2)=θ
v
(y=t
ILD
).
, 0
) (
0
=
= x
m
dx
x dθ
θ
m
(x=±L/2)= θ
J,
(2.13)
In addition, the two boundary conditions to solve Eq. 2.10 are the temperature rise θ
J
at
the top of the via (y=t
ILD
) and zero at the bottom of the via (y=0).
θ
v
(y=t
ILD
)= θ
J,
θ
v
(y=0)=0. (2.14)
The temperature rise along the wire and within the via can then be solved from Eq. 2.9
and Eq. 2.10 as
( )
( )
( )
( )
( )


.

\

− + =
m H
m H
m
m H m
m H
m H
J m
L L
L x
k
L q
L L
L x
x
,
,
2
,
' ' '
,
,
2 cosh
cosh
1
2 cosh
cosh
θ θ , (2.15)
( )
( )
( )
v H ILD
v H
J v
L t
L y
y
,
,
sinh
sinh
θ θ =
( )
( )


.

\

− +
− +
v H ILD
v H ILD v H
v
v H v
L t
L y t L y
k
L q
,
, ,
2
,
' ' '
sinh
] / ) sinh[( sinh
1
, (2.16)
Chapter 2: Analytical Interconnect Thermal Model
30
with L/2≤ x≤ L/2 and 0≤ y≤ t
ILD
. It is instructive to note that we can define q
m
’’’
L
H,m
2
/k
m
(=
j
rms,m
2
ρ
m
L
H,m
2
/k
m
=
j
rms,m
2
ρ
m
Hw/k
ILD
S
m
) in the second term of Eq. 2.15 as ∆T
1D
, which
represents the temperature rise in the wire obtained from simplified 1D heat conduction
model where via effect is not included. Thereafter, the wire temperature rise, θ
m
(x), from
Eq. 2.15 can be interpreted as the combined contribution from via selfheating and the
thermal impedance caused by the via (first term) and the Joule heating in the wire cor
rected with the help of via effect (second term). The ratio of θ
J
/∆T
1D
has other important
indications as will be explained in a later section. To determine θ
J
, we need to realize the
continuous flow of heat at the junction of the wire and the via, which can be shown as the
continuity equation
ILD
t y
v
v v L x
m
m m
dy
y d
A k
dx
x d
A k
= ± =
= −
) ( ) (
2 /
θ θ
. (2.17)
Substituting Eq. 2.15 and Eq. 2.16 into Eq. 2.17, θ
J
can then be solved as
) / coth( ) 2 / tanh(
) 2 tanh(
,
,
,
,
,
' ' '
,
v H ILD
v H
v v
m H
m H
m m
m H m m H m
J
L t
L
A k
L L
L
A k
L L q L A
+
= θ
) / coth( ) 2 / tanh(
)] ( csc ) [coth(
,
,
,
,
, ,
' ' '
,
v H ILD
v H
v v
m H
m H
m m
v H ILD v H ILD v v H v
L t
L
A k
L L
L
A k
L t h L t q L A
+
−
+
. (2.18)
Effect of variation in electrical resistivity ρ with temperature was found to be small
for practical situations and is ignored in this model. It should be noted that we should
certainly include the temperature coefficient of resistivity (TCR) for electromigration
tests where very high current densities are employed. It is simple to do so because
Section: 2.3 Interconnect Thermal Model and Assumptions
31
electromigration tests are usually conducted on isolated wires and not multilevel struc
tures. Several papers have been devoted to study the case under electromigration test
condition [36, 37]. The focus of our work, however, is to incorporate the via effect and to
predict temperature distributions along multilevel interconnects under normal operation
conditions. Therefore, the resistivity used in this work is assumed to be a constant at
ρ(T
Die
) to a first order approximation.
The last pieces of the puzzle in this model are the shape factors S
m
and S
v
, which
accommodate 3D heat spreading and different boundary conditions. The commonly used
Bilotti’s equation [29] is not adopted in this work to account for the deviations from 1D
heat flow. This is due to the fact that it assumes a single heat source, whereas, in typical
IC layout, there are multiple heat sources due to parallel metal wire array. A new expres
sion of heat spreading factor, S’
m,
is therefore derived here for uniformly separated
infinite number of parallel wires and will be used in the following analysis throughout
this work. For the worst case scenario, all metal wires are assumed to carry the maxi
mum RMS current density and separated by spacing d. As shown in Fig. 2.8, the Joule
heat transfers downward as well as spreads laterally in the ILD. Assuming the lateral
spreading to increase linearly with vertical coordinate, the spreading thermal resistance,
R
spr
, can be derived as

.

\

+
=
+
=
∫
w
d w
k y w
dy
k
R
ILD
d
ILD
spr
ln
2
1
2
1
2
0
. (2.19)
Chapter 2: Analytical Interconnect Thermal Model
32
where d is the spacing between wires and is shared by the two wires for heat dissipation.
Then, the total thermal resistance of ILD, R
th,ILD
, can be calculated by combing the
spreading resistance with the volume resistance,
d w
d
t
k d
d w
k
R
ILD
ILD ILD
ILD th
+
−
+ 
.

\
 +
=
2
1
ln
2
1
,
. (2.20)
t
ILD
w
d
Metal Wires
d/2
heat
dissipation
path
t
ILD
Rrect
R
spr
d w
d/2
d/2
heat
spreading
1D Model Quasi 2D
t
ILD
w
d
Metal Wires
d/2
heat
dissipation
path
t
ILD
w
d
Metal Wires
d/2
heat
dissipation
path
t
ILD
t
ILD
Rrect
R
spr
d w
d/2
d/2
heat
spreading
1D Model Quasi 2D 1D Model Quasi 2D
Fig. 2.8: Geometry used for calculating R
spr
, R
rect
and the spreading factor S’
m
. Cross
sectional view is shown on the left hand side. Quasi 2D heat conduction is
used to correct 1D heat conduction model. Space between any two wires is
shared for heat dissipation.
Section: 2.3 Interconnect Thermal Model and Assumptions
33
On the other hand, R
th
,
ILD
can be also expressed as
m ILD
ILD
effective ILD
ILD
ILD th
wS k
t
w k
t
R
'
,
= = . (2.21)
By comparing Eq. 2.20 and Eq. 2.21, , S’
m
can be obtained as
1
2
2
1
'
−


.

\

+
−
+ 
.

\
 +
=
d w
d
t
t
w
w
d w
t
w
S
ILD
ILD ILD
m
. (2.22)
The next step is to convert spreading factor S’
m
to shape factor S
m.
By the definition of
shape factor [38],
ILD th
m
m ILD m
R
x
x k S x q
,
) (
) ( ) (
θ
θ = = (2.23)
where q(x) is the heat transfer rate from the metal wire to the dielectric medium. Hence, ,
R
th
,
ILD
can be manipulated as,
ILD m
ILD th
k S
R
1
,
= (2.24)
Again, by comparing Eq. 2.21 and Eq. 2.24, the conversion between S
m
and S’
m
can be
readily obtained as,
m
ILD
m
S
t
w
S ' ) ( = (2.25)
and thus,
1
]
2 /
ln
2
1
[
−


.

\

+
−
+ 
.

\
 +
=
d w
d t
w
d w
S
ILD
m
(2.26)
Chapter 2: Analytical Interconnect Thermal Model
34
It should be noted that, depending on different layout and operating conditions, S
m
could
have different expressions than Eq. 2.26. However, all the equations derived here can
still be valid as long as the appropriate S
m
is determined by either an appropriate analyti
cal expression or extracted from simulation. The S
v
is adopted from [38] for a vertical
cylinder in a semiinfinite medium attached to a constant temperature surface,
)] / 4 /[ln( 2 D t S
ILD v
π = . (2.27)
After the shape factors are installed in Eq. 2.11 and Eq. 2.12, the effect of the via separa
Polymer, Analytical
Polymer, ANSYS
Oxide, ANSYS
Oxide, Analytical
Polymer, Analytical
Polymer, ANSYS
Oxide, ANSYS
Oxide, Analytical
Fig. 2.9: Temperature profile along the Cu wires with 100 µm via separation.
H = t
ILD
= 0.8 µm, w = d = 0.3 µm, and j
rms,m
= 3.7E6 A/cm
2
. k
polymer
=
0.3 W/mK and k
oxide
=1.2 W/mK are assumed here.
Section: 2.4 Impact of Via Effect on Effective kILD
35
tion and heat spreading on the temperature profile along metal wire can be captured
completely by Eq. 2.15. As can be observed form Fig. 2.9, the result from analytical
expression is shown to be within 3% agreement with the 3D finite element thermal
simulation using ANSYS.
2.4 Impact of Via Effect on Effective k
ILD
2.4.1 Via correction factor
It should be noted that as predicted from Eq. 2.11 and validated from Fig. 2.9, the
thermal healing length, L
H,m
, in a wire is longer if the ILD has lower thermal conductivity,
(k
polymer
= 0.3 W/mK v.s. k
oxide
=1.2 W/mK). Consequently, the via effect is more impor
tant for lowk insulators. By defining a via correction factor (η), the via effect can be
incorporated into the effective thermal conductivity of ILD, k
ILD,eff
, which can then be
used in place of the nominal k
ILD
in the conventional thermal equations. An analytical
expression for k
ILD,eff
incorporating the via effect is now derived here. ∆T
ave
is defined as
the average temperature rise in one metal layer and it can be expressed as
m eff ILD
m m rms ILD th m elec m ILD th m ave
S k
Hw
j R R I R q T
,
2
, , ,
2
,
) ( ρ = = = ∆ . (2.28)
On the other hand, ∆T
ave
can also be obtained from Eq. 2.15 as
( ) ( )
∫ ∫ − −
= − = = ∆
2
2
2
2
0 ,
) (
1 1
L
L
m
L
L
ave m ave
dx x
L
dx T x T
L
T θ θ
Chapter 2: Analytical Interconnect Thermal Model
36
(
(
(
(
(
¸
(
¸


.

\

− +


.

\

= )
2 /
2 /
tanh
1 (
2
2
tanh
) (
,
,
,
,
2
,
2
,
m H
m H
m H
m H
m ILD
m m rms
J
m ILD
m m rms
L
L
L
L
L
L
L
L
S k
Hw j S k
Hw
j
ρ
θ
ρ
(2.29)
Comparing Eq. 2.28 and Eq. 2.29, via correction factor, η, can be deduced as
( )
m H
m H
D
J
L L
L L
T
,
,
1
2 /
2 / tanh
) 1 ( 1
−
∆
− − =
θ
η , (2.30)
which yields two important results.
η /
, ILD eff ILD
k k = (2.31)
and η
D ave
T T
−
∆ = ∆
1
. (2.32)
2.4.2 Hot spot location
As implied by Eq. 2.31 and Eq. 2.32, via effect will help increase the effective
thermal conductivity of the ILD and, therefore, alleviate interconnect temperature rise if
η<1. On the other hand, if via self heating and thermal impedance caused by the via is
excessive, η>1 may occur and via effect will be detrimental to the wire. In addition, hot
spots can occur within the vias if η>1. The essential criterion to differentiate these two
domains is to evaluate the ratio of θ
J
/∆T
1D
, as indicated by Eq. 2.30. With θ
J
/∆T
1D
,<1, a
beneficial via effect is guaranteed. This critical condition can be further shown to be
1 ) cosh( ) )( )( ( 1
,
2
,
,
' ' '
' ' '
<
(
(
¸
(
¸
−
v H
ILD
v H
m H
m
v
v
m
L
t
L
L
k
k
q
q
, (2.33)
or
1 ) cosh( ) )( )( )( ( 1
,
2
,
2
,
<
(
(
¸
(
¸
−
v H
ILD
v
m
m
v
v
m
v rms
m rms
L
t
A
A
S
S
j
j
ρ
ρ
. (2.34)
Section: 2.4 Impact of Via Effect on Effective kILD
37
Fig. 2.10: Via correction factor, η, with two different D/W ratios, are plotted for three
types of dielectrics. Dielectrics with lower nominal k
ILD
experience lower η
even with longer via separation. Wires with shorter via separation feel
stronger via effect.
Fig. 2.11: Effective ILD thermal conductivity increases with decreasing via separation.
The lower the nominal k
ILD
, the longer the L
H,m
, and hence, the stronger is the via
effect.
Chapter 2: Analytical Interconnect Thermal Model
38
From Eq. 2.33 and Eq. 2.34, various interconnect parameters can be quickly evaluated to
determine the hot spot location (in the wire or within via) and the nature of via effect
(alleviate or enhance wire temperature rise). To illustrate the effect of the via, η and
k
ILD,eff
are plotted against via separation with two via diametertowire width ratios, D/w,
in Fig. 2.10 and Fig. 2.11, respectively. Geometries of the interconnect structure are
taken from ITRS [5] 65nm technology node for global wires. It can be observed that,
under the condition η<1, incorporation of the via effect results in increased k
ILD,eff
espe
cially for ILD materials with lower nominal thermal conductivity. This fact can explain
why the interconnect temperature is not as high as commonly assumed when lowk ILD
is implemented in the advanced interconnect structure. Fig. 2.11 also shows that the via
effect diminishes rapidly for the portion of wire beyond thermal healing length from each
end, the L
H,m
’s for oxide, polymer and air are about 5µm, 10um and 30um as calculated
from Eq. 2.11. On the other hand, as shown in Fig. 2.10, for the case of D/w=0.1, the
cross section of the via is much smaller than that of the wire which results in substantial
via heating. The temperature profile along the interconnect is shown in Fig. 2.12. In this
case, short via separation suffers more heat backflow from the vias. However, in the
typical VLSI interconnect structure, D/w is generally greater than 0.5, so that the hot spot
locates in the middle of the wire and the via effect is always beneficial.
Section: 2.5 Temperature in Multilevel Metal Layers
39
2.5 Temperature in Multilevel Metal Layers
2.5.1 Multilevel interconnect formula
Following the previous argument, Joule heat generated in metal wires is considered
to dissipate only through the heat sink attached to the Si substrate. This assumption is
fairly legitimate due to the fact that the chip is usually encapsulated with insulation
materials. Therefore, all the heat generated in the upper metal levels has to transfer
through the lower metal levels to the substrate. With ∆T
i1, i
defined as the average
0
0.5
1
1.5
2
2.5
3
0
10 20 30 40 50
D/W = 0.1
D/W = 0.2
D/W = 0.3
D/W = 0.5
D/W = 0.7
D/W = 1
Location along Interconnect [um]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
[
°
C
]
via via
0
0.5
1
1.5
2
2.5
3
0
10 20 30 40 50
D/W = 0.1
D/W = 0.2
D/W = 0.3
D/W = 0.5
D/W = 0.7
D/W = 1
Location along Interconnect [um]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
[
°
C
]
0
0.5
1
1.5
2
2.5
3
0
10 20 30 40 50
D/W = 0.1
D/W = 0.2
D/W = 0.3
D/W = 0.5
D/W = 0.7
D/W = 1
Location along Interconnect [um]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
[
°
C
]
via via
Fig. 2.12: The figure shows the temperature profile along the wire from the middle of
the interconnect to the via at the right end. The hot spots usually locate in
the wire except for extremely small via dimensions.
Chapter 2: Analytical Interconnect Thermal Model
40
temperature rise between metal layers i1 and i, the temperature rise at the top layer for
an Nlevel metal interconnect can be obtained as
i
N
i
i th
N
i
i i substrate N N
Q R T T T T
∑ ∑
= =
−
= ∆ = − = ∆
1
,
1
, 1
(2.35)
j j j m
N
i j
j rms
N
i i i m i ILD
i i
H j
S k
w
α ρ
α
η
,
2
,
1 , ,
∑ ∑
= =
≅ (2.36)
where Q
i,
is the accumulated heat dissipated through the i
th
level interconnects and α
i
is
the metal coverage of the i
th
level metal layer. For the case when the via effect is ne
glected, η
i
is set to be 1. As can been seen from Eq. 2.35, there is more heat flowing
through the lower levels since Q
i
represents the sum of all the heat generated from i
th
layer to N
th
layer. As a result, a substantial temperature rise will occur in local wires if
the effect of the dense via population is not taken into account. For the following demon
stration of the importance of the via effect, some reasonable values of via separation are
assigned to each of the 10 metal layers with polymer used as the ILD in a 65 nm technol
ogy node structure. In addition, a worst case current density, j
rms
, of 2.1e6 A/cm
2
is
assumed for all wires and 50% metal coverage is assumed for each metal layer. First, as
shown in Fig. 2.13, the overall temperature rise is much lower with the help of the vias.
Second, it can be observed that the temperature distribution among metal layers is quite
different in these two cases. Ignoring the via effect results in large temperature jump in
the lower layers and then the temperature rise levels off. On the other hand, with the via
effect considered, there is hardly any temperature rise in the lower levels even when
Section: 2.5 Temperature in Multilevel Metal Layers
41
k
polymer
is only one fourth of k
oxide
. Most of the temperature rise is attributed to the upper
metal layers with long via separation. Therefore, from the thermal design point of view,
global interconnects are more problematic . The concern of RC delay in the global wires
may get worse with this additional temperature effect.
2.5.2 Interconnect temperature rise trend
The trend of interconnect temperature rise due to Joule heating is investigated in Fig.
2.14, curve (a). It is interesting to see that even with increasing current density and lower
thermal conductivity from the lowk ILDs used for the more advanced technology nodes,
the interconnect temperature rise will reach a plateau and then drop. This phenomena can
Metal Layer Number
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
∆
T
[
°
C
]
0
5
10
15
20
25
30
35
40
45
1 2 3 4 5 6 7 8 9 10
via effect ignored
via effect included
Metal Layer Number
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
∆
T
[
°
C
]
0
5
10
15
20
25
30
35
40
45
1 2 3 4 5 6 7 8 9 10
via effect ignored
via effect included
Fig. 2.13: Temperature rise distribution along metal layers from substrate to top metal
level. In the case of via effect included, the via separations assigned to the
metal layers, from 1
st
to 10
th
levels, are 5, 10, 30, 50, 100, 150, 200, 300,
500 and 1000 metal pitches,respectively, based on 65nm technology node.
Chapter 2: Analytical Interconnect Thermal Model
42
be understood by realizing that the via separation keeps reducing due to scaling. The
same functionality can be reached within a shorter distance. On the other hand, the
magnitude of current density and the dielectric constant of the lowk insulators are
believed to encounter a bottleneck due to reliability concern at 65nm technology node.
Several scenarios are studied here to identify their impact on interconnect temperature, as
shown in Fig. 2.14. Overall, the interconnect temperature rise will be less if either the
trend of increasing current density or lowerk materials stops at 65nm nodes, while all the
other aspects of scaling continue.
Technology Node [nm]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
∆
T
[
°
C
]
130 90 65 45 22
0
5
10
15
20
25
30
scenario (a)
scenario (b)
scenario (c)
scenario (d)
Technology Node [nm]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
∆
T
[
°
C
]
130 90 65 45 22
0
5
10
15
20
25
30
scenario (a)
scenario (b)
scenario (c)
scenario (d)
Fig. 2.14: Trend of interconnect temperature rise along technology nodes for different
scenarios. (a) both j and k
ILD
scale according to ITRS. (b) j scales with
ITRS, but k
ILD
stops scaling at 65nm node. (c) j stops scaling at 65nm
node, while k
ILD
continues scaling. (d) both j and k
ILD
stops
scaling at 65
nm node
Section: 2.6 Summary
43
2.6 Summary
In conclusion, a compact analytical thermal model has been presented to evaluate the
spatial thermal distribution and the average interconnect temperature rise under the
influence of vias. Both the via effect and heat spreading have been taken into account to
ensure accurate predictions. This model provides a quick and accurate interconnect
temperature rise estimation as well as a comprehensive thermal design guideline. It will
be essential for high performance chip designer being able to comprehend the thermal
impact on the circuit performance and reliability implication in the early design stage. It
has been shown that with the help of vias as efficient thermal paths, the effective thermal
conductivity of the ILD materials can be significantly higher than their nominal values if
via separation is comparable to the characteristic thermal length. The interconnect
temperature can be substantially lower than that predicted from overly simplified 1D
thermal model. Therefore, the thermal problem associated with lowk insulators is not as
bad as it appears.
Additionally, a closed form thermal model incorporating the via effect has been for
mulated to estimate the temperature rise of interconnects in multilevel metal arrays. It is
shown that via effect must be considered in the thermal analysis of interconnect struc
tures. It is observed that global interconnects would suffer much higher temperature rises
than local interconnects due to the much longer via separation. Beyond the 45nm node
closer packing of vias will alleviate the temperature rise problem.
45
Chapter 3
Thermal Impact on Interconnect Design
3.1 Introduction
Thermal effects are very important in determining both reliability and performance
of interconnects. Accurate integrity estimation of onchip interconnect temperature rises
caused by joule heating is essential for high performance chip design because the wire
currentcarrying capability is stringently limited by interconnect temperature. Thermal
effects impact the interconnect design in the following ways. First, the wire slows down
with higher temperature due to increasing metal electrical resistivity. For the temperature
coefficient of resistivity (TCR) of Cu given in [39], wire resistance goes up by 510% as
interconnect temperature rises by 1020°C. As a result, the delay reduction expected
from the introduction of lowk dielectrics, which have invariably poor thermal conductiv
ity, will be discounted to some extent. Second, most of interconnect failure mechanisms
are temperature related including electromigration (EM). EM lifetime, meantime tofail
Chapter 3: Thermal Impact on Interconnect Design
46
(MTF), decreased exponentially with inverse interconnect temperature according to
Black’s equation [19]. It is reported that MTF could be reduced by 90% when intercon
nect temperature rises from 25°C to 52.5°C [32]. Therefore, traditionally, the current
density design limits developed by reliability engineers for circuit designers are very
conservative. The interconnect temperature is commonly limited to ~105°C for
electromigration life time considerations [40, 41].
Having realized the importance of the temperature rise on interconnects, many ef
forts have been devoted to estimate the interconnect temperature [24, 31] and propose
“selfconsistent” analysis for allowed current density [30, 42]. However, without consid
ering via effect properly, the estimated interconnect temperature rise can be much higher
than twice as high compared to the realistic case [43, 44]. The temperature and conse
quent performance/reliability predictions with those overly simplified thermal analysis
will deviate even more profoundly with more advanced technology nodes as the via effect
is more effective for lowerk dielectrics and the via separation is shortened with scaling
[45]. Extending the work from Chapter 2, our multilevel interconnect thermal model
incorporating the via effect enables more realistic circuit timing simulation and reliability
assessment without being excessively conservative. In this chapter, the impact of Joule
heating on the scaling of deep submicron Cu/lowk interconnects will be investigated in
detail. [46]. In addition, traditional wisdom requires the onchip interconnect temperature
to be no more than 5°C above the silicon temperature (100°C for a typical high perform
ance microprocessor) for EM lifetime consideration. However, the prospects of rising
Section: Thermal Effect on Interconnect Metrics
47
interconnect temperature and the need for greater current carrying capability are not
compatible. Although it has been proposed to consider EM lifetime and wire selfheating
simultaneously in generating EM guidelines [23, 30, 42], the lack of realistic interconnect
temperature estimation makes the proposals fruitless. In a later section of this chapter,
combined with our new interconnect thermal model, a comprehensive EM evaluation
methodology is proposed from the view of coupled performance and reliability analysis
[45].
The remainder of this chapter is organized as follows. In Section 3.2, we discuss
how the interconnect metrics are influenced by thermal effect. Section 3.3 addresses the
foreseeable impact from interconnect scaling trend on wire temperature due to Joule
heating. Section 3.4 describes the definition of a ‘reasonable’ worstcase scenario of
interconnect thermal analysis and applies it to evaluate the thermal effect on Cu/lowk
interconnects. In Section 3.5, we perform a coupled assessment of EM reliability and
current drivability for global interconnects for 22130 nm technology nodes. Section 3.6
investigates the trend of interconnect temperature rise under various scaling scenarios.
Finally, we summarize and conclude in Section 3.7.
3.2 Thermal Effect on Interconnect Metrics
It suffices to mention here that a nonnegligible and increasing resistance of the
wires leads to a j
2
ρ (j is the current density) power dissipation in the form of heat. This
Chapter 3: Thermal Impact on Interconnect Design
48
raises the temperature of the interconnects above the device temperature, especially if
lower dielectric constant materials, which are also invariably poorer heat conductors, are
used.
3.2.1 RC delay
The delay of the wires can be well approximated by the product of resistance (R)
and the capacitance of the wire (C), if inductive effects are not important. The wire
capacitance typically has three components: interlevel, C
ILD
, intermetal (within the
same level between metal lines), C
IMD
, (Fig. 3.1) and the fringe component. To model
the RC delay of the wires it is imperative to accurately model both the resistance and the
capacitance accurately.
H
W
C
ILD
C
IMD
AR=H/W
Fig. 3.1: Schematic showing the intermetal and the interlevel components of capaci
tance. Also showing aspect ratio (AR).
Section: Thermal Effect on Interconnect Metrics
49
It is well known that the electrical resistivity, ρ
m
,is a function of wire temperature,
and can be modeled as,
)] ( 1 )[ ( ) (
0 0
T T T T
m m
− + = β ρ ρ (3.1)
here β being the temperature coefficient of resistivity (TCR), and has the value of 4.3e
3/K for Cu and 4.5e3/K for Al. In terms of interconnect capacitance, to gain better
insight of the role of C
ILD
and C
IMD
, let us ignore any fringe capacitance for this moment.
The total interconnect capacitance per unit length can then be simply expressed as:
C
total
=2(C
ILD
+C
IMD
), (3.2)
where C
ILD
=ε
ILD
/2AR and C
IMD
=ε
IMD
AR, (3.3)
with AR denotes wire aspect ratio, defined by H/w, where H is the metal thickness
(height) and w is the width of the wire. ε represents the respective premittivity for IMD
and ILD. Thickness of the ILD is approximately the same as the height of the metal wire.
The factor of 2 in the denominator for C
ILD
accounts for the overlap with orthogonal
wires on adjacent levels. The length of overlap is taken to be half the length of the
interconnect based on the assumption that wire width is half the pitch. First, we notice
that, due to the high wire AR, the intralevel (linetoline) capacitance, C
IMD
,
is dominant
.
Second, since most of the heat is flowing downwards toward the heat sink, the ILD
assumes the major contribution to the thermal impedance and IMD being a minimal
factor. Therefore, heterogeneous dielectric schemes (with different dielectric materials
for ILD and IMD), should be exploited to optimize RC delay and subside thermal effect,
such as use lowk dielectrics for IMD and keep silicon dioxide for ILD.
Chapter 3: Thermal Impact on Interconnect Design
50
3.2.2 Dynamic power consumption
The second metric of importance, the power dissipation due to interconnects, is a
result of charging and discharging its capacitance and is given by the dynamic power
dissipation formula
f V C S P
2
int w int
= (3.4)
Here, S
w
is the switching activity factor representing the probability of a particular
interconnect switching during a clock cycle, C
int
is the total interconnect capacitance, V is
the voltage to which the interconnect charges and f is clock frequency. Thus, at a given
technology node, the interconnect power is heavily dependent on its total capacitance.
3.2.3 Cross talk noise
. Cross talk is proportional to the ratio of the intermetal to the total capacitance [47]
of the wire. Thus, from the cross talk perspective it is more beneficial to lower just the
intermetal capacitance. Hence, again, an heterogeneous dielectric approach will be
appropriate to lower coupling noise while contain interconnect temperature rise.
3.2.4 Electromigration reliablity
Electromigration is wildly regarded a major failure mechanism of VLSI intercon
nects[28, 48]. The current through metal wires leads to metal atom migration due to
momentum exchange between electrons and metal atoms [19]. This migration, over time,
leads to depletion of enough material so as to initially increase the wire resistance and
finally cause an open circuit [49]. On the other hand, it also causes excess metal atoms to
Section: Thermal Effect on Interconnect Metrics
51
accumulate at a different location along the wire, which in extreme cases can cause a
short to the adjacent wire through metal hillocks. The primary factors, which influence
electromigration can be divided into those related to the physical structure of the metal
wire and those related to the conditions of operation. Certain crystal orientations of thin
films (metal wire) are more conducive to preventing electromigration. Among the
conditions of operation, temperature and the current density play the most important role
in dictating electromigration, as modeled by the well known Black’s equation, given by
Eq. 3.5,
) / exp( T k Q j A MTF
B
n −
= (3.5)
where MTF is the meattimetofailure (typically for 0.1% of accumulative failure or 10
years lifetime), A is a constant that is dependent on the geometry and microstructure of
the interconnect, j is the DC or average current density, the exponent n is typically 2
under normal operation conductions. The activation energy Q for Cu has been reported
in the range of 0.51eV [5052], k
B
is the Boltzmann’s constant, and T
m
is the metal wire
temperature. For a middle of the range Q value of 0.75eV, MTF will drop by 50% when
interconnect temperature rises from 100°C to 110°C. Usually, the maximum allowable
current density is limited by the goal that the electromigration lifetime will achieve 10
years with interconnect temperature maintained at or below 105°C.
Chapter 3: Thermal Impact on Interconnect Design
52
3.3 Scaling Trend of Joule Heating
As VLSI technology advances, interconnects have become the limiting factor to IC
chip performance [3]. Aggressive interconnect scaling has resulted in increasing current
density, more metal levels, and introduction of low dielectric constant (lowk) materials.
The growing demand of higher current driving capability and the aggressive shrinking of
metal pitch has resulted in significant current density rise in the wires (Table 3.1). Wire
width is half the wire pitch for all cases if not defined specifically, n is the number of
layers in each tier, ε
r
,
ILD
and k
ILD
are the relative dielectric constant and thermal conduc
tivity of the interlevel dielectrics (ILD) for each technology node.
So explosive has the growth been, according to ITRS [5], the rise of interconnect
current density will outpace the average chip power density by a factor of two for high
performance microprocessor throughout the technology nodes (Fig. 3.2).
Local Tier Semiglobal Tier Global Tier
Tech.
Node
ε
r
,
ILD
k
ILD
[W/mK]
n
Wire
Pitch
[nm]
Wire
AR
Via
AR
n
Wire
Pitch
[nm]
Wire
AR
Via
AR
n
Wire
Pitch
[nm]
Wire
AR
Via
AR
130 3.3 0.7 2 350 1.6 1.6 4 450 1.6 1.4 2 670 2.0 1.8
90 2.8 0.45 2 210 1.7 1.7 4 265 1.7 1.5 3 460 2.1 1.9
65 2.5 0.36 2 150 1.7 1.7 4 195 1.8 1.6 4 290 2.2 2.0
45 2.1 0.25 2 105 1.8 1.8 4 135 1.8 1.6 4 205 2.3 2.1
22 1.8 0.17 2 50 2.0 2.0 4 65 2.0 1.8 5 100 2.5 2.3
Table 3.1: Interconnect parameters for 130 nm to 22nm technology nodes based on ITRS[5].
Section: 3.3 Scaling Trend of Joule Heating
53
Furthermore, Cu resistivity will increase due to barriers, surface scattering and skin
effect [53, 54]. As a result, interconnect joule heating is becoming nonnegligible. In
addition, thermal conductivity of lowk dielectrics is decreasing rapidly with the reduc
tion of dielectric constant (Table 3.1). The combination of greater heat generation and
thermal impedance is leading to a continuous increasing interconnect temperature and
consequent impact on wire delay and reliability wires is fast emerging as an urgent issue.
Therefore, the impact of the interconnect “thermal”scaling trend on wire temperature due
to Joule heating requires immediate attention. The work presented in this chapter is for
the 65nm technology node with all parameters quoted from the newly updated ITRS
’01[5]. We assign the first two metal layers as local tier, the following four layers as
semiglobal tier and the remaining layers as global tier. The substrate temperature, T
ref
, is
22 45 65 90 130 180
0
5
10
15
20
25
30
0
1
2
3
4
5
Technology Node [nm]
P
o
w
e
r
D
e
n
s
i
t
y
[
W
/
c
m
2
]
J
m
a
x
[
M
A
/
c
m
2
]
22 45 65 90 130 180
0
5
10
15
20
25
30
0
1
2
3
4
5
Technology Node [nm]
P
o
w
e
r
D
e
n
s
i
t
y
[
W
/
c
m
2
]
J
m
a
x
[
M
A
/
c
m
2
]
Fig. 3.2: Trends of chip power density and interconnect J
max
along technology
nodes suggested by ITRS [5]. Chip power density is calculated by total
power of the chip divided by chip size.
Chapter 3: Thermal Impact on Interconnect Design
54
assumed to be fixed at 100°C (which will be true if twophase microchannel cooler is
employed in the future [55] ), and the temperature of the top global wire is T
m
=
T
ref
+∆T
joule heating
.
3.4 Temperature Effect on Cu/lowk Interconnects
3.4.1 Definition of worst case condition
To provide robust thermal analysis for interconnects, it is important to identify a rea
sonable worst case scenario. Previous work has attempted to evaluate the thermal
characteristics of interconnects neglecting vias as an effective heat conduction path [4].
However, ignoring vias in heat transfer predicts unrealistically high temperature rises
even with moderate current densities. This is especially true for dielectrics with lower
thermal conductivities [44, 46]. Therefore, this work includes the via effect in the
analysis to obtain more realistic results. Case (1) in Fig. 3.3 shows the widely used
condition, i.e., all wires flowing the same current density and the via effect ignored. Case
(2) represents an isolated global wire. Although there is no additional heat source be
tween this wire and the substrate, the thermal impedance is higher in the absence of lower
level metal. Case (3) has the same current condition as case (1) but with vias taken into
account for all metal layers. Reasonable via separations are assigned for each metal level
from level one to ten: 5, 20, 50, 100, 150, 200, 300 ,500, 1000 and 2500 interconnect
pitches. As shown in Fig. 3.4, case (1) has the highest temperature rise (∆T
joule heating
).
Section: Temperature Effect on Cu/lowk Interconnects
55
But we will disregard this condition as unrealistic as vias are ignored. Case (3) is worse
than case (2) and we will use this condition for the following analysis.
Case (1) Case (2)
Case (3)
vias
Case (1) Case (2)
Case (3)
vias
Case (3)
vias
Fig. 3.3: Configurations of the three thermally worst case scenarios. Current flows
through all wires at all metal layers in cases (1) and (3).
0
20
40
60
80
100
120
140
160
180
0 1 2 3 4
case (1)
case (2)
case (3)
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
[
°
C
]
Current Density , J [MA/cm
2
]
Fig. 3.4: Temperature on the toplevel interconnects rises rapidly with increasing
current density.
Chapter 3: Thermal Impact on Interconnect Design
56
3.4.1 Thermal impact on Cu/lowk Interconnects
Fig.3.5 compares the thermal performance of several lowk materials. The materials
properties are shown in Table 3.2. As the figure shows, interconnects with lower dielec
tric constant materials exhibit significantly higher temperature rises, thus higher
Conductor Dielectrics Parameter
Cu Al W Air Aerogel Polyimide HSQ FSG Oxide
k [W/mK] 400 240 180 0.03 0.17 0.3 0.54 1.05 1.2
ρ [µΩcm] 2.2 3.6 10      
εr    1.0 1.7 2.5 3.1 3.7 4.0
Table 3.2: Materials properties of various conductors and dielectrics used in this study. k
[W/mK] is thermal conductivity, ρ [µΩcm] is electrical resistivity near
100ºC and ε
r
is the relative dielectric constant
Current Density , J [MA/cm
2
]
M
e
t
a
l
T
e
m
p
e
r
a
t
u
r
e
T
m
[
°
C
]
10 0
15 0
20 0
25 0
30 0
0. 5 1 1. 5 2 2 .5 3 3. 5 4
ai r (!
r
= 1. 0; k= 0. 0)
aerogel ( !
r
=1. 7; k=0. 17)
po l yi mide (!
r
=2. 2; k= 0. 3)
HS Q(!
r
=3 .1; k =0 .54)
FSG (!
r
=3 .75; k=1. 05)
ai rgap (!
r
= 2. 2; k=1. 2)
Fig. 3.5: Temperature of top global interconnects rises sharply for Lowk dielec
trics. ε
r
: relative dielectric constant, k: thermal conductivity [W/Km].
Section: Temperature Effect on Cu/lowk Interconnects
57
interconnect temperature, T
m
. The product of R
met
C
total
per unit length
for these dielec
trics is shown in Fig. 3.6 as a function of current density. As the figure shows, the lowk
advantage of reduced capacitance can be offset by the increased temperature rise due to
poor thermal conductivity. It should be noted that for the case of airgap scheme (ILD:
SiO
2
and IMD: Air), the R
met
C
total
is relatively constant through the range of current
density and it is even better than polyimide. This is because airgaps reduce the dominat
ing linetoline capacitance while leaving the SiO
2
ILD intact for better thermal
conductivity. Since electromigration mean time to failure (MTF) is exponentially de
20
40
60
80
100
120
140
0.5 1 1.5 2 2.5 3 3.5 4
FSG
H SQ
a i r  g a p
p o l y i mi d e
a e r o g e l
a i r
Cu r ren t Den sit y , J [ MA/cm
2
]
R
m
e
t
C
t
o
t
a
l
[
p
s
e
c
/
m
m
]
Fig. 3.6: RC delay is strong function of current density on the wires because of
Joule heating. The lower the dielectric constant, the stronger the Joule
heating and greater RC degradation.
Chapter 3: Thermal Impact on Interconnect Design
58
pendent on wire temperature, the maximum allowable current density for a given tem
perature rise is evaluated in Fig. 3.7 for different lowk materials.
3.5 Delay and Reliability Optimization
With the possibility that much fatter wires may be used in the global tier to reduce
IR drop, it is important to know the thermal impact of using low aspect ratio lines on the
expected performance and reliability. In addition, ILD thickness should be evaluated to
assess the tradeoffs of thermal impedance and capacitance. The contours of constant
R
met
C
total
, in Fig.3.8 show the optimization of the wire aspect ratio (AR) and the ILD
0
1
2
3
4
5
6
1 2 3 4
delt a T=50 C
delt a T=30 C
delt a T=20 C
delt a T=10 C
delt a T=5 C
C
u
r
r
e
n
t
D
e
n
s
i
t
y
[
M
A
/
c
m
2
]
Dielectric Constant ε
r
Fig. 3.7: Maximum allowed current density is limited by the maximum allowed ∆T
on the metal wires. The constraint is more stringent for lowk dielectrics.
Section: 3.5 Delay and Reliability Optimization
59
thickness for delay consideration. The value of H/S (=0.32µm/0.15µm) is fixed, where H
is metal height and S is wire spacing. It demonstrates that Joule heating can increase
delay of the interconnect by as much as 15%. Fig. 3.9 shows constant MTF contours,
with the temperature effect included. The MTF can never reach 50% of MTF at 105°C
which is the wire temperature specified in ITRS. We can conclude from Fig. 3.8 and 3.9
that although fat wires can provide better speed performance, but they are also more
subject to electromigration failure since the temperature is higher.
0
0.5
1
1.5
2
2.5
3
0.1 0.2 0.3 0.4 0.5
100ps
80ps
60ps
40ps
20ps
G
l
o
b
a
l
W
i
r
e
A
s
p
e
c
t
R
a
t
i
o
(
A
R
)
ILD thickness [µm]
Fig. 3.8: Constant R
met
C
total
contour plots of global level wiring as functions of wire
aspect ratio (AR) and ILD thickness. Solid curves represent the case ther
mal effect neglected. Dashline curves include the influence of thermal
effect. (H/S=0.32/0.15 µm).
Chapter 3: Thermal Impact on Interconnect Design
60
3.6 Impact of Joule Heating on Scaling Trend
Table 3.3 shows the coupled analysis of delay and electromigration MTF for vari
ous technology nodes. Column 2 shows the maximum current density specified in ITRS
and column 3 shows the resultant temperature of the top global wires. The R
met
C
total
delay
is shown in column 4 and the corresponding MTF is shown in Fig. 3.10 as the solid
curve. The achievable MTF is about 90% at 130nm node, but drops sharply for the
following technology nodes. The MTF(T
m
) is compared to the MTF(105°C). To confine
the wire temperature at 105°C, the current density has to be reduced as shown in column
0
0.5
1
1.5
2
2.5
3
0.1 0.2 0.3 0.4 0.5
0.45
0.4
0.35
0.3
0.25
0.2
MTF/MTF(Tm=105°C)
ILD thickness [µm]
G
l
o
b
a
l
W
i
r
e
A
s
p
e
c
t
R
a
t
i
o
(
A
R
)
Fig. 3.9: Constant normalized electromigration MTF, MTF(Tm)/MTF(105°C),
contour plots of global level wiring as functions of wire aspect ratio
(AR) and ILD thickness. (H/S=0.32/0.15 µm).
Section: 3.6 Impact of Joule Heating on Scaling Trend
61
5 and resultant current drive, compared to the maximum current density specified in
ITRS, drops drastically as can be observed in column 6. To relax the temperature rise to
10°C above substrate, the current drive can be improved as shown in column 7. We notice
that in both cases trying to match certain specific wire temperatures (105°C and 110°C),
the corresponding MTF rises above the required MTF, as shown in Fig.3.10. It implies
that this approach is overly conservative. Since the MTF of electromigration is also
inversely proportional to the square of current density, as modeled by the wellknown
Black’s equation, we can achieve the expected MTF by optimizing both the current
density and wire temperature, which are related by Joule heating. The result is shown in
the last three columns of the table. The optimal current density is much closer to the
value specified in ITRS. On the other hand, although the wire temperature is higher than
105°C, as shown in the last column, the MTF is on target, as shown in the flat line in
Fig.3.10.
Table 3.3: Coupled evaluation of electromigration reliability and performance for global
interconnects for 22130 nm technology nodes. T
m
is the top global wire temperature
with all the heat sources, including substrate and Joule heating from all metal levels un
derneath, taken into account.
Chapter 3: Thermal Impact on Interconnect Design
62
3.7 Summary
In conclusion, a detailed analysis of the impact of Joule heating on the characteris
tics of future Cu/lowk interconnects is presented using a realistic full chip model with
via effect included. Thermal effects can severely degrade both reliability and speed
performance. Optimization with various interconnect parameters is provided. Joule
heating will limit scaling of current density and use of lowk materials. Global wires will
be more problematic with higher operating temperature and careful consideration is
imperative.
0
20
40
60
80
100
0
1
2
3
4
5
6
J max , I T RS
J max , T m=105 C
J max , T m=110 C
J max , M T F=100%
22 45 65 90 130
M
T
F
a
c
h
i
e
v
e
d
(
%
)
M
T
F
a
c
h
i
e
v
e
d
(
x
)
Technology Node
Fig. 3.10: The solid curve shows the MTF (%) that can be achieved under the
current density specified in ITRS. The brokenline curves show the
MTF under various wire temperature (Tm) criterions and the values re
fer to the right axis.
63
Chapter 4
SPICEBased ElectroThermal
Simulation Methodology
4.1 Motivation
In Chapter 2, we analyzed the thermal effects in interconnects using analytical
thermal models and demonstrated the strong influence of vias on the temperature
distribution in metal lines. Nevertheless, without solving complex differential equations,
the electrothermal simulation approach is desirable to facilitate quick estimation of
temperature rises and to investigate thermal coupling effects between wires in order to
provide thermal design guidelines for advanced interconnect structures [56, 57].
Teng [58] has used lumped thermal circuit model to predict interconnect
temperature. However, without including thermal capacitance, the model is not capable
of transient analysis. This chapter presents a fast SPICE based 3D electrothermal
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
64
simulation methodology to characterize thermal effects due to Joule heating in high
performance Cu/lowk interconnects under steadystate and transient stress conditions.
The results demonstrate excellent agreement with experimental data and those using
Finite Element (FE) thermal simulations (ANSYS). The effect of vias, as efficient heat
sinking paths to alleviate the temperature rise in the metal wires, is included in our
analysis to provide more accurate and realistic thermal diagnosis. The simulation
methodology has also been applied to quantify the use of dummy thermal vias as
additional heat transfer paths to lower the temperature rise in the metal wires for the first
time. Furthermore, thermal coupling between wires is evaluated and found to be
significant. Finally, the impact of metal wire aspect ratio on interconnect thermal
characteristics is discussed. All the dimensions of the interconnect structure in this paper
were taken from the 100 nm technology node based on the ITRS [59]. The remainder of
this chapter is organized as follows. To serve as a starting point, in Section 4.2, we give
an overview of our thermal simulation methodology. Sections 4.3 & 4.4 describe the
impact of via effect of interconnect thermal characteristics under steadystate and
transient conditions, respectively. Finally, we summarize and conclude in Section 4.5.
4.2 SPICEBased Thermal Modeling
In order to provide robust thermal analysis for deep submicron Cu/lowk
interconnects, it is very desirable to have an efficient 3D simulation methodology to
Section: 4.2 SPICEBased Thermal Modeling
65
estimate the temperature profiles in the metal wires and evaluate the thermal coupling
between them. Based on the thermalelectrical analogy (Fig. 4.1), a 3D RC distributed
thermal circuit model has been developed, as shown in Fig. 4.2. This thermal network
can be easily implemented and simulated using SPICE in the same manner as an
electrical circuit network by simply employing the proper counterparts as illustrated in
Fig. 4.1. For example, the thermal resistance along the metal wire per unit length, R
m
,
and the thermal resistance of the insulator, R
i
, per segment can be computed as,
wH k
R
M
m
l ∆
= (4.1)
Thermal Electrical
Temperature T [K] Voltage V [V]
Heat Q [J] Charge Q [C]
Heat transfer rate q [W] Current i [A]
Thermal resistance R
T
[K/W] Electrical resistance R [V/A]
Thermal capacitance C
T
[J/K] Electrical capacitance C [C/V]
Governing equations
SteadyState condition
Temperature Rise Voltage Difference
Transient condition
Heat diffusion RC transmission line
2
∇
= T R C
T
t
T T
∂
∂
2
∇
= V RC
V
t
∂
∂
T
R q T = ∆ R i V = ∆
Fig. 4.1: ThermalElectrical analogous parameters.
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
66
and
w k
t
R
I
I
i
∆ ∆
=
l
(4.2)
where ∆ℓ is the unit length of metal wire, w is the width of metal wire, H is metal height,
t
I
is the thickness of insulator, and ∆w is the width of the chosen insulator segment. k
M
and k
I
are the thermal conductivity of metal wire and insulator, respectively.
Furthermore, the thermal capacitance along the metal wire per unit length, C
m
, and the
thermal capacitance of the insulator per segment, C
i
, can be shown as,
) (
,
l ∆ = wH c C
M M p m
ρ (4.3)
and ) (
,
w t c C
I I I p I
∆ ∆ = l ρ (4.4)
ILD
Via
Neighbor
Metal Wires
Heated
Metal Wire
M
n
M
n+1
M
n1
IMD
Fig. 4.2: 3D thermal circuit RC transmission line model for transient thermal
analysis of interconnect structures.
Section: 4.2 SPICEBased Thermal Modeling
67
where c
p
,
M,
ρ
M
and c
p
,
I,
ρ
I
are the specific heat and density of metal wire and insulator,
respectively. Inclusion of the lateral thermal impedance in the model captures the heat
spreading effect and thermal coupling from nearby interconnects. Therefore, there is no
need to add any datafitting modification in the circuit model as in [58]. In addition, RC
transmission lines are used to model the heat diffusion due to the similarity of the
governing equations as illustrated in Fig. 4.1 and transient thermal effects in
interconnects can, thus, also be conveniently analyzed. To account for the temperature
dependence of the metal resistivity, the heat generation (q) in each segment of the wire
has been modeled as a voltage (temperature) controlled current source,
) ) ( ( ) (
2
wH
T wH j q
rms
l ∆
= ρ (4.5)
where j
rms
is the rootmeansquare current density flowing through the wire and
ρ(T)(=ρ
0
(1+βT)) is the temperature dependent metal resistivity. ρ
0
is the metal resistivity
at 0°C and β is the temperature coefficient of the resistivity (TCR). The main advantage
of this SPICEbased methodology is that once layout data is available, the thermal
analysis can be quickly done. HSPICE was used for the 3D simulations of the
distributed thermal RC circuits in this work. This technique was validated by comparing
with both experimental data [60] and simulation results carried out by ANSYS, a finite
element (FE) simulation package. It will be shown in the following sections that it
exhibits excellent accuracy within 5% agreement.
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
68
4.3 SteadyState Analysis
4.3.1 Impact of Via Separation on Effective k
ILD
In the steadystate analysis, thermal capacitance can be removed from our RC thermal
circuit model for the obvious reason. The resultant simplified model is shown in Fig. 4.3.
The temperature profiles along the embedded Cu wires with 100µm via separation for
two different dielectrics are shown in Fig. 4.4. Simulated temperature profile along Cu
interconnect with current density J=3.7x10
6
A/cm
2
and via separation of 100 µm.
Current is flowing in alternate wires, i.e. current is not flowing in the nearest, third
S
W 2
V i a
M
n
M
n 1
V
i
a
S
e
p
a
r
a
t
i
o
n
W 1
(a)
(b)
Fig. 4.3: (a) Interconnect configuration. (b) Correspondent distributed thermal circuitry.
Section: 4.3 SteadyState Analysis
69
nearest, fifth nearest and so on neighboring wires but current is flowing in the second
nearest, forth nearest and so on neighboring wires. Geometries used in the simulations
are metal height H=0.8 µm, wire width (w)=wire spacing (S)=0.3µm, and ILD thickness
t
ILD
=0.8 µm. The interconnect geometries used in the simulations are stated in the
caption and the materials properties are listed in Table 4.1. The diameter of the vias is
assumed to be the same as the width of the wire for the damascene process. Since the
0
2
4
6
8
10
12
14
16
18
0 20 40 60 80 100
Polymer, ANSYS
Polymer, HSPICE
SiO2, ANSYS
SiO2, HSPICE
Location along metal wire [µm]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
[
°
C
]
0
2
4
6
8
10
12
14
16
18
0 20 40 60 80 100
Polymer, ANSYS
Polymer, HSPICE
SiO2, ANSYS
SiO2, HSPICE
Location along metal wire [µm]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
[
°
C
]
Fig. 4.4: Validation of HSPICE thermal simulation with Finite Element
Analysis (ANSYS) data.
Conductor Dielectrics Parameters
Cu Al Oxide Polymer Air
k [W/mK] 400 220 1.2 0.3 0.03
ρ [kg/m3] 8933 2720 2220 1380 0.87
Cp [J/kgK] 385 900 745 1195 1014
ρ [µΩcm] @ 0°C 1.67 2.66   
TCR × 103/K 4.3 4.5   
εr   4.0 2.5 1.0
Table 4.1: Materials properties used in this work
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
70
substrate, to which a heat sink is usually attached, is assumed to be the sole heat
dissipation path to the outside ambient, only heat conduction downwards is considered in
this interconnect thermal modeling. Heat convection to the ambient air is ignored by the
application of adiabatic boundary condition on the four side walls and top of the chip.
This is a reasonable assumption because, in general, the chip is enclosed by thermally
insulated package materials. Owing to the lack of available experimental data for such
small spatial resolution under steadystate current, the results from SPICE simulations are
compared against with those using finite element thermal simulation (ANSYS), and it
demonstrates excellent agreement (Fig. 4.4). However, because of the wide familiarity of
SPICE to the circuit design community, this approach, as compared to traditional finite
element simulation, can be adopted more easily and it requires much less effort to set up
as well as less CPU time to run. It can be observed in Fig. 4.4 that the temperature
profiles along the wire for dielectrics with lower thermal conductivity (k
polymer
= 0.3 W/m
K compared to k
SiO2
=1.2 W/mK) is more curved. This is due to the larger ratio between
the thermal conductivity of Cu and the dielectric and hence a stronger influence from vias
is expected. The distance over which via effect is important can be roughly estimated by
thermal healing length L
H,
(=k
M
t
ILD
H
/k
ILD
s)
1/2
[45], which is 14 µm and 30 µm for SiO
2
and polymer respectively, where k
M
, is the metal thermal conductivity, H
is the metal
wire thickness, separated from the underlying layer by ILD of thickness t
ILD
and thermal
conductivity k
ILD
and
s is the heat spreading factor to accommodate the deviation from 1
D heat flow between a metal wire and the underlying layer. It can be thought that within
Section: 4.3 SteadyState Analysis
71
the range of L
H
from vias, heat generated will flow through the vias to the underlying
layer. Beyond L
H
, heat will flow through the ILD and the via effect is diminished. It can
be observed that the lower the nominal k
ILD,
the longer the L
H
and the stronger the via
effect. Since the via separation for the local interconnects is generally much shorter than
L
H
, via effect can not be ignored in local interconnects. Even at the global level, large
portions of the wires are still under the influence of vias. Therefore, the effective thermal
conductivity of lowk dielectrics is considerably higher than the nominal values with the
help of vias, as shown in Fig. 4.5, where the k
ILD,eff
is plotted against via separation for
three different ILD materials. Knowing the temperature profile (e.g. Fig. 4.4), the k
ILD,eff
is obtained by taking the average temperature of the wire and adjusting from the nominal
0.01
0.1
1
10
100
0 100 200 300 400 500
ILD: SiO2
ILD: polymer
ILD: air
Via Separation [µm]
k
I
L
D
,
e
f
f
[
W
/
m
K
]
k
polymer
k
air
k
SiO2
0.01
0.1
1
10
100
0 100 200 300 400 500
ILD: SiO2
ILD: polymer
ILD: air
Via Separation [µm]
k
I
L
D
,
e
f
f
[
W
/
m
K
]
k
polymer
k
air
k
SiO2
Fig. 4.5: Effective ILD thermal conductivity increases with decreasing via
separation. The lower the nominal k
ILD
, the longer the L
H
, and hence, the
stronger is the via effect.
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
72
k
ILD
accordingly
.
It can be observed that incorporation of via effect results in increased
k
ILD,eff
especially for ILD materials with lower nominal thermal conductivity. This fact
can explain why the interconnect temperature is not as high as commonly assumed when
lowk ILD is implemented in advanced interconnect structures and, thus, via effect has to
be included for accurate interconnect temperature estimation. As concluded from Fig. 4.4
and 4.5, significant difference in temperature profiles along the wires and in the
maximum temperature rise can arise between the realistic situation of heat dissipation in
the presence of vias and the overly simplified case that ignores via effect. In addition, it
should be noted that the steeper temperature profile for the lowk dielectrics can
exacerbate electromigration due to the larger temperature gradient.
4.3.2 Thermal Coupling Effects
Thermal coupling between wires in the same layer and different layers has become an
important issue due to the scaling of wire spacing. It is very difficult to solve the
coupling effect analytically. However, our SPICEbased electrothermal methodology
provides a very convenient solution. To investigate the parallel thermal coupling effect
between wires, the configuration shown in Fig. 4.3(a) is examined and symmetrical
boundary conditions applied. The interconnect structure used here is the same as that
used for Fig. 4.4, but with wire width fixed at 0.3 µm, and wire spacing varied from 0.1
µm to 2 µm. Furthermore, polymer is used as the insulation material. In Fig. 4.6(a), the
middle wire is carrying a current density of J
0
(=1.4x10
6
A/cm
2
), which is the maximum
current density specified for the 100 nm technology node from ITRS [58]. The tempera
Section: 4.3 SteadyState Analysis
73
(b)
(
T

T
0
)
/
T
0
(
%
)
50
0
50
100
150
200
0 0.2 0.4 0.6 0.8 1
S=0.1um
S=0.3um
S=0.6um
S=0.8um
S=1.0um
S=2.0um
Current Densit y in Neighboring Wire, J [x J
0
]
(
T

T
0
)
/
T
0
(
%
)
50
0
50
100
150
200
0 0.2 0.4 0.6 0.8 1
S=0.1um
S=0.3um
S=0.6um
S=0.8um
S=1.0um
S=2.0um
Current Densit y in Neighboring Wire, J [x J
0
]
200
0
200
400
600
800
1000
1200
1400
0 0.2 0.4 0.6 0.8 1
S=0.1um
S=0.3um
S=0.6um
S=0.8um
S=1.0um
S=2.0um
(
T

T
0
)
/
T
0
(
%
)
Current Density in Neighboring Wire, J [x J
0
]
200
0
200
400
600
800
1000
1200
1400
0 0.2 0.4 0.6 0.8 1
S=0.1um
S=0.3um
S=0.6um
S=0.8um
S=1.0um
S=2.0um
(
T

T
0
)
/
T
0
(
%
)
Current Density in Neighboring Wire, J [x J
0
]
current
flowing
no current
flowing
(i)
(ii)
S1
S2
(c)
current
flowing
no current
flowing
(i)
(ii)
S1
S2
(c)
wire
(a)
Fig. 4.6: (a) T is defined as the maximum temperature in the middle wire. Reference
temperature T
0
is the maximum temperature in the middle wire with no current
flowing in the nearest neighboring wires and with equal line width and spacing (S).
The current density in the middle wire is J
0,
1.4x10
6
A/cm
2
. Line width is kept at
0.3µm in this simulation. (b) The current density in the middle wire is one third of
J
0
. Line width is 0.3µm. The temperature of wires carrying lower current density
is strongly affected by the heat coupled from the neighboring wires. (c) Top view
of interconnect structures, with different spacing, used in this simulation and
current flow patterns showing current flowing on alternate wires.
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
74
ture variation due to thermal coupling is normalized to T
0
, which is the maximum
temperature in the middle wire when there is no current flowing in the nearest
neighboring wires and with equal wire width and wire spacing, S (=0.3µm in this
simulation). It can be seen that both the current density of the neighboring wire and the
line spacing greatly affect the temperature of the middle wire. In the case of equal line
width and spacing (0.3µm), the temperature of the middle wire can be increased by
nearly 100%. The negative values are due to the symmetric boundary condition chosen
in this simulation and the temperature variation is normalized to T
0,
which is with S equal
to w. Fig. 4.6(c) depicts the top view of two current patterns and can be used to
understand the results. With S1 < S2, more current is flowing in Fig. 4.6(c)(i) for a given
area, thus causing higher thermal coupling effect to the middle wire. The coupling effect
can be even more drastic when the wire is carrying a low current density, which
represents a typical signal line. In this case, only one third of J
0
is flowing in the middle
wire and the thermal coupling effect is shown in Fig. 4.6(b). The strong thermal coupling
may cause resistance variations and further enhance crosstalk problem.
4.4. Transient Stress Analysis
Apart from normal circuit operating conditions, ICs also experience highcurrent
stress events, the most crucial of them being electrostatic discharge (ESD), which causes
accelerates thermal failures [61]. Semiconductor industry survey has indicated that ESD
Section: Transient Stress Analysis
75
is one of the major causes of the failure of ICs [62]. Furthermore, I/O circuitry also
exposes to high current events since it interfaces between multiple power supplies [63].
Therefore, it is important to be able to analyze interconnects heating under transient stress
conditions.
4.4.1 Analytical Model vs. SPICEBased Simulation Methodology
In general, a highcurrent shortpulse (J > 10MA/cm
2
, and pulse width t
pulse
< 200
ns) usually causes much higher ∆T
max
, the maximum temperature rise on the wire,
due to
more severe selfheating than under normal operating conditions and the heat diffusion is
limited to the immediate materials in contact with the metal line. Again, only heat
conduction is considered in this analysis. The governing heat diffusion equation of
temperature rise during brief transients can be written down based on energy
conservation law as [60, 64],
t
T
C
wH
T i
z
T
k
z y
T
k
y x
T
k
x
eff
rms
IMD ILD M
∂
∂
= +
∂
∂
∂
∂
+
∂
∂
∂
∂
+
∂
∂
∂
∂ ) (
) ( ) ( ) (
2
ρ
(4.6)
where
1
] 2 2 [
−
+ + ≅ t H s c t w s c wH c C
IMD l IMD IMD ILD V ILD ILD M M eff
α ρ α ρ ρ
(4.7)
and c
M
, c
ILD
, c
IMD
and ρ
M
, ρ
ILD
, ρ
IMD
are the specific heat and density of metal, ILD and
IMD materials respectively. S
v
and S
l
are shape factors in vertical and lateral directions to
account for threedimensional heat spreading. α
ILD
and α
IMD
are the thermal diffusion
coefficients of ILD and IMD materials respectively. C
eff
can be understood as the
effective thermal capacitance per unit length to absorb the heat. Due to the nature of heat
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
76
diffusion process, C
eff
is increasing with time as heat diffuses through larger volume.
Therefore, temperature rise on the wire under current pulse i
rms
with duration t
pulse
can be
shown as,
eff
C
E
T = ∆ (4.8)
1
2
] 2 2 [
) (
−
+ + ≅ t H s c t w s c wH c
wH
t T i
IMD l IMD IMD ILD V ILD ILD M M
pulse rms
α ρ α ρ ρ
ρ
(4.9)
where E is the pulse energy. It should be noted that, unlike in the case of normal steady
state operation condition, the temperature dependence of ρ(T) absolutely can not be
ignored because of the large temperature rise. Together with the timedependent C
eff
,
there is no easy analytical solution available for Eq. 4.6. Therefore, an efficient
simulation methodology is desirable and will be discussed in details in the following. In
contrast with the unsolvable nature of the analytical analysis, SPICEbased simulation
methodology provides a convenient and efficient transient thermal analysis. RC
transmission lines are used to model the heat diffusion (Fig. 4.2) due to the similarity of
the respective governing equations as illustrated in Fig. 4.1. The distributed thermal
resistance and thermal capacitance components are constructed, using Eq 4.1Eq. 4.4, to
form thermal RC network. Heat generation, calculated by Eq. 4.5, is realized as voltage
(temperature) controlled current sources as shown in Fig. 4.2. After substituting these
thermal quantities into the correspondent electrical identities in the thermal circuit model,
the temperature along the interconnect, manifested as node voltage, can be easily
Section: Transient Stress Analysis
77
obtained through SPICE simulation. This technique was validated by comparing with
experimental data from [60] as shown in Fig. 4.7, where it shows excellent agreement.
SiO2
Cu
TaN
3 um
SiO
2
Al
TiN
0.6um
3 um
0.05um
0.15um
0.01 um
0.6um
SiO2
Cu Cu
TaN
3 um
SiO
2
Al
TiN
0.6um
3 um
0.05um
0.15um
0.01 um
0.6um
0
200
400
600
800
1000
0 2 4 6 8
t=0.6um, exp.data
t=0.6um, simulation
t=1.2um, exp. data
t=1.2um, simulation
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Current Density, J [x10
7
A/cm
2
]
0
200
400
600
800
1000
0 2 4 6 8
t=0.6um, exp.data
t=0.6um, simulation
t=1.2um, exp. data
t=1.2um, simulation
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Current Density, J [x10
7
A/cm
2
]
(a)
(b)
Fig. 4.7: (a) Schematic cross section of Al and Cu interconnects with cladding layers
used in model validation Fig. 4.7(b) and for simulations in Fig. 4.8.
(b) Validation of HSPICE thermal simulation with experimental data from
[60]. Al interconnect test structure shown in Fig. 4.7(a) is used. Two different
wire thickness , t, are under examination.
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
78
4.4.2. Al vs. Cu Interconnects
With the interconnect structure of Al and Cu wires described in Fig. 4.7(a),
temperature rise is simulated and the comparison is shown in Fig. 4.8. It should be noted
that the simulation is done for a 3D parallel metal array, but only the wire under
examination has a current pulse flowing through it because event like ESD usually strikes
very few wires at a time. The materials properties again are listed in Table 4.1. As can
be observed from Fig. 4.8, for the same cross section, current density, and surrounding
dielectric, Cu wires would experience lower ∆T
max
than Al wires, due to their higher
0
200
400
600
800
1000
1200
0 2 4 6 8 10
Al, t=0.6um
Cu, t=0.6um
Al, t=1.2um
Cu, t=1.2um
Current Density, J [x10
7
A/cm
2
]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
0
200
400
600
800
1000
1200
0 2 4 6 8 10
Al, t=0.6um
Cu, t=0.6um
Al, t=1.2um
Cu, t=1.2um
Current Density, J [x10
7
A/cm
2
]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Fig. 4.8: For the same wire cross sections, shown in Fig. 4.7(a), Cu wire shows much
lower temperature rise compared to Al wire under a 200 ns pulse stress.
Symbols represent simulation data.
Section: Transient Stress Analysis
79
thermal conductivity and thermal capacity, and most importantly, due to their lower
resistivity. This along with higher melting point (~1100
0
C) than that of Al (~660
0
C)
would provide more thermal margin to Cu interconnects. However, ∆T
max
can still be
high when metal dimensions are scaled down and lowk dielectrics are incorporated.
Therefore, it is prudent to study the effect of vias on the thermal characteristics of Cu
wires to make reliable use of them in deep submicron designs.
4.4.3 Impact of Via Separation and Lowk Dielectrics
Before continuing our discussion, we would like to remind the readers again that, if
it is not otherwise mentioned, the interconnect dimensions used throughout the following
sections are taken from global interconnect levels from the 100 nm technology node
based on the ITRS [58] with metal height H=0.8 µm, wire width w=wire spacing
S=0.3µm, and ILD thickness t
ILD
=0.8 µm. If a local interconnect is simulated, then the
dimensions are metal height H=0.26 µm, wire width w=wire spacing S=0.13µm, and ILD
thickness t
ILD
=0.32 µm.
Fig. 4.9 compares the normalized spatial temperature distribution along an
interconnect line with polymer as insulation material and a via separation of 100 µm,
subjected to transient current pulses of 200 ns and 2 µs duration. It can be observed that
the temperature rise profile for the 2 µs pulse is more gradual due to the increased
influence of vias for longer diffusion time, which results in longer diffusion lengths. The
heat diffusion length, L
D
∝ (αt)
1/2
, is 5 µm and 16 µm for the 200 ns and 2 µs pulse
durations respectively, where α is the thermal diffusivity of the interconnect materials.
Thermal diffusion length L
D
can be interpreted as the distance over which via effect is
prominent, which is a function of time and should not be confused with the thermal
healing length L
H
under steadystate condition as mentioned in Chapter 2 and Section 4.3.
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
80
0
0.2
0.4
0.6
0.8
1
1.2
0 10 20 30 40 50 60 70 80 90 100
200ns pulse
2us pulse
Location along Interconnect [um]
N
o
r
m
a
l
i
z
e
d
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
0
0.2
0.4
0.6
0.8
1
1.2
0 10 20 30 40 50 60 70 80 90 100
200ns pulse
2us pulse
Location along Interconnect [um]
N
o
r
m
a
l
i
z
e
d
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
Fig. 4.9: Simulated normalized temperature (∆T/∆T
max
) profile along global Cu/low
k (polymer) interconnect for two pulse durations with current density
J=4x10
7
A/cm
2
.
0.1
1
10
100
1000
0 2 4 6 8 10
via sep.: 10um
via sep.: 20um
via sep.: 35um
via sep.: 50um
via sep.:100um
Time after ESD pulse passed [us]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
0.1
1
10
100
1000
0 2 4 6 8 10
via sep.: 10um
via sep.: 20um
via sep.: 35um
via sep.: 50um
via sep.:100um
Time after ESD pulse passed [us]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Fig. 4.10: The temperature decay after a 200 ns current pulse with J=8x10
7
A/cm
2
, for
Cu/lowk (polymer) global wires. The decay is facilitated by the presence
of vias.
Section: Transient Stress Analysis
81
L
D
is a function of time under transient condition and L
H
is only a function of materials
properties and is not varying with time. Fig. 4.10 plots the temperature decay after a 200
ns pulse. It can be observed that the temperature decays more rapidly with vias placed
closer. This shortens the high temperature span that the interconnect would experience
and thus reduces thermal problems. ∆T
max
vs. via separation for different dielectrics is
shown for high current pulse duration of 200 ns in Fig. 4.11(a) and 100 ns in Fig. 4.11(b).
∆T
max
is higher for the 200 ns pulse due to higher pulse energy, which results in greater
Joule heating. However, both Fig. 4.11(a) and Fig. 4.11(b) indicate that ∆T
max
would be
reduced dramatically for smaller via separation. For larger via separation ∆T
max
saturates
since the effect of vias diminishes. This is dictated by thermal diffusion length (L
D
),
which can be interpreted as the distance over which heat generated in the wire flows
through the via. It can also be observed in Fig. 4.11(b) that for shorter pulse duration the
differences between ∆T
max
for various dielectrics is much smaller. This is due to the fact
that heat does not have sufficient time to diffuse through the surrounding dielectrics.
The maximum temperature rise, at the end of a 200 ns pulse, ∆T
max
, is shown for
global (Fig. 12(a)) and local interconnects (Fig. 12(b)) for different dielectrics as a
function of the current density. It can be observed that while the temperature rises
sharply and lowk dielectrics show worse situation for a typical global line with via
separation of 100 µm, the temperature rise is significantly alleviated for local
interconnects due to a smaller via separation of 1 µm. Even if air is used as both the ILD
and the IMD dielectric (worst case thermal scenario), no significantly higher ∆T
max
is
observed for the local interconnects. This observation is further validated in Fig. 13 with
1 µm via separation, which shows that the spatial temperature distributions for various
dielectrics are similar and ∆T
max
is nearly independent of the dielectric material. This
suggests that they are all within L
D
and via effect dominates the thermal characteristics.
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
82
0
200
400
600
800
0 25 50 75 100 125 150 175 200
ILD & IMD: SiO2
ILD & IMD: Polymer
ILD & IMD: Air
ILD: SiO2, IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Via Separation [um]
0
200
400
600
800
0 25 50 75 100 125 150 175 200
ILD & IMD: SiO2
ILD & IMD: Polymer
ILD & IMD: Air
ILD: SiO2, IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Via Separation [um]
0
50
100
150
200
250
300
0 50 100 150 200
ILD & IMD: SiO2
ILD & IMD: Polymer
ILD & IMD: Air
ILD: SiO2, IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Via Separation [um]
0
50
100
150
200
250
300
0 50 100 150 200
ILD & IMD: SiO2
ILD & IMD: Polymer
ILD & IMD: Air
ILD: SiO2, IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Via Separation [um]
(a)
(b)
Fig. 4.11: (a) ∆T
max
of global interconnects for a 200 ns current pulse with
J=6x10
7
A/cm
2
. ∆T
max
increases with via separation and saturates
when via effect has diminished. (b) For a short pulse of 100 ns, with
the same current density as in Fig. 4.11(a), ∆T
max
of the global Cu
interconnect is much less dependent on the dielectric materials.
Section: Transient Stress Analysis
83
0
20
40
60
80
100
0.E+00 4.E+07 8.E+07 1.E+08 2.E+08 2.E+08
ILD & IMD: SiO2
ILD & IMD: Polymer
ILD & IMD: Air
ILD: SiO2, IMD:Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Current Density, J [ A/cm
2
]
0
20
40
60
80
100
0.E+00 4.E+07 8.E+07 1.E+08 2.E+08 2.E+08
ILD & IMD: SiO2
ILD & IMD: Polymer
ILD & IMD: Air
ILD: SiO2, IMD:Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Current Density, J [ A/cm
2
]
0
200
400
600
800
1000
1200
1400
0 2 4 6 8
ILD & IMD: SiO2
ILD & IMD: Polymer
ILD & IMD: Air
ILD: SiO2, IMD:Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Current Density, J [x10
7
A/cm
2
]
0
200
400
600
800
1000
1200
1400
0 2 4 6 8
ILD & IMD: SiO2
ILD & IMD: Polymer
ILD & IMD: Air
ILD: SiO2, IMD:Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Current Density, J [x10
7
A/cm
2
]
(a)
(b)
Fig. 4.12: (a) ∆T
max
of Cu global interconnect with 100 µm via separation under a 200
ns current pulse. Temperature rises sharply with current density. (b) ∆T
max
of
Cu local interconnect with 1 µm via separation under a 200 ns current pulse.
Temperature rises are much lower and nearly independent of the surrounding
dielectric materials.
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
84
4.4.4 Impact of Dummy Thermal Vias
Dummy thermal vias, which conduct heat but are electrically isolated, can be
installed in Cu/lowk structure to lower the temperature rise. Since a highcurrent short
pulse (J > 10MA/cm
2
, and t
pulse
< 200 ns) usually causes much higher ∆T
max
due to more
severe selfheating than under normal operating conditions and the heat diffusion is
limited to the immediate materials in contact with the metal line, the transient condition
will define the most stringent via separation requirement for lowering the temperature of
the metal wires. It should be noted that the simulation methodology developed here is
quite general, and can be easily extended to study steadystate stress conditions by using
longer pulses. The advantage of the dummy via effect can be demonstrated by using the
0
2
4
6
8
10
12
14
16
18
0 0.2 0.4 0.6 0.8 1
ILD & IMD: SiO2
ILD & IMD: Polymer
ILD & IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Location along Interconnect [um]
0
2
4
6
8
10
12
14
16
18
0 0.2 0.4 0.6 0.8 1
ILD & IMD: SiO2
ILD & IMD: Polymer
ILD & IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Location along Interconnect [um]
Fig. 4.13: Temperature profiles along local Cu interconnect under a 200 ns pulse
are gradual for all the dielectric materials and ∆T
max
are about the same
with J=1x10
8
A/cm
2
.
Section: Transient Stress Analysis
85
0
200
400
600
800
0 20 40 60 80 100
ILD & IMD: Air
ILD & IMD: Polymer
ILD & IMD: SiO2
Via Separation [um]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
(a)
0
100
200
300
400
500
600
0 20 40 60 80 100
ILD & IMD: SiO2
ILD & IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Via Separation [um]
0
100
200
300
400
500
600
0 20 40 60 80 100
ILD & IMD: SiO2
ILD & IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Via Separation [um]
(b)
0
200
400
600
800
0 20 40 60 80 100
ILD & IMD: Air
ILD & IMD: Polymer
ILD & IMD: SiO2
Via Separation [um]
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
(a)
0
100
200
300
400
500
600
0 20 40 60 80 100
ILD & IMD: SiO2
ILD & IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Via Separation [um]
0
100
200
300
400
500
600
0 20 40 60 80 100
ILD & IMD: SiO2
ILD & IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Via Separation [um]
(b)
0
100
200
300
400
500
600
0 20 40 60 80 100
ILD & IMD: SiO2
ILD & IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Via Separation [um]
0
100
200
300
400
500
600
0 20 40 60 80 100
ILD & IMD: SiO2
ILD & IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Via Separation [um]
(b)
Fig. 4.14: (a) ∆T
max
of Cu/air and Cu/polymer can match ∆T
max
of
Cu/oxide global
interconnect if dummy thermal vias are added every 20 µm and 30 µm,
respectively, with t
pulse
= 200 ns and J=6x10
7
A/cm
2
. (b) Cu/air with
thermal vias every 20 um shows nearly the same ∆T
max
as that of Cu/oxide
global interconnect with 100 µm via separation under a 200 ns and J=6x10
7
A/cm
2
current pulse.
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
86
lowestk dielectric, air, in ULSI interconnects despite its poor thermal properties. In the
case of global interconnect, Fig. 4.14(a) shows that thermal vias would be required
approximately every 20 µm in Cu/air and every 30 µm in Cu/polymer interconnect
structures to match the temperature rise of Cu/SiO
2
. For the purpose of comparison,
under normal steady state operating condition with the J
max
specified in the ITRS [17],
the thermal via separation can be much more relaxed based on a simple analytical
evaluation. The temperature profiles for Cu/SiO
2
with via separation of 100 µm, and for
Cu/air with via separation of 20 µm, experiencing the same ∆T
max
under a 200 ns pulse
are shown in Fig. 4.14(b). It can be observed that the temperature rise along the Cu/air
wire varies spatially and that the average temperature is much lower than that of Cu/SiO
2
wire, resulting in reduced thermal problems and may relax the requirement for thermal
via separation.
4.4.5. Impact of Interconnect Aspect Ratio
Finally, the effect of wire aspect ratio on the thermal characteristics is evaluated.
For the same metal thickness, wires with smaller aspect ratio would suffer higher ∆T
max
because of the smaller surface areatovolume ratio (Fig. 4.15(a)). For the same cross
section area, indicating same current capability, a larger perimeter would result in lower
∆T
max
by offering
larger area for heat to diffuse out of metal wires, as shown in Fig.
4.15(b). However, for embedded air gap (ILD: SiO
2
and IMD: Air), ∆T
max
increases
slightly with aspect ratio. This can be explained by the fact that with higher aspect ratio,
embedded air gap interconnect structure would have increasing area contacted by air.
Fig. 15(a) and (b) can be used to provide thermal design guidelines for interconnect.
Section:
87
300
350
400
450
500
550
600
650
700
750
0 0.5 1 1.5 2 2.5 3 3.5
ILD& IMD: SiO2
ILD& IMD: Polymer
ILD& IMD: Air
ILD: SiO2, IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Metal Wire Aspect Ration (AR)
300
350
400
450
500
550
600
650
700
750
0 0.5 1 1.5 2 2.5 3 3.5
ILD& IMD: SiO2
ILD& IMD: Polymer
ILD& IMD: Air
ILD: SiO2, IMD: Air
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
m
a
x
[
°
C
]
Metal Wire Aspect Ration (AR)
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
x
[
°
C
]
300
350
400
450
500
550
600
650
700
750
0 0.5 1 1.5 2 2.5 3 3.5
ILD&IMD: SiO2
ILD&IMD: Polymer
ILD&IMD: Air
ILD: SiO2, IMD: Air
Metal Wire Aspect Ration (AR)
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
,
∆
T
x
[
°
C
]
300
350
400
450
500
550
600
650
700
750
0 0.5 1 1.5 2 2.5 3 3.5
ILD&IMD: SiO2
ILD&IMD: Polymer
ILD&IMD: Air
ILD: SiO2, IMD: Air
Metal Wire Aspect Ration (AR)
(a)
(b)
Fig. 4.15: (a) For the same metal thickness (2.5 µm), global interconnects with
lower aspect ratio (AR), shows higher ∆T
max
due to lower surface areato
volume ratio. t
pulse
= 200 ns and J=6x10
7
A/cm
2
. (b) For the same metal
wire cross section area (5 µm
2
), the impact of aspect ratio (AR) and
dielectric structure strategy is shown. ∆T
max
peaks at aspect ratio=1 due to
the smallest perimeters. t
pulse
= 200 ns and J=6x10
7
A/cm
2
.
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
88
4.5 Conclusions
In this chapter, a simple SPICEbased 3D thermal circuit simulation methodology
for steadystate and transient stress conditions is presented which allows quick evaluation
of various Cu/lowk interconnect structures. The SPICE simulation results show excellent
agreement with rigorous finite element simulations and experimental data and can be
employed for accurate reliability and performance analysis. It is demonstrated that via
density strongly affects the spatial temperature distribution as well as the maximum
temperature rise in interconnects and should be considered in interconnect design. In
fact, the effective thermal conductivity of ILDs can be significantly higher than the
nominal value if via separation is comparable to the thermal characteristic length.
Additionally, the impact of dummy thermal vias on the thermal characteristics of
interconnects has been presented. It is shown that by optimal spacing of dummy thermal
vias the Cu/lowk structures can have the same thermal capability as Cu/SiO
2
.
Furthermore, it has been demonstrated that thermal coupling can be significant for wires
in densely packed structure. The impact of wire aspect ratio (AR) on the thermal
characteristics has also been shown to be important for the thermal design of deep sub
micron interconnect structures. Finally, as technology keeps scaling, thermal effects
should be carefully evaluated not only to address reliability concerns, but also for
accurate interconnect performance analysis. It has been shown that, to accurately
estimate the temperature rise and the thermal characteristics of interconnect, the effect of
via as efficient thermal paths must be considered properly.
89
Chapter 5
Thermal Analysis of 3D ICs
5.1 3D Integration: Background and Motivation
Interconnect RC delay is increasingly becoming the dominant factor determining
the performance of advanced ICs. On the other hand, SystemonaChip (SoC) designs
are often driven by the evergrowing demand for increased system functionality, heading
to dramatic increase in chip area and the use of numerous digital blocks with many long
global wires. The three dimensional (3D) IC architecture has emerged as a unique
solution to alleviate the interconnect delay problem [6567]. Additionally, 3D ICs
provide a vehicle for heterogeneous integration to realize SoC designs, e.g., to incorpo
rate such diverse combinations as embedded DRAM, highperformance logic, analog, RF
ICs and optical interconnects [6872], as schematically illustrated in Fig. 5.1. Further
more, the noise generated by the interference between different embedded circuit blocks
containing digital and analog circuits can be minimized.
Chapter 5: Thermal Analysis of 3D ICs
90
3D integration (schematically illustrated in Fig. 5.2) to create the vertical stacking
of multilevel active layers of Si ICs is a concept that may significantly improve deep
submicron interconnect performance, increase transistor packing density, and reduce chip
area and power dissipation. Each Si layer in the 3D structure may have its own dedi
cated or shared interconnect network. Each of these layers are connected together
through Vertical InterLayer Interconnects (VILICs) and common global interconnects as
shown in Fig. 5.2. Souri [73] has demonstrated the basic concept for this 3D analysis, as
depicted in Fig. 5.3. A general representation of a wirepitch limited 2D IC is consid
ered as consisting of a number of logic blocks. By migrating to a 3D structure it is
assumed that logic blocks can be rearranged in some fashion so as to occupy any number
Fig. 5.1: Schematic of a 3D chip showing integrated heterogeneous technologies.
Section: 5.1 3D Integration: Background and Motivation
91
of active layers of Si. Such an arrangement makes the vertical dimension available for
logic block interconnectivity. For instance, a global wire in the 2D IC connecting two
logic blocks across the chip and contributing to the chip size can now be replaced with a
VILIC connecting the same two blocks, which can be arranged vertically stacked on top
of another. This VILIC is characterized by its much shorter length and smaller contribu
tion to chip size as compared to the original global wire. By performing such
Fig. 5.2: Schematic representation of 3D integration with multilevel wiring network
and VILICs. T1: first active layer device, T2: second active layer device,
Optical I/O device: third active layer I/O device. M’1 and M’2 are for T1,
M1 and M2 are for T2. M3 and M4 are shared by T1, T2, and the I/O de
Chapter 5: Thermal Analysis of 3D ICs
92
replacements across the entire interconnect network, a significant fraction of lateral wires
can thus be replaced with VILICs which ultimately reduces the horizontal wiring re
quirement and chip size and prevents the interconnect delay problem from dominating IC
performance.
However, there is concern of poor heat dissipation and consequent chip tempera
ture rise in 3D ICs due to increased power density and lower thermal conductivity of
inter layer dielectrics (ILD) [24]. The majority of the thermal power dissipated in
integrated circuits arises due to signal switching. This heat is typically conducted
2D
1 2
3 4
5
3D
2 active layers
3 1
2 4
5
Fig. 5.3: Horizontal interconnects are replaced with VILICs, reducing wiring
requirement, chip area and interconnect delays.
Section: 5.1 3D Integration: Background and Motivation
93
through the silicon substrate to the package and then to the ambient by a heat sink. With
multilayer device designs, devices in the upper layers will also generate a significant
fraction of the heat. Furthermore, all the active layers will be insulated from each other
by layers of dielectrics (LTO, HSQ, polyimide etc.) which typically have much lower
thermal conductivity than does Si [74, 75]. Assuming low thermal conductivity of the
ILDs, the heat dissipation issue may become even more acute for 3D ICs and may cause
degradation in device performance, and reduction in chip reliability due to increased
junction leakage, electromigration failures, and by accelerating other failure mechanisms
[76].
In this chapter, a detailed thermal analysis of high performance 3D ICs is pre
sented under various integration schemes [77]. The analysis presented here is the
culmination of collaborative work with Shukri Souri at Stanford University. In Section
5.2, a complete thermal model including power consumption due to both transistors and
interconnect joule heating from multiple strata is presented. The vertical links (VILICs)
and vias have much higher thermal conductivity and hence can effectively reduce the
thermal resistance caused by the ILD layers. Ignoring the effect of these structures can
result in overly pessimistic estimations predicting unacceptably high 3D chip tempera
tures. Section 5.3 shows schematically the power analysis of 3D ICs. Tradeoffs
between power, performance, chip area and thermal impact are evaluated. Section 5.4
discusses the thermal impact on heterogeneous 3D integration and the main results are
summarized in Section 5.5.
Chapter 5: Thermal Analysis of 3D ICs
94
5.2 3D Thermal Modeling
The significant increase in device switching speed and current density in the inter
connect layers, together with increasing Cu resistivity, due to barriers, surface scattering
and skin effect, have augmented heat generation in IC chips. Exacerbated by the lowk
dielectrics with poor thermal conductivity, thermal effects will not only lead to higher
interconnect temperature in 2D ICs but also impact the device temperature in various
active layers in 3D ICs (Fig. 5.4).
Package
+
Heat Sink
Bonding Interface
Si Substrate
Si _2
Si _3
interconnect
interconnect
interconnect
Package
+
Heat Sink
Bonding Interface
Si Substrate
Si _2
Si _3
interconnect
interconnect
interconnect
Fig. 5.4: Schematic of multilevel 3D IC with a heat sink attached to Si substrate.
Section: 5.2 3D Thermal Modeling
95
As seen in Fig. 5.5, even for the case of two active layers, the ratio of thermal resis
tance caused by the ILD layers (R
ILD
) to the required package (including glue layers, heat
sink) thermal resistance (R
pkg
) increases rapidly for future technology nodes. The re
quired R
pkg
is the maximum allowed value which gives the maximum junction
temperature specified in the ITRS [58]. With multiple active layers, R
ILD
will become the
dominant factor to determine temperature rise in 3D ICs.
35 50 70 100 130 180
0
0.5
1
1.5
2
2.5
3
3.5
0
5
10
15
20
25
30
35
R
e
q
u
i
r
e
d
R
p
k
g
[
c
m
2
C
/
W
]
R
I
L
D
/
R
p
k
g
(
%
)
Technology Node [nm]
35 50 70 100 130 180
0
0.5
1
1.5
2
2.5
3
3.5
0
5
10
15
20
25
30
35
R
e
q
u
i
r
e
d
R
p
k
g
[
c
m
2
C
/
W
]
R
I
L
D
/
R
p
k
g
(
%
)
Technology Node [nm]
Fig. 5.5: The required package thermal resistance, R
pkg
, to achieve the maximum
junction temperature specified in ITRS and the ratio of R
ILD
and R
pkg
vs.
technology nodes.
Chapter 5: Thermal Analysis of 3D ICs
96
The analytical expression derived based on first principles, to evaluate temperature
rise in 3D structure is given below [77]:
where:
• T
amb
: ambient temperature.
• M : number of strata.
• N
m
: number of metal levels in the m
th
stratum.
• mn : the n
th
interconnect level in the m
th
stratum.
• t
ILD
: thickness of ILD.
• k
ILD
: thermal conductivity of ILD materials.
• s : heat spreading factor [45].
• η : via correction factor, 0≤ η≤ 1 [45].
• j
rms
: rootmeansquare value of current density flowing in the wires.
• ρ : electrical resistivity of metal wires.
• H : thickness of metal wires.
• Φ : total power consumption of m
th
stratum, including power consumed by active
layer and interconnect joule heating.
• R : thermal resistance of glue layer and Si layer for each of the stratum, with R
1
represents the total thermal resistance of package, heat sink and Si substrate.
Via effect is incorporated in the expression by the via correction factor η (0≤η≤1),
with k
ILD,eff
= k
ILD
/η, where k
ILD,eff
is the effective thermal conductivity of ILD with the
)} ( ] ) ( {[
1
2
,
1 1 ,
,
_ ∑ ∑ ∑ ∑ ∑
= + = = = =
Φ + Φ + + =
M
m k
k m
M
m j
j
N
n i
mn mn rms mn
N
m
N
n mn mn ILD
mn ILD
amb N Si
R H j
s k
t
T T
m m
ρ η
Temperature rise caused by ILDs
Temp. Rise caused by
PKG, glue layer, Si sub.
)} ( ] ) ( {[
1
2
,
1 1 ,
,
_ ∑ ∑ ∑ ∑ ∑
= + = = = =
Φ + Φ + + =
M
m k
k m
M
m j
j
N
n i
mn mn rms mn
N
m
N
n mn mn ILD
mn ILD
amb N Si
R H j
s k
t
T T
m m
ρ η
Temperature rise caused by ILDs
Temp. Rise caused by
PKG, glue layer, Si sub.
Section: 5.2 3D Thermal Modeling
97
help of via effect and k
ILD
is the nominal thermal conductivity with via effect ignored
[45]. Power consumption due to both active (device) layers and interconnect joule
heating are included. This expression can be better understood by comparing it with the
Elmoredelay model following an electricalthermal analogy (Fig. 5.6).
The model has been validated by comparing it with full chip thermal simulations
done using ANSYS [24]. The twolayer 3D structure with wafer bonding technique
used for the validation is shown in Fig. 5.7. Since it is too complicated to construct the 3
D structure with thousands of vias in ANSYS, the validation is done for the case where
via effect is ignored (η = 1). The results obtained from analytical expressions show
excellent agreement with ANSYS (Fig. 5.8), where J
rms
is 4.85x10
5
A/cm
2
and power
Q
3
ElmoreDelay Analogy
Q
1
Q
2
R
3
R
2
R
1
∆T
3
= Q
3
R
3
+(Q
3
+Q
2
)R
2
+(Q
3
+Q
2
+Q
1
)R
∆T
2
= (Q
3
+Q
2
)R
2
+(Q
3
+Q
2
+Q
1
)R
1
∆T
1
= (Q
3
+Q
2
+Q
1
)R
1
Fig. 5.6: ElmoreDelay model – electrical analogy for the thermal mode em
ployed [45].
Chapter 5: Thermal Analysis of 3D ICs
98
density of each active layer is assumed to be 0.615 W/cm
2
. Interconnect parameters are
quoted from the ITRS 100 nm technology node. For the case where via effect is in
cluded, the via separations assigned to the metal layers, from 1
st
to 8
th
levels, are 1, 5, 15,
30, 50, 80, 200 and 500 µm, respectively. However, the analytical model takes much less
computation time and provides key design insight. Furthermore, as shown in Fig. 5.8, via
effect greatly helps with heat dissipation and the resultant temperatures are much lower
as compared to previous works [24, 67, 78]. Therefore, it is crucial to include via effect
for thermal analysis of 3D ICs.
Cu Cu Air
Si_1
Si_2
Interstratum
interconnect
Vias
Metal wires
and ILD
Cu Cu Air
Si_1
Si_2
Interstratum
interconnect
Vias
Metal wires
and ILD
Fig. 5.7: Schematic of 3D structure fabricated by wafer bonding using Cu
pad thermocompression.
Section: 5.3 Power Analysis of 3D ICs
99
5.3 Power Analysis of 3D ICs
In general, performing any comparison between 2D and 3D configurations of
the same IC is problematic. At issue is the lack of common ground on which to perform
an objective comparison. For instance, the operating frequency can be maintained
constant, yet the chip area and hence power densities must change. Conversely, if the
chip areas are assumed invariant, then the operating frequencies are necessarily different.
Such an assumption serves well for a performance comparison. Comparing power
dissipation and temperature rises as a result of migration from 2D to 3D, on the other
0
20
40
60
80
100
0 2 4 6 8 10 12 14 16 18
ANSYS, vias ignored
Analytical, vias ignored
Analytical, vias included
T
e
m
p
e
r
a
t
u
r
e
R
i
s
e
a
b
o
v
e
S
i
S
u
b
s
t
r
a
t
e
(
C
)
Si_1M
11
M
12
M
13
M
14
M
15
M
16
M
17
M
18
Si_2M
21
M
22
M
23
M
24
M
25
M
26
M
27
M
28
Fig. 5.8: Temperature distribution along the vertical layers from the Si substrate
surface (Si_1) to top metal level of the second stratum. The 3D structure
is shown in Fig. 6. M
11
represents the first metal level in stratum 1, etc.
Chapter 5: Thermal Analysis of 3D ICs
100
hand, is a different matter. To address this thermal comparative issue, the analysis in this
section is again focused on the ITRS projection at the 50nm technology node. However,
several 3D integration cases and scenarios are explored to cover an adequate space for
the comparison.
A summary of this wirepitch limited 2D case and different 3D integration
cases used in this comparative study is listed in Table 5.1. All the data in this table are
calculated based on the 50nm technology node and the thermal resistance of the package
is assumed to be 2.15cm
2
°C/W for a constant supply voltage of 0.6V from ITRS projec
tions for 2D ICs at the 50 nm node. The data in the 2D column represents the standard
2D IC. 3D, Case 1, is a special 3D integration case in that memory and logic from the
2D are each dedicated to separate active layers without any modifications to the wiring.
The resulting chip area, A
c
, is determined by the larger logic area and power dissipation is
unchanged relative to the 2D case. The remaining four 3D cases are obtained, and
compared to 2D, by modifying the chip wiring and assuming the memory is interspaced
with logic. Their characteristics are summarized below:
• 3D, Case 2: Equal f
c
and decreased A
c
;
• 3D, Case 3: Equal f
c
and A
c
;
• 3D, Case 4: 2f
c
and equal A
c
;
• 3D, Case 5: Equal A
c
with f
c
determined by maintaining 2D P
Total
.
where A
c
is the chip area, f
c
is the operating frequency and P
total
is the total power dissipa
tion. The different characteristics of these 3D integration cases give rise to different
Section: 5.3 Power Analysis of 3D ICs
101
power dissipation results as summarized in Table 5.1. The dynamic power dissipation
components considered are due to logic, interconnect (local, semiglobal and global),
clock distribution and repeaters and are calculated using P
Dynamic
=1/2αCV
dd
2
f
c
where α is
the activity factor (assumed to be 0.1), V
dd
is the supply voltage obtained from ITRS, f
c
is
the operating frequency and C is the capacitance. Other power dissipating components
include memory, I/O pads and static components, such as leakage and shortcircuit
currents, are all combined under P
Other
.
2D 3D, Case 13D, Case 2 3D, Case 3 3D, Case 4 3D, Case 5
Active Layers 1 2 2 2 2 2
f
c
(MHz) 3000 3000 3000 3000 6000 3559
Feature Size (nm) 50 50 50 50 50 50
Chip Area (cm
2
) 8.17 4.25 4.51 8.17 8.17 8.17
Memory Area (cm
2
) 3.92 3.92 3.92 3.92 3.92 3.92
Logic Area (cm
2
) 4.25 4.25 5.1 12.42 12.42 12.42
P
Logic
(W) 34.8 34.8 34.8 34.8 69.6 41.28
P
Local
(W) 17.4 17.4 17.44 20.66 10.44 6.19
P
SemiGlobal
(W) 14.63 14.63 6.89 8.16 30.68 18.2
P
Global
(W) 6.96 6.96 4.18 5.63 11.78 6.99
P
Clock
(W) 34.8 34.8 22.97 27.21 56.93 33.76
P
Repeaters
(W) 45.24 45.24 29.7 35.19 73.6 43.65
P
Other
(W) 20.17 20.17 20.17 20.17 40.34 23.93
P
Total
(W) 174 174 136.15 151.82 293.37 174
Power Density Per
Active Layer (Wcm
2
) 21.30 20.47 15.09 9.29 17.95 10.65
Table 5.1: Comparison of power dissipation due to logic, interconnect, clock distribution
and repeaters for 2D and 3D ICs with 2 active layers for ITRS 50nm technology
node. 3D IC cases are presented for comparison by varying the chip area, A
c
, and
operating frequency, f
c
, and represent the same 2D IC (conserving feature size,
number of transistors and functionality) converted to 3D with 2 active layers
Chapter 5: Thermal Analysis of 3D ICs
102
The capacitance, C, is calculated for each component to determine the associated
power dissipated. For P
Logic
, the device capacitance is calculated by considering gate
oxide capacitance, overlap capacitance and junction capacitance all of which are calcu
lated from ITRS data [58]. Interconnect capacitances for the local, semiglobal and
global tiers are found from the wirelength distribution and the dimensions of the wire
pitches for each tier [73]. Clock distribution capacitances are calculated using the
BACPAC model proposed in [79] by considering a buffered HTree model. Power
dissipated by repeaters is calculated based on the driver capacitances and the number of
repeaters. P
Other
is determined in the 2D case to be the sum of remaining components to
achieve the ITRS projected total power dissipation for this generation. Since this compo
nent is assumed dominated by dynamic dissipation, it is considered linearly dependent on
the operating frequency for all 3D cases.
In 3D Case 2, the total power dissipation is seen to decrease primarily due to the re
duction in the wiring requirement thus reducing the interconnect power dissipation,
reducing number of required repeaters and reducing the clock distribution network. 3D
Case 3 is associated with a larger chip area which requires longer interconnect lines, a
larger number of repeaters and clockdistribution network all of which increase the power
dissipation. However, lower power density is achieved due to larger chip area. 3D Case
4 shows a dramatic increase in the power dissipated primarily due to the significant
increase in operating frequency. 3D Case 5 illustrates the increase in the operating
Section: 5.3 Power Analysis of 3D ICs
103
frequency if the chip area and the power dissipation requirements are maintained constant
to 2D.
Although the total power consumption as shown in Table 5.1 is reduced by going
from 2D to 3D ICs due to the reduction in the interconnect and the clock network
related capacitance, the heat removal capability could deteriorate as the upper active
layers experience a longer heat dissipation path to the heat sink. Fig. 5.9 compares the
temperature rise for different 3D integration scenarios. The lowest temperature achieved
is that for 3D Case 3. Here, although the wiring has been reduced by migrating to 3D,
the operating frequency and the chip area have been maintained constant as compared to
2D. The power density, then, is significantly lower than in the 2D case giving rise to a
lower die temperature. It should be noted that, in most of the cases, 3D ICs have similar
0
20
40
60
80
100
120
140
160
2D 3D,
case1
3D,
case2
3D,
case3
3D,
case4
3D,
case5
M
a
x
i
m
u
m
T
e
m
p
e
r
a
t
u
r
e
(
C
)
Fig. 5.9: Comparison of temperature performance among 2D ICs and five dif
ferent twoactivelayer 3D ICs scenarios
Chapter 5: Thermal Analysis of 3D ICs
104
temperature rise but have the advantage of either reduced chip area (Case 2) or increased
operating frequency (Case 5). In the case of equal chip area and operating frequency
(Case 3), lower temperature than 2D ICs can be achieved. Note that to double the
operating frequency (Case 4), temperature will invariably increase. Nevertheless. the
temperature is still much lower than that estimated with via effect ignored [24, 67, 78].
Furthermore, it is desirable to put memory in close proximity to the logic circuitry
in order to reduce latency in high performance microprocessors. 3D IC technology
provides an excellent opportunity to stack memory and logic. The power consumption in
onchip memory is generally less than 10% of total power consumption and the area
occupied by memory and logic are comparable at 50 nm node. With these assumptions,
several schemes are developed, and applied in the following analysis to 3D Case 4.
These schemes are illustrated in Fig. 5.10 where four 3D stack schemes are shown, each
with a different configuration of memory and logic. Fig. 5.10 shows four 3D stack
schemes along with their temperature performance. It can be observed that with logic in
the bottom active layer and memory in the upper layer (Scheme 1), the resultant tempera
ture rise is the lowest. On the other hand, stacking logic parts back to back will
experience much higher temperature rise.
Section: 5.3 Power Analysis of 3D ICs
105
Logic
Memory
(1)
M
L
(2)
L M
L M
(3)
L M
M L
(4)
0
50
100
150
200
250
Scheme 1 Scheme 2 Scheme 3 Scheme 4
Max. Temp. @Si_1
Max. Temp. @Si_2
M
a
x
i
m
u
m
T
e
m
p
e
r
a
t
u
r
e
(
C
)
Fig. 5.10: Thermal capability of high performance 3D ICs (Case 4) for four
different 3D logicmemory integration schemes.
1
1.5
2
2.5
3
3.5
4
0 100 200 300 400 500
Temperature rise above 290 K, ∆T [K]
N
o
i
s
e
F
i
g
u
r
e
,
N
F
(
d
B
)
1
1.5
2
2.5
3
3.5
4
0 100 200 300 400 500
1
1.5
2
2.5
3
3.5
4
0 100 200 300 400 500
Temperature rise above 290 K, ∆T [K]
N
o
i
s
e
F
i
g
u
r
e
,
N
F
(
d
B
)
Fig. 5.11: The noise performance of a typical low noise amplifier (LNA) under
the temperature effect is evaluated. NF is assumed to be 2dB @
290K.
Chapter 5: Thermal Analysis of 3D ICs
106
5.4 Thermal Impact on Heterogeneous 3D Integration
Heterogeneous integration of RF circuit in 3D ICs, e.g., with a microprocessor to
facilitate wireless communication, has attracted major interest with the intent to isolate
substratecoupled noise between digital and analog components [67]. Evaluated in Fig.
5.11 is the noise figure, indexing RF performance, that deteriorates rapidly with tempera
ture rise, imposing more stringent design requirements for 3D integration.
Employing optical interconnects in the global signaling or clocking could eliminate
many problems associated with large multiGHz chips like reducing timing skew, power
and area for clock distribution [80]. An initial investigation to look at the thermal aspects
of the required optical components is shown below. In an onchip optical interconnect
system, photodetectors are responsible for opticaltoelectrical signal conversion. Usu
ally, the detector photocurrent is ~100µA. Even the dark current (noise) increases rapidly
with temperature (Fig. 5.12), a phototodark current ratio of 100 is still retained, which is
enough for reliable communications. In a dense optical interconnect, the receivers are
not noiselimited, but indeed gainlimited aiming at low power.
Focusing on the transmitting components like lasers and modulator diodes, the three
major temperatureinduced changes in performance are drop in quantum efficiency,
degradation of laser threshold current, and shift in transmitter wavelength. Employing
strained multiple quantum well (MQW) active regions with effective facet coatings [81],
Section: 5.3 Power Analysis of 3D ICs
107
[83] and a bimetallic heatsink [84] were explored to get around the detrimental wave
length shift. Recently, a quantumdot laser with a maximum operating temperature of
160°C in pulse mode [85] has been demonstrated.
The last major components are the waveguiderelated devices like modulators, optical
attenuators, and electric field sensors, etc. For a waveguide (attenuator) with an InGaAsP
MQW active layer and InP claddings, a negligible change in absorption coefficient (an
index for fiber loss) has been demonstrated from about 20 to 120°C [86]. Even up to
170°C, the absorption loss is still tolerable.. From the above analysis, the incorporation
1.E11
1.E10
1.E09
1.E08
1.E07
1.E06
1.E05
25 125 225 325
Temperature [C]
I
d
a
r
k
(
A
)
Fig. 5.12: Temperature effect on optical receiver performance is evaluated. A
typical example of TiSi metalsemiconductormetal (MSM) PD dark
current vs. ambience temperature is shown here.
Chapter 5: Thermal Analysis of 3D ICs
108
of optical interconnect into the 3D architecture may require some more device design
efforts for higher temperature operation.
5.5 Conclusions
In conclusion, a compact analytical thermal model to evaluate temperature distribu
tion in 3D ICs including via effect is presented. It is demonstrated that via effect must
be considered to evaluate the thermal capability of 3D ICs. Temperature performance of
various 3D integration schemes has been examined thoroughly. It is shown that with
careful thermal designs, 3D ICs can have the same thermal capability as that of 2D ICs.
For the high performance case where higher temperature is not avoidable, better circuit
design and advanced package solution will be necessary. Finally, examples of 3D
heterogeneous integration of RF circuits and optical interconnects have been explored
from the thermal view point.
109
Chapter 6
Conclusion
6.1 Summary
Due to scaling trends of VLSI technology, Joule heating at the interconnect levels
is becoming nonnegligible. As a result, not only will thermal effects be a major reliabil
ity concern, but also the increase of wire electrical resistivity with temperature can
degrade the expected speed performance. In this thesis, both a compact analytical
formula and a convenient SPICEbased simulation methodology are developed to evalu
ate the temperature rise on interconnects under steadystate and transient stress
conditions. The results demonstrate excellent agreement with experimental data and with
Finite Element (FE) thermal simulations (ANSYS). With the effects of vias, as efficient
heat transfer paths, being taken into account, more realistic and accurate interconnect
thermal analysis is finally possible. Significant differences in temperature distribution
and maximum temperature rise are observed between the realistic situation of heat
Chapter 6: Conclusion
110
dissipation with vias and the overly simplified case that ignores the via effect. Further
more, it shows that the effectiveness of vias in reducing the temperature rise is highly
dependent on the via separation and the dielectric materials used. The effective thermal
conductivity of ILDs can be significantly higher than the nominal value if via separation
is comparable to the thermal characteristic length. Therefore, the thermal problem
associated with lowk insulators is not as bad as it might appear. It should be noted that
global interconnects would suffer much higher temperature rises than local interconnects
due to the much longer via separation and further away from heat sink. Additionally, the
thermal advantage gained by using dummy thermal vias in advanced Cu/lowk intercon
nects is quantified, which may offers a solution to alleviate hot interconnect phenomena.
The impact of Joule heating on the scaling trends of advanced VLSI interconnects
has been evaluated in detail. Coupled analysis of delay and electromigration MTF for
various technology nodes suggests that Joule heating will limit scaling of current density
and use of lowk materials. It shows that the interconnect Joule heating can strongly
affect the maximum operating temperature of the global wires which will, in turn, con
strain the scaling of current density to mitigate electromigration and, thus, greatly
degrade the expected speed improvement from the use of lowk dielectrics. Optimization
with various interconnect parameters is performed to provide thermal design guide
line/insight in the early design phase.
Finally, potential bottlenecks and opportunities of future heterogeneous three di
mensional (3D) ICs with various integration scenarios are identified from the thermal
Section: Future Work
111
point of view. It is shown that under certain scenarios, 3D ICs can actually lead to better
thermal performance than planar (2D) ICs. Tradeoffs among power, performance, chip
real estate and thermal impact for 3D ICs is evaluated.
6.2 Future Work
In our analytical thermal model under steadystate normal operation condition, we
assume the wire resistivity is at a constant value. Although errors introduced because of
this simplification are minimal, more accurate results can be obtained by including the
temperature dependency, carrier scattering and skin effect etc. This will lead to a non
closed form formula, and require a few iterations. However, with the help of computer,
this process should not be too difficult.
On the other hand, we consider the heat sink as the sole thermal dissipation path to
the outside ambient and assume adiabatic boundary conditions on the top and on all the
four side walls of the chip, with the reasoning that the chip is encapulated by some
insulating materials. These assumptions are fairly accurate with wire bonding package
and still quite reasonable with the current flipchip technology. However, with the
increasing number of bumps for each generation, more heat will be dissipated through
these thousands of metal balls. The effect of this additional heat transfer path should be
investigated to ensure valid thermal analysis.
Chapter 6: Conclusion
112
Lastly, the thermal modeling and insight developed in this work should not be lim
ited to interconnect level; it can be easily extended to device level and chip level with
certain modification. The thermal effects will continue to be a major issue for years to
come in chip design, hopefully, more accurate and convenient methodologies can be built
upon this work to facilitate the continued advancement of semiconductor industry
113
Bibliography
[1] G. E. Moore, “Cramming More Components onto Integrated Circuits,” Electronics,
vol.38, no.8, April 19
th
1965.
[2] C. Hu, “MOSFET Scaling in the Next Decade and Beyond,” Semiconductor Inter
national, pp. 105114, 1994.
[3] M. T. Bohr, “Interconnect ScalingThe Real Limiter to High Performance ULSI,”
IEDM Tech. Dig., pp. 241244, 1995.
[4] K. C. Saraswat and F. Mohammadi, “Effect of Interconnection Scaling on Time
Delay of VLSI Circuits,” IEEE Trans. On Electron Devices, vol. ED29, pp. 645
650,1982.
[5] The International Technology Roadmap for Semiconductors (ITRS), 2001
[6] M. T. Bohr and Y. A. ElMansy, "Technology for Advanced HighPerformance
Microprocessors," IEEE Trans. Electron Devices, vol 45, no. 3, pp. 620625, 1998.
[7] T. N. Theis, “The Future of Interconnection Technology”, IBM Journal of R&D,
Vol. 44, 2000.
[8] R. Ho, K. Mai, and M. Horowitz, “The Future of Wires,” Proceedings of the IEEE,
vol.89, no.4, pp. 490504, April 2001.
[9] S. Venkatesan, A. V. Gelatos, S. Hisra, B. Smith, R. Islam, J. Cope, B. Wilson, D.
Tuttle, R. Cardwell, S. Anderson, M. Angyal, R. Bajaj, C. Capasso, P. Crabtree, S.
Das, J. Farkas, S. Filipiak, B. Fiordalice, M. Freeman, P. V. Gilbert, “A High Per
formance 1.8V, 0.20 µm CMOS Technology with Copper Metallization,” IEDM
Tech. Dig., pp. 769772, 1997.
[10] N. Rohrer, C. Akrout, M. Canada, D. Cawthron, B. Davari, R. Floyd, S. Geissler, R.
Goldblatt, R. Houle, P. Kartschoke, D. Kramer, P. McCormick, G. Salem, R.
Schulz, L. Su, L. Whitney, “A 480MHz RISC microprocessor in a 0.12 µm L
eff
Error! No text of specified style in document.
114
CMOS Technology with Copper Interconnects,” Int. SolidState Circuits Conf.,
Tech. Digest, pp. 240241,1998.
[11] E. M. Zielinski, S. W. Russell, R. S. List, A. M. Wilson, C. Jin, K. J. Newton, J. P.
Lu, T. Hurd, W. Y. Hsu, V. Cordasco, M. Gopikanth, V. Korthuis, W. Lee, G.
Cerny, N. M. Russell, P. B. Smith, S. O'Brien, R. H. Havemann, “Damascene Inte
gration of Copper and UltraLowk Xerogel for High Performance Interconnects,”
Tech. Dig. IEDM, pp. 936938,1997.
[12] B. Zhao, D. Feiler, V. Ramanathan, Q. Z. Liu, M. Brongo, J. Wu, H. Zhang, J. C.
Kuei, D. Young, J. Brown, C. Vo, W. Xia, C. Chu, J. Zhou, C. Nguyen, L. Tsau, D.
Dornish, L. Camilletti, P. Ding, G. Lai, B. Chin, M. Johnson, J. Turner, “A Cu/low
k Dual Damascene Interconnect for High Performance and Low Cost Integrated
Circuits,” Symp. VLSI Technology, Tech. Digest, pp. 2829,1998.
[13] F. Pollack, “New Challenges in Microarchitecture and Compiler Design”, PACT
2000, Oct 2000.
[14] P. K. Ko, “HotElectron Effects in MOSFET’s,” Ph.D. dissertation, Univ. of
California, Berkeley, 1982.
[15] C. Fiegna, H. Iwai, T. Wada, T. Saito, E. Sangiorgi, and B. Ricco, “Scaling the
MOS Transistor below 0.1 µm: Methodology, Device Structure, and Technology
Requirements,” IEEE Trans. Electron Devices, vol. 41, no. 6, p. 941, June 1994.
[16] J. Ida, M. Yoshimaru, T. Usami, A. Ohtomo, K.Shimokawa, A. Kita and M. Ino,
“Reduction of Wiring Capacitance with New Low Dielectric SiOF Interlayer Film
for High Speed/Low Power Subhalf Micron CMOS,” VLSI Technology Sympo
sium, Dig. Tech. Papers, pp. 5960,1994.
[17] B. P. Shieh, B. Shieh, K. C. Saraswat, J.P. McVittie, S. List, S. Nag, M. Islamraja,
R. H. Havemann, "AirGap Formation During ILD Deposition to Lower Intercon
nect Capacitance," IEEE Electron Device Lett., vol. 19, no. 1, January 1998, pp.
16~18.
Error! No text of specified style in document.
115
[18] D. H. Cho, Y. S. Eo, M. H. Seung, N. H. Kim, J. K. Wee, O. K. Kwon, and H.S.
Park, “Interconnect Capacitance, Crosstalk, and Signal Propagation Delay for 0.35
µm CMOS Technology,” IEDM Tech. Dig., pp. 619622,1996.
[19] J. R. Black, “Electromigration – A Brief Survey and Some Recent Results,” IEEE
Transactions on Electron Devices, vol. ED16, pp. 338347,1969.
[20] G. Yoh and F. N. Najm, "A Statistical Model for Electromigration Failures," IEEE
2000 1st International Symposium on Quality Electronic Design, San Jose, CA,
March 2022, pp. 4550, 2000.
[21] K. Banerjee, A. Amerasekera, G. Dixit, and C. Hu, “The Effect of Interconnect
Scaling and Lowk Dielectric on the Thermal Characteristics of the IC Metal,”
Tech. Dig. IEDM, pp. 6568,1996.
[22] K. Banerjee, A. Mehrotra, A. SangiovanniVincentelli, and C. Hu, “On Thermal
Effects in Deep Submicron VLSI Interconnects,” 36th ACM Design Automation
Conference, pp. 885891,1999.
[23] Danqing Chen, Erhong Li, Elyse Rosenbaum and SungMo Kang, “Interconnect
Thermal Modeling for Determining Design Limits on Current Density,” Computer
Aided Design of Integrated Circuits and Systems, IEEE transactions on, vol. 19, is
sue 2, pp. 197205 Feb. 2000.
[24] S. Im and K. Banerjee, “Full Chip Thermal Analysis of Planar (2D) and Vertically
Integrated (3D) High Performance ICs,” Tech. Dig. IEDM, pp. 727730, 2000.
[25] Y. S. Ju and K.E. Goodson, “Shorttimescale Thermal Mapping of Interconnects,”
Proc. IEEE I.R.P.S, pp. 320324,1997.
[26] S. Ramaswamy, C. Duvvury, and S. Kang, “EOS/ESD Reliability of Deep Sub
micron NMOS Protection Devices,” Proc. IEEE IRPS, pp. 284291, 1995.
[27] R. V. Andrews, “Solving Conductive Heat Transfer Problems with Electrical
analogue Shape Factor,” Chemical Engineering Process, vol. 51, No. 2, pp. 6771,
1955.
Error! No text of specified style in document.
116
[28] P. R. Gray and D. J. Hamilton, “Analysis of electrothermal integrated circuits,”
IEEE J. SolidState Circuits, vol. SC6, pp. 814, Feb. 1971.
[29] A. A. Bilotti, “Static temperature distribution in IC chips with isothermal heat
sources,” IEEE Trans, Electron Devices, vol. ED21, p. 217226, 1974.
[30] W. R. Hunter, “Selfconsistent solutions for allowed interconnect current density –
Part I: Implications for technology evolution, “ IEEE Trans, Electron Devices, vol.
ED44, pp. 304309, 1997.
[31] W. R. Hunter, “Selfconsistent solutions for allowed interconnect current density –
Part II: Application to design guidelines, “ IEEE Trans, Electron Devices, vol. ED
44, pp. 310316, 1997.
[32] S. Rzepka, K. Banerjee, E. Meusel and C. Hu, “Characterization of selfheating in
advanced VLSI interconnect lines based on thermal finite element simulation,”
IEEE Trans. on Components, Packaging and Manufacturing Technology – part A ,
vol. 21, No.3, pp. 406411, 1998.
[33] J. T. Trattles, A. G. O’Neill, and B.C. Mecrow, “Threedimensional finiteelement
investigation of current crowding and peak temperatures in VLSI multilevel inter
connections,” IEEE Trans. Electron Devices, vol. 40, No. 7, pp. 13441347, 1993.
[34] K. A. Gardner, “Efficiency of extended surfaces,” Trans. ASME, vol. 67, pp. 621
631, 1945.
[35] John H. Lau, Low Cost Flip Chip Technologies, McGrawHill, Inc., NY, pp. 314,
2000.
[36] H. Katto, M. Harada, and Y. Higuchi, “Waferlevel Jramp and Jconstant electromi
gration testing of conventional & SWEAT patterns assisted by a thermal and
electrical simulator,” Proc. of the IRPS, p298305, 1991.
[37] H. A. Schafft, “Thermal analysis of electromigration test structure,” IEEE Trans,
Electron Devices, vol. ED34, pp. 664672, Mar. 1987.
[38] F. P. Incropera and D. P. DeWitt, Introduction of heat transfer, New York: John
Wiley & Sons, pp. 192194, 2002.
Error! No text of specified style in document.
117
[39] Virgil A. Sandborn, Resistance temperature transducers, Metrology Press. Fort
Collins, Colorado, pp. 511, 1972.
[40] W. R. Hunter, “The implications of selfconsistent current density design guide
lines: comprehending electromigration and joule heating for interconnect
technology evolution,” IEDM Dig. of Tech. Papers, pp. 483486, 1995.
[41] W. Y. Shin, M.C. Chang, R.H. Havemann, and J. Levine, “Implications and solu
tions for joule heating in high performance interconnects incorporating lowk
dielectrics,” Digest of 1996 Symposium on VLSI Technology, June, pp. 8384, 1997.
[42] K. Banerjee and A. Mehrotra, "Coupled Analysis of Electromigration Reliability
and Performance in ULSI Signal Nets," IEEE International Conference on Com
puterAided Design (ICCAD), San Jose, CA, November 48, pp. 158164, 2001
[43] T. Y. Chiang, K. Banerjee, and K. C. Saraswat, “Compact modeling and SPICE
based simulation for electrothermal analysis of multilevel ULSI,” ICCAD, pp 165
172, 2001.
[44] T. Y. Chiang, B. Shieh and K. Saraswat, "Impact of Joule Heating on Scaling of
Deep SubMicron Cu/LowK Interconnects," Digest of 2002 Symposium on VLSI
Technology, Honolulu, Hawaii, June, pp. 141142, 2002.
[45] T. Y. Chiang, K. Banerjee, and K. C. Saraswat, "Analytical Thermal Model for
Multilevel VLSI Interconnects Incorporating Via Effect," IEEE Electron Device
Letters, Vol. 23, No. 1, pp. 3133, January, 2002.
[46] T. Y. Chiang, and K. Saraswat, "Closedform Analytical Thermal Model for
Accurate Temperature Estimatin of Multilevel ULSI Interconnects " Digest of 2003
Symposium on VLSI Circuit, Kyoto, Japan. June, pp. 275278, 2003.
[47] W. J. Dally and J. W. Poulton, Digital Systems Engineering, Cambridge University
Press, NY, 1998.
[48] G. Yoh and F. N. Najm, "A Statistical Model for Electromigration Failures," IEEE
2000 1st International Symposium on Quality Electronic Design, San Jose, CA,
March 2022, pp. 4550, 2000.
Error! No text of specified style in document.
118
[49] B. K. Liew, N. Cheung, and C. Hu, “Projecting interconnect electromogration
lifetime for arbrtrary current waveforms,” IEEE Trans. Electron Devices, vol. 37,
pp. 13431351, May 1990.
[50] C. K. Hu and J.M.E. Harper, “Copper interconnect: fabrication and reliability,”
VLSI Technology, Systems, and Applications, 1997. Proceedings of Technical Pa
pers, 1997 International Symposium on, pp.18 – 22, June 1997.
[51] C. K. Hu, B. Luther, F.B. Kaufman, J. Hummel, C. Uzoh, D. J. Pearson, “Copper
interconnection integration and reliability,” Thin solid Films, 262, pp. 8492, 1995.
[52] Jiang Tao, Nathan W. Cheung, and C. Hu, “Electromigration characteristics of
copper interconnects,” IEEE Electron Device Lett., vol. 14, no. 5, pp. 249251, May
1993.
[53] P. Kapur, G. Chandra and K.C. Saraswat, “Technology and reliability constrained
future copper interconnectsPart I: Resistance Modeling,” IEEE Trans. Electron
Devices, vol. 49, no. 4, pp. 590597, April 2002.
[54] P. Kapur, “Scaling induced performance challenges/limitations of onchip metal
interconnects and comparisons with optical interconnects,” Ph.D. dissertation, Stan
ford, 2002.
[55] Linan Jiang; JaeMo Koo; Shulin Zeng; Mikkelsen, J.C.; Lian Zhang; Peng Zhou;
Santiago, J.G.; Kenny, T.W.; Goodson, K.E.; Maveety, J.G.and Tran, Q.A., “Two
phase microchannel heat sinks for an electrokinetic VLSI chip cooling system,”
Semiconductor Thermal Measurement and Management, Seventeenth Annual IEEE
Symposium, pp.153 – 157, March 2001
[56] T. Y. Chiang, K. Banerjee, and K. C. Saraswat, “Effect of via separation and lowk
dielectric materials on the thermal characteristics of Cu interconnects,” Tech. Dig.
IEDM, pp. 261264, 2000.
[57] T. Y. Chiang, and K. C. Saraswat, “Impact of vias on the thermal characteristics of
deep submicron Cu/lowk interconnects,” Tech. Dig., Symp. VLSI Technology, pp.
141142, 2001
Error! No text of specified style in document.
119
[58] C. C. Teng, Y.K. Cheng, E. Rosenhaum and S. M. Kang, “iTEM: a temperature
dependent electromigration reliability diagnosis tool,” IEEE Trans. Computer
Aided Design, vol. 16, No. 8, pp. 882893, Aug. 1997.
[59] The International Technology Roadmap for Semiconductors, 1999.
[60] K. Banerjee, A. Amerasekera, N. Cheung, and C. Hu, “Highcurrent failure model
for VLSI interconnects under shortpulse stress conditions,” IEEE Electron Device
Lett., vol. 18, No. 9, pp. 405407, 1997.
[61] C. Duvvury and A. Amerasekera, “ESD: A pervasive reliability concern for IC
technologies,” Proc. IEEE, vol. 81, pp. 690702, May 1998.
[62] T. Green, “A review of EOS/ESD field failures in military equipment,” in Proc.
10
th
, EOS/EDS symp., pp. 714, 1998.
[63] M. J. Pelgrom and E. C. Dijkmans, “A 3/5 V compatible I/O buffer,” IEEE J. Solid
Statae Circuits, vol. 30, no. 7, pp. 823825, 1995.
[64] F. P. Incropera and D. P. DeWitt, Introduction to heat transfer. New York: John
Wiley & Sons, pp. 6165, 2002.
[65] S. J. Souri and K. C. Saraswat, “Interconnect performance modeling for 3D inte
grated circuits with multiple Si layers,” Int. Interconnect Technology Conf.
Proceedings, pp. 2426, 1999.
[66] S. J. Souri, K. Banerjee, A. Mehrotra, and K. C. Saraswat, “Multiple Si layer ICs:
motivation, performance analysis, and design implications,” Proc. 37
th
ACM De
sign Automation Conf., pp. 873880, 2000.
[67] K. Banerjee, S. J. Souri, P. Kapur, and K. C. Saraswat, "3D ICs: A Novel Chip
Design for Improving Deep Submicron Interconnect Performance and Systemson
Chip Integration," Proceedings of the IEEE, Special Issue, Interconnections Ad
dressing The Next Challenge of IC Technology, Vol. 89, No. 5, pp. 602633, May
2001.
[68] Akasaki, "ThreeDimensional IC Trends", Proceedings of the IEEE, Dec. 1986
Error! No text of specified style in document.
120
[69] N. Takahashi, N. Senba, Y. Shimada, I. Morisaki, K. Tokuno, “Threedimensional
memory module”, IEEE Trans. on Advanced Packaging, Feb. 1998
[70] Y. Akasaka and T. Nishimura, “Concept and basic technologies for 3D IC struc
ture,” IEDM Tech. Dig., pp. 488491,1986.
[71] T. Nishimura, Y. Inoue, K. Sugahara, S. Kusunoki, T. Kumamoto, S. Nakagawa,
M. Nakaya, Y. Horiba, and Y. Akasaka, “Three dimensional IC for high perform
ance image signal processor,” IEDM Tech. Dig., pp. 111114, 1987.
[72] A. V. Krishnamoorthy, A. L. Lentine, K. W. Goossen, J. A. Walker, T. K. Wood
ward, J. E. Ford, G. F. Aplin, L. A. D'Asaro, S. P. Hui, B. Tseng, R. Leibenguth, D.
Kossives, D. Dahringer, M. F. Chirovsky, D. A. B. Miller, “3D integration of
MQW modulators over active submicron CMOS circuits: 375 Mb/s transimpedance
receivertransmitter circuit,” IEEE Photonics Technology Letters, vol. 7, no. 11, pp.
12881290, 1995.
[73] S. Souri, “3D ICs interconnect performance modeling and analysis,” Ph.D. disserta
tion, Stanford, 2003.
[74] K. E. Goodson and Y. S. Ju, “Heat conduction in novel electronic films,” Annu.
Rev. Mater. Sci., 29: pp. 261293, 1999.
[75] K. Banerjee, A. Mehrotra, A. SangiovanniVincentelli, and C. Hu, “On thermal
effects in deep submicron VLSI interconnects,” Proc. 36
th
ACM Design Automa
tion Conference, pp. 885891, 1999.
[76] K. Banerjee, Ph.D. Dissertation, University of California at Berkeley, 1999.
[77] TingYen Chiang, S. J. Souri, Chi On Chui, K. C. Saraswat, “Thermal analysis of
heterogeneous 3D ICs with various integration scenarios,” Electron Devices Meet
ing, 2001. IEDM Technical Digest. 2001
[78] S. A. Kuhn, M. B. Kleiner, P. Ramm, and W. Weber, “Thermal analysis of verti
cally integrated circuits,” in IEDM Technical Digest, pp.487490, 1995
[79] D. Sylvester, Chenming Hu, “Analytical modeling and characterization of deep
submicrometer interconnect,” Proc. IEEE, pp. 634664, May 2001.
Error! No text of specified style in document.
121
[80] D. A. B. Miller, “Optical interconnects to silicon;” Selected Topics in Quantum
Electronics, IEEE Journal of, Volume: 6, Issue: 6 ,Nov.Dec. pp.1312 – 1317,
2000.
[81] A. Kasukawa, N. Iwai, N. Yamanaka, N. Yokouchi, “Very high characteristic
temperature and constant differential quantum efficiency:1.3µm GaInAsPInP
strainedlayer quantumwell lasers by use of temperature dependent reflectivity
(TDR) mirror,”Selected Topics in Quantum Electronics, IEEE Journal of, Volume:
1, Issue: 2, pp. 293300, Jun 1995.
[82] CE Zah, R. Bhat, B. N. Pathak, F. Favire, Wei Lin Wang, M.C.
Andreadakis, N.C. Hwang, D.M. Koza, M.A. TeinPei Lee, Zheng Wang
D. Darby, D. Flanders, J. J. Heieh, “Highperformance uncooled 1.3µm AlxGayIn
1xyAs/InP strainedlayerquantumwell lasers for subscriber loop applications,”
Quantum Electronics, IEEE Journal of, Volume: 30, Issue: 2, pp. 511523, Feb.
1994.
[83] D. A. Cohen, Daniel A. Cohen, Mark E. Heimbuch, and Larry A. Coldrene, “Re
duced temperature sensitivity of the wavelength of a diode laser in a
stressengineered hydrostatic package,” APL, vol. 69, no. 4, pp. 455457, 1996
[84] D. A. Cohen et al., OFC ’97 Tech. Dig., pp. 303304, 1997.
[85] F. Klopf, R. Krebs, J. P. Reithmaier and A. Forchel, “Hightemperature operating
1.3µm quantumdot lasers for telecommunication,” IEEE PTL, vol. 13, no. 8, pp.
764766, 2001.
[86] M. Allard, R. A. Masut, M. Boudreau, “Temperature determination in optoelec
tronic waveguide modulators,” IEEE JLT, vol. 18, no. 6, pp. 813818, 2000.
©Copyright by TingYen Chiang 2004 All Rights Reserved
ii
I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy
___________________________________________ Professor Krishna C. Saraswat, Principal Adviser
I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy
__________________________________________ Professor Robert W. Dutton
I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy
_________________________________________ Dr. James P. McVittie Approved for the University Committee on Graduate Studies:
_________________________________________
iii
iv .
It shows that the effectiveness of vias in reducing the temperature rise in interconnects is highly dependent on the via separation and the dielectric materials used. The simulation v . overly pessimistic estimation of the interconnect temperature will lead to overly conservative approach. As a result. dynamic power consumption and crosstalk noise in advanced technology. a variety of lowk materials have been introduced to reduce the RC delay. Hence. On the other hand. the increasing thermal impedance further exacerbates temperature rise in interconnects.Abstract The scaling of VLSI structures leads to continuous increase in current density that results in ever greater interconnect Joule heating. The results demonstrate excellent agreement with experimental data and those using Finite Element (FE) thermal simulations (ANSYS). is included in our analysis for the first time to provide more accurate and realistic thermal diagnosis. Together with the poor thermal conductivity of such materials and more metal levels added. not only will thermal effects be a major reliability concern. In addition. but also the increase of resistivity with temperature can degrade the expected speed performance. as efficient heat transfer paths to alleviate the temperature rise in the metal wires. This research proposes both compact analytical models and fast SPICE based 3D electrothermal simulation methodology to characterize thermal effects due to Joule heating in high performance Cu/lowk interconnects under steadystate and transient stress conditions. The effect of vias. performing a more realistic thermal modeling and analysis of interconnects is critical.
It is shown that under certain scenarios. Coupled analysis of reliability and delay. thus. is performed to optimize interconnect structures such as wire aspect ratio and ILD thickness. performance.methodology has also been applied to quantify the use of dummy thermal vias as additional heat sinking paths and possible solution to hot wires. chip real estate and thermal impact for 3D ICs are evaluated. under the influence of thermal effects. Finally. Tradeoffs among power. vi . It shows the interconnect Joule heating can strongly affect the maximum operating temperature of the global wires which will. constrain the scaling of current density to mitigate electromigration and. greatly degrade the expected speed improvement from the use of lowk dielectrics. 3 ICs can actually lead to better thermal performance than D planar (2D) ICs. potential bottlenecks and opportunities of future heterogeneous three dimensional (3D) ICs with various integration scenarios are identified from the thermal point of view. in turn. The impact of Joule heating on the scaling trends of advanced VLSI interconnects has been evaluated in detail.
Pianetta. and to Dr. Dutton. Kaustav Banerjee and Sungjun Im whose generosity vii . I would like to thank my adviser Professor Krishna C. I would like to thank members of my Oral and Reading Committees. It is a great pleasure to acknowledge the people who have supported me throughout these years. Kelly for his continuous help and encouragement. I also thank Dr. First and foremost. Professor Saraswat is not only an admirable educator. I feel very fortune to have worked with Professor Saraswat. My gratitude also goes out to Professor Kenneth E. Being one of the most friendly. Shukri Souri. Saraswat for his guidance. and Dr. to Dr. I also thank Professor Michael A. whose sense of outofbox thinking and professionalism I thoroughly enjoyed. both in my research and in my startup dream. but also a lifetime mentor. Professor Robert W. who provides a creative academic atmosphere for the students and always shows his unconditional support to the students in all aspects. Jim McVittie for their enthusiastic discussion and valuable feedback. and who made possible the thermal analysis of 3D ICs through his 3D ICs performance study. Professor Piero A. The experience I have gained under his guidance is invaluable and has left an indelible mark on my self and persona. I am forever grateful his unwavering confidence in me throughout my time at Stanford. Pawan Kapur for numerous entertaining discussions and invaluable brain storming sessions. rewarding and memorable experience. Goodson for his time and effort acquainting me with thermal analysis and. downtoearth professors.Acknowledgements I am deeply indebted to many people who made my Stanford years an enlightening. insight and patience.
Mayur Joshi. Netherlands. My colleagues from various research groups have been terrific and my work has benefited tremendously from interactions with the past and present members. Dr. Carmen Miraflor and Maureen Rochford for working hard to get me back to U. viii . Dr. Mar Hershenson. Dr. Dr. Richard Dasher for coordinating and coming along for a SPIE trip to Philips. Dr. Rohan D. Richard Chang. during the SPIE trip. Abhijit Jayant Pethe. I thank Irene Sweeney for great administrative support. YuLi Hsueh. Shahram Alibeik. Marci Liao. TsuJu Chin. Pinkesh Sachdev. I would also like to thank Dr. I would like to thank Ashish Chopra and Alper Tekin for all the dreams of riches and early retirement and for our TAAV team which aimed to be the leader in the physical analysis segment of electronic design automation with its breakthrough full chip thermal simulation technology. Dr. Hoon Cho. Tejas Krishnamohan. Amol Joshi. Rohit Shenoy. Alvin Loke. Dr. I am very thankful for the assistance and friendship of Dr. Dr. I especially thank Professor Tom Lee. Kambiz Kaviani. ChiOn Chui. Dr. David Bang. In addition.000 prize in the Stanford Entrepreneur’s challenge business plan competition was thrilling. work. when my green card got stolen in Amsterdam.S. YiChang Lu. Although it was short. Ray Chen. Dr. the wonderful experience of writing business plan. presenting our ideas in front of venture capitalists and winning the $10. Ali Kemal Okyay.of discussions and challenges elevated my intellectual awareness of my own work. Netherlands. Dr. Xiao Ning Qi. Steve Jurichich. Nabeel Ibrahim. Pranav Kalavade. Donghyun Kim. I would like to thank MARCO Interconnect Focus Center for funding my Ph. Dr. Ben Shieh and Junfeng Xu. Dr. Hoyeol Cho.D. Ammar Nayfeh. Kekatpure.
and I am grateful to be part of the cast. ChingJung. their children. TingChun and SherFeng. none of this would have been possible without the support.Without question. sacrifices. To my grandparents. I am forever grateful. Thanks to my precious son. my eternal gratitude goes to my mother and father. moral and otherwise. MinChuan and HoChen. for all your support in a time of need. ix . Without all her unparalleled love and support through the years. To my brother and sister. loving face to go home to after a tough day. for being always there for me. the greatest blessing in my heart. are the starting actors. This dissertation. of family and loved ones. for all their years. ChingJung and Eric. commitment and hard work towards each one of us. for my beloved wife. for always greeting me with a smiling. I am eternally indebted and dedicate this work. Most of all. its accomplishment through all the obstacles. To Aunt Chuling and her family. To them. Eric. whose constant inspiration and unreserved support have made it all possible. For your friendship and taking care of our parents when I am far away from home. my deepest appreciation. I would not go back to pursue Ph.D. is a manifestation of their dedication. Finally. degree and this thesis would never have been finished. There will be another book. I thank you for your endless love and encouragements.
x .
2 Multilevel interconnect Formula Interconnect Temperature Rise Trend 2.1 2.2 Motivation Preliminary Observations 2.3.4.3 First glance at via effect 13 13 16 17 22 22 24 27 35 35 36 39 39 41 43 Interconnect Thermal Model and Assumptions 2.4 Background Thermal Implication of Interconnect Scaling Via Effect Dissertation Organization 1 1 4 8 10 Chapter 2 Analytical Interconnect Thermal Model 2.5 Temperature in Multilevel Metal Layers 2.1 1.2 Via Correction Factor Hot Spot Location 2.3.2 2.1 2.1 2.3 Energy Conservation Law Conduction Dominant Heat Transfer Steady State Thermal Modeling 2.4.5.3.2.1 2.6 Summary xi .3 1.4 Impact of Via Effect on Effective k ILD 2.2 1.1 2.5.Contents Chapter 1 Introduction 1.
1 4.5 3.1 3.2 4.3 4.1 Analytical Model vs.4.4.7 Delay and Reliability Optimization Impact of Joule Heating on Scaling Trend Summary Chapter 4 SPICEBased ElectroThermal Simulation Methodology 4.3.4.4 RC Delay Dynamic Power Consumption Cross Talk Noise Electromigration Reliability 50 52 54 54 56 58 60 62 45 45 47 48 50 Scaling Trend of Joule Heating Temperature Effect on Cu/lowk Interconnects 3.3.4.4 4.3.2.3 Motivation SPICEBased Thermal Modeling Steady State Analysis 4.1 3.Chapter 3 Thermal Impact on Interconnect Design 3.2 Introduction Thermal Effect on Interconnect Metrics 3.3 3.4.2 Impact of Via Separation on Effective k ILD Thermal Coupling Effect 63 63 64 68 68 72 4.5 Al vs.3 3.4.2.4 Transient Stress Analysis 4.6 3.3.4 3. Cu Interconnects Impact of Via Separation and Lowk Dielectrics Impact of Dummy Thermal Vias Impact of Interconnect Aspect Ratio xii 74 75 78 79 84 86 .1 3.4. SPICEBased Simulation Methodology 4.1 4.2 Definition of Worst Case Condition Thermal Impact on Cu/lowk Interconnects 3.2 4.2 3.
2 5.5 Conclusions 88 Chapter 5 Thermal nalysis of 3D IC 5.5 3D Integration: Background and Motivation 3D Thermal Modeling Power Analysis of 3D ICs Thermal Impact on Heterogeneous 3D Integration Conclusions 89 89 94 99 106 108 Chapter 6 Conclusions 6.4.4 5.1 6.1 5.2 Summary Future work 109 109 111 Bibliogrphy 113 xiii .3 5.
xiv .
.................101 xv .......................... clock distribution and repeaters for 2 and 3D ICs with 2 active layers for ITRS 50nm D technology node.............................................................. and operating frequency............................................... number of transistors and functionality) converted to 3D with 2 active la yers.....1: Evolution of characteristics for Intel® commercial processors..52 Table 3..............1: Interconnect parameters for 130 nm to 22nm technology nodes based on ITRS [5] .61 Table 4.....56 Table 3........1: Materials properties used in this work..........List of Tables Table 1.................... k [W/mK] is thermal conductivity................. fc........ ............ Ac.......3 Table 3... interconnect..................................69 Comparison of power dissipation due to logic....3: Coupled evaluation of electromigration reliability and performance for global interconnects for 22130 nm technology nodes.. 3D IC cases are presented for comparison by varying the chip area....... ? [µOcm] is electrical resistivity near 100ºC and er is the relative dielectric constant............................................2: Materials properties of various conductors and dielectrics used in this study.. ............1: Table 5.... and represent the same 2D IC (conserving feature size....
xvi .
. ..4: Both dielectric constant and thermal conductivity of ILD materials decrease with advanced technology nodes [5]...1: Fig....... It can be observed that there is no significant difference between the two simulation conditions and temperature rise in the via is minimum..........3: Two extreme simulation conditions: (a) no current flows through via..........List of Figures Fig... (b) all the current flowing through the wire will also flow through the via and cause heat generation..................... .. ................. 100 µm (middle) and 5 µm (bottom) for case (a) and (b) of Fig.. 1.............. 2........9 Fig............1: Fig. 2............2 Evolution of power density for Intel processor [13] (Courtesy of Intel).... from a conducting metal wire embedded in a dielectric medium.... ..4 Schematic showing the hierarchy of metal levels for distribution of interconnects in modern ICs............. so there is no heat generation in the via..7 Fig..5: SEM photo of local level interconnect... 2..... ........3: Predictions for device and wire delays from ITRS 2001 [5].. It should be noted that.............23 xvii ..18 Fig...... 1........................................ ..14 Heat generated by Joule–heating is conducted downwards through both ILD and vias......................2: Fig......... .......4: Temperature distribution along the metal wire with via separation of 250 µm (top)....19 Fig....3........................ shows energy balance.. 2............. 1...... 2....................6 Fig...........5: A differential control volume..21 Fig.... 2......2: Typical multilevel interconnect structure with heat sink attached to substrate. 1........... the via helps transfer heat to underlying layer and reduce the temperature rise on the wire................... in both cases.... L represents the separation between vias. 1............. Figure taken from ITRS 2001 [5]..
....... .. 50............. ..... and jrms..... and hence...........respectively. 2.... the via separations assigned to the metal layers........................39 Fig.....41 xviii ...... the majority of heat will conduct to outside ambient through substrate rather than to printed circuit board (PCB) [35]...........11: Effective ILD thermal conductivity increases with decreasing via separation...... 2......12: The figure shows the temperature profile along the wire from the middle of the interconnect to the via at the right end. Rrect and the spreading factor S’m... the stronger is the via effect..... particularly with fan........ Dielectrics with lower nominal kILD experience lower η even with longer via separation........ .. 2. Wires with shorter via separation feel stronger via effect... 2.....27 Fig....m = 3.........13: Temperature rise distribution along metal layers from substrate to top metal level.... 150.......................... 100........Fig..... 2... based on 65nm technology node..8: Geometry used for calculating Rspr...... representing parallel metal wire array............................ with two different D/W ratios............2 W/mK are assumed here.... Crosssectional view is shown on the left hand side.... w = d = 0.. the longer the LH.7: Geometry used to derive interconnect thermal model.................8 µm...3 W/mK and koxide =1.. 500 and 1000 metal pitches............. 200.................... .10: Via correction factor............................. Quasi 2D heat conduction is used to correct 1 heat conduction model...... 10.........37 Fig......26 Fig......32 Fig...7E6 A/cm2.. The lower the nominal kILD..............m.................. 2.6: With the help of heat sink.9: Temperature profile along the Cu wires with 100 µm via separation...............................37 Fig..... 2. The hot spots usually locate in the wire except for extremely small via dimensions........ ....... are 5......... ..... 30......34 Fig..................... η.. 300... kpolymer = 0.. H = tILD = 0.. 2..... are plotted for three types of dielectrics............. In the case of via effect included... from 1st to 10th levels............. Space b D etween any two wires is shared for heat dissipation................. .......... ...3 µm...........
....................53 Fig............... 3.................48 Fig..... ................ while kILD continues scaling.. The constraint is more stringent for lowk dielectrics. 3...... 3......... (b) j scales with ITRS.........................................7: Maximum allowed current density is limited by the maximum allowed ∆T on the metal wires..............59 xix ... The lower the dielectric constant... ........................................57 Fig......... 3.... the stronger the Joule heating and greater RC degradation................... 3...6: RC delay is strong function of current density on the wires because of Joule heating.....1: Schematic showing the intermetal and the interlevel components of capacitance....4: Temperature on the toplevel interconnects rises rapidly with increasing current density..... ......... (a) both j and kILD scale according to ITRS.......... Current flows through all wires at all metal layers in cases (1) and (3).. (d) both j and kILD stops scaling at 65 nm node...... Dashline curves include the influence of thermal effect.................... ................ but kILD stops scaling at 65nm node.............32/0.... Chip power density is calculated by total power of the chip divided by chip size.....5: Temperature of top global interconnects rises sharply for Lowk dielectrics....................... ........ 3...58 Fig... Solid curves represent the case thermal effect n eglected...............2: Trends of chip power density and interconnect Jmax along technology nodes suggested by ITRS [5].....8: Constant RmetCtotal contour plots of global level wiring as functions of wire and ILD thickness...... 3..14: Trend of interconnect temperature rise along technology nodes for different scenarios......15 µm)...... 2.....56 Fig.......... (c) j stops scaling at 65nm node.................................................................. εr: relative dielectric constant........55 Fig.. 3............. Also showing aspect ratio (AR)............................ ................42 Fig..... ....Fig........ (H/S=0.55 Fig................. ...............3: Configurations of the three thermally worst case scenarios.. k: thermal conductivity [W/Km]. ...
......3µm............. the longer the LH....... The current density in the middle wire is J0...15 µm)...............3µm in this simulation.................. The temperature of wires carrying lower current density is strongly affected by the heat coupled from the neighboring wires.......... (H/S=0........ Line width is 0.... with xx ...Fig........ The lower the nominal kILD......................62 Fig................................68 Fig....3: (a) Interconnect configuration......10: The solid curve shows the MTF (%) that can be achieved under the current density specified in ITRS.... (b) The current density in the middle wire is one third of J0.... 4.....66 Fig.1: Fig.................. .......................................... Line width is kept at 0.60 Fig.......... 4. ... the stronger is the via effect............... 4.............. Reference temperature T0 is the maximum temperature in the middle wire with no current flowing in the nearest neighboring wires and with equal line width and spacing (S). ... 4.... 3....................................................4x106 A/cm2.. and hence........ 1..........5: Effective ILD thermal conductivity increases with decreasing via separation.. 4....... 4...............6: (a) T is defined as the maximum temperature in the middle wire..65 3D thermal circuit RC transmission line model for transient thermal analysis of interconnect structures................4: Validation of HSPICE thermal simulation with Finite Element Analysis (ANSYS) data...............9: Constant normalized electromigration MTF...2: ThermalElectrical analogous parameters...69 Fig.32/0. contour plots of global level wiring as functions of wire aspect ratio (AR) and ILD thickness......................... (b) Correspondent distributed thermal circuitry.... MTF(Tm)/MTF(105°C)........71 Fig. The brokenline curves show the MTF under various wire temperature (Tm) criterions and the values refer to the right axis .................... . .. (c) Top view of interconnect structures................................................. 3.............
77 Fig. Al interconnect test structure shown in Fig..... ∆Tmax of the global Cu interconnect is much less dependent on the dielectric materials..... 4.....83 xxi ...............................................8: For the same wire cross sections.... 4...7(b) and for simulations in Fig.....82 Fig... 4.......................73 Fig.............. used in this simulation and current flow patterns showing current flowing on alternate............80 Fig..... ...........78 Fig......................................... Two different wire thickness ................................ ∆Tmax increases with via separation and saturates when via effect has diminished........................7: (a) Schematic cross section of Al and Cu interconnects with cladding layers Fig.10: The temperature decay after a 200 ns current pulse with J=8x107 A/cm2.........11(a)...................................... 4. 4...................9: Simulated normalized temperature (∆T/∆Tmax) profile along global Cu/lowk (polymer) interconnect for two pulse durations with current density J=4x107A/cm2..... Temperature rises sharply with current density...... shown in Fig.. 4..... Symbols represent simulation data.................... 4...... The decay is facilitated by the presence of vias.7(a) is used.............. (b) ∆Tmax of Cu local interconnect with 1 µm via separation under a 200 ns current pulse.... 4...... (b) Validation of HSPICE thermal simulation with experimental data from [60]..........80 Fig.................................different spacing..................8....... t........... .......... (b) For a short pulse of 100 ns.....11: (a) ∆Tmax of global interconnects for a 200 ns current pulse with J=6x107 A/cm2................. ... Cu wire shows much lower temperature rise compared to Al wire under a 200 ns pulse stress.......................12: (a) ∆Tmax of Cu global interconnect with 100 µm via separation under a 200 ns current pulse.......... 4............................... with the same current density as in Fig.. 4................................ ................... Temperature rises are much lower and nearly independent of the surrounding dielectric materials............... for Cu/lowk (polymer) global wires. are under examination......7(a)..... 4.......
. chip area and interconnect delays .......90 Schematic representation of 3D integration with multilevel wiring network and VILICs........... (b) For the same metal wire cross section area (5 µm2). reducing wiring requirement.3: Horizontal interconnects are replaced with VILICs..........94 The required package thermal resistance......... 4...... 5.........14: (a) ∆Tmax of Cu/air and Cu/polymer can match ∆Tmax of Cu/oxide global interconnect if dummy thermal vias are added every 20 µm and 30 µm... 5... respectively...........................................5: Schematic of multi.... T1: first active layer device............................... to achieve the maximum junction temperature specified in ITRS and the ratio of RILD and Rpkg vs.... . T2.......5 µm)..... M1 and M2 are for T2... 4......15: (a) For the same metal thickness (2....................... Rpkg......... M’1 and M’2 are for T1.......... with tpulse = 200 ns and J=6x107 A/cm2.. tpulse = 200 ns and J=6x107 A/cm2 ..84 Fig.. the impact of aspect ratio (AR) and dielectric structure strategy is shown.......... 4......Fig.......91 Fig................................... T2: second active layer device........... 5.........2: Schematic of a 3D chip showing integrated heterogeneous technologies . and the I/O device............. 5............ global interconnects with lower aspect ratio (AR).......... (b) Cu/air with thermal vias every 20 um shows nearly the same ∆Tmax as that of Cu/oxide global interconnect with 100 µm via separation under a 200 ns and J=6x107 A/cm2 current pulse ..... technology nodes .95 xxii ...13: Temperature profiles along local Cu interconnect under a 200 ns pulse are gradual for all the dielectric materials and ∆Tmax are about the same with J=1x108 A/cm2 ........ .. M3 and M4 are shared by T1...... ∆Tmax peaks at aspect ratio=1 due to the smallest perimeters....level 3D IC with a heat sink attached to Si substrate................ 5......92 Fig....1: Fig... shows higher ∆Tmax due to lower surface areatovolume ratio......4: Fig......................... Optical I/O device: third active layer I/O d evice.......87 Fig... tpulse = 200 ns and J=6x107 A/cm2.85 Fig.........
.. 5.. .................12: Temperature effect on optical receiver performance is evaluated.... 6............................... A typical example of TiSi metalsemiconductormetal (MSM) PD dark current vs......... .......... M11 represents the first metal level in stratum 1...... 5........... 5.....6: ElmoreDelay model – electrical analogy for the thermal mode employed [45]..............97 Fig................11: The noise performance of a typical low noise amplifier (LNA) under the temperature effect is evaluated..................................103 Fig.......................................Fig...............10: Thermal capability of high performance 3 ICs (Case 4) for four different 3D D logicmemory integration schemes............8: Temperature distribution along the vertical layers from the Si substrate surface (Si_1) to top metal level of the second stratum.............98 Fig...................... 5.................... ..............105 Fig... ambience temperature is shown here..... NF is assumed to be 2dB @ 290K....9: Comparison of temperature performance among 2D ICs and five different twoactivelayer 3D ICs scenarios.... ......................................................... 5.. The 3 structure is D shown in Fig....7: Schematic of 3D structure fabricated by wafer bonding using Cu pad thermocompression....................107 xxiii ......... 5.. etc.... 5......99 Fig.......105 Fig..
` xxiv .
Consequently. while the rapid increase in functional density and chip size leads to longerdistance communication in the global levels. due to longer wires and smaller wire pitch. between devices and between circuit blocks is becoming more complex and a challenging problem. The connection of miniaturized and closely packed transistors requires re duced wire crosssections in the local levels. the interconnect delay.Chapter 1 Introduction 1.1 Background The unprecedented prosperity of information technology has driven semiconductor industry to evolve at an incredible rate. greater number of transistors. supported by onchip interconnects. 2]. With the aggressive scaling of VLSI technology. The growing demand for higher performance in integrated circuits (ICs) results in faster device switching speed. the communication. increased functional density and larger chip size [1. now plays a key role and becomes bottleneck in the con1 .
1). As illustrated by the International Technology Roadmap for Semiconductors (ITRS) [5]. interconnect RC delay is dominating the chip performance in advanced technology nodes (Fig.Chapter 1: Introduction tinued improvements in ICs density and speed [3. To mitigate this adversescaling trend of signal propagation (RC) delay. 4]. For instance. 1. hierarchical interconnect structure has been adopted to incorporate several metal levels.1: Predictions for device and wire delays from ITRS 2001 [5]. considerable work has gone into overcoming the interconnect limitations. Relative Delay Technology Node [nm] Fig. This interconnect architecture along with the 2 . so that long global interconnects can be routed to higher tier and maintain larger crosssections to minimize wire resistance [68]. 1.
Cu has replaced Al as the interconnect metal for its lower resistivity [9.1 Background demand for more wiring has resulted in an increasing number of interconnect metal levels.000 421 6 Table 1. In addition. However.2 1 1982 286 1. the interconnect reliability due to electromigration and thermal effects is quickly emerging as a serious integration issue.000. gate stress and leakage [14. interconnects As a result.18 µm 1000 221.5 µm 6 120.1). 10] and low dielectric constant (lowk) materials has been pursued as an alternative to silicon dioxide to reduce interconnect capacitance [11.100. has significantly increased the current density in interconnects. hot electrons.000 217 3 2002 Itanium 2® 0.2) [13]. are farther away from the substrate. the fast increasing power density in ICs (Fig.Section: 1.8 µm 60 3.7 2 1993 Pentium® 0. In addition.000 68. where heat sink is usually attached.1: Evolution of characteristics for Intel® commercial processors (Table 1. 3 . 15]. although all these tremendous efforts have salvaged the attentiongetting interconnect delay problem. Year Processor Technology Node Frequency (MHz) Transistors Chip Area (mm2) Metal Layers 1972 8008 10 µm 0. 12]. together with the scaling trend toward lower power supply (Vdd) to reduce short channel effect. interconnects generally experience higher average temperature than transistors do. 1. Furthermore.2 3500 15.
but also the concern of reliability and thermal effect of interconnect can potentially become another serious system design constraint. 1.2: Evolution of power density for Intel processor [13] (Courtesy of Intel). it is imperative to understand and analyze the emerging VLSI “hot wire” phenomena to handle the impact of interconnect thermal effect in early design phase. as semiconductor technology strives to keep up with Moore’s Law [1]. It is not difficult to imagine that not only will interconnect delay be a major showstopper for continuous ICs performance advancement. 1.2 Thermal Implication of Interconnect Scaling Thermal effects are an inseparable aspect of electrical power distribution and signal transmission through interconnect nets due to Joule heating (or selfheating) caused by 4 .Chapter 1: Introduction Sun’s Surface Rocket Nozzle Nuclear Reactor Fig. In short.
starting from the local at the lowest level to the global at the top most levels. in the interconnects is determined by the product of Joule heating power dissipation and the thermal impedance from the wires to the substrate. The global interconnects are responsible for long distance communication on a chip and have a larger cross sectional area to minimize resistance. on top of the substrate temperature. the local interconnects have the shortest range and the tightest cross sectional dimensions. These are the signaling interconnects. The higher level interconnects are used for global routing and they can span across an entire chip and may be as long as 23 cm in modern highperformance chips. Traditionally. Modern ICs have multiple levels of interconnects to accommodate their large numbers. Interconnects can be further subdivided into three categories according to the range of their lengths and their cross section dimension.2 Thermal Implication of Interconnect Scaling current flow in the wire. 1. Vias are the metal fillings enabling interlevel wire connections.Section: 1. The interconnects (or wires) in a modern IC chip can be broadly characterized into three groups according to the functions they perform. Before proceeding to further discussion. The lower level interconnects are used in local routing and they connect transistors that are a few microns apart. This is depicted in Fig.3 [5]. The wires are separated by inter layer dielectrics (ILD) from level to level and isolated by intermetal dielectrics (IMD) within the same level. These are the local. The temperature rise. the clock distribution interconnects and the power and ground supply distribution interconnects. Whereas. semiglobal/intermediate and the global interconnects. silicon dioxide (SiO 2 ) has been 5 . it is instructive to be familiar with basic onchip interconnect structure and terminology.
1. 6 . 1. However.Chapter 1: Introduction Fig. these lowk dielectrics inevitably have very poor thermal properties (Fig. the choice of the d ielectrics in ICs for its superior properties. the increasing thermal impedance further exacerbates temperature rise in interconnects. Compounded with the poor thermal conductivity of such materials and more metal levels added with more advanced technology nodes.4 [5]). Figure taken from ITRS 2001 [5]. including organic and inorganic films. have been fiercely pursued since a decade ago for the benefit of reducing wire capacitance. as well as the use of porosity and airgaps [1618]. Unfortunately.3: Schematic showing the hierarchy of metal levels for distribution of interconnects in modern ICs. crosstalk noise and dynamic power consumption. a variety of low dielectric constant materials.
4: Both dielectric constant and thermal conductivity of ILD materials decrease with advanced technology nodes [5]. The reduction in supply voltage slows down the power consumption at the transistor level.8 0.2 Thermal Conductivity Thermal Conductivity Dielectric Constant 3 2 1 0 35 50 70 100 130 180 1 0. The interconnect 7 . The typical supply current is in the 100A range in modern chips and is expected to increase for future processors. Although the power dissipated by Joule heating is not a significant portion of the total chip power consumption yet.6 0.2 0 [W/mK]] [W/mK]] Technology Node [nm] Fig. however. the power consumption in the chips is rapidly becoming unmanageable as clock frequency continues to climb and chip temperature has been always a major concern to further chip advancement. 1. the interconnects are separated from the substrate by dielectrics with low thermal conductivity. the resultant increased current requirement leads to greater interconnect Joule heating. and heat cannot be removed efficiently Therefore the temperature rise in interconnects can be nonnegligible.2 Thermal Implication of Interconnect Scaling 4 1. On the other hand.4 0.Section: 1.
performing a more realistic thermal analysis of interconnects is critical. to some extent. overly pessimistic estimaTherefore. the impact of interconnect temperature rise. accurate temperature estimation is extremely important to regulate the design rules and. With the aggressive scaling trend. it is essential not to underestimate On the contrary. tion of the interconnect temperature will lead to overly conservative design. but also constrain the speed performance due to limited current drive capability. thermal effect will limit the maximum allowable current density in the wires to limit the temperature rise in interconnects. the resultant greater temperature rise. It can be envisioned that. 20]. 1.3 Via effect Since the lifetime of interconnects has an exponential dependence on the inverse metal temperature owing to electromigration [19. 8 . hence. the thermal effect may not only be a major reliability concern.Chapter 1: Introduction structures with decreasing critical dimension will inevitably experience everhigher current density and.
Consequently. 1. Poor thermal conductivity of lowk insulators has been a major concern to cause substantial rise in interconnect temperature. the effect of vias (Fig.5). Neglecting via effect on interconnect temperature estimation can lead to intolerable prediction errors. which have much higher thermal conductivity than the dielectrics and therefore can serve as efficient heat dissipation paths. However.Section: 1. has not been adequately addressed in those simplified thermal model [23. This research proposes both compact analytical models and fast SPICE based 3D electrothermal simulation methodology to 9 .3 Via effect Metal wire Transistors Metal via Fig.5: SEM photo of local level interconnect Several publications have addressed the issue of lowk dielectrics and their impact on interconnect temperature. reliability and performance [21. if via effect is ignored. the predicted temperature rise of interconnects can be much higher than the reality. 1. 22]. 24].
which will enable us to quantify the impact of via effect and various interconnect parameters on the interconnect temperature rise.4 Dissertation Organization With the pursuit of realistic evaluation in mind. The impact of Joule heating on the scaling trends of advanced VLSI interconnects has been evaluated in detail. the thermal characteristics of various Cu/lowk schemes are quantified and their effects on electromigration reliability and interconnect delay is determined. The model is then applied to estimate the temperature rise of densely packed multilevel interconnects. 1. The results demonstrate excellent agreement with experimental data and those using Finite Element (FE) thermal simulations (ANSYS). The effect of vias. is included in our analysis for the first time to provide more accurate and realistic thermal diagnosis. 10 . Chapter 2 derives the analytical interconnect thermal models to provide building blocks. Through a combination of extensive electrothermal simulation and 2D field solver for capacitance calculation. via effect is included in all the thermal models and simulations.Chapter 1: Introduction characterize thermal effects due to Joule heating in high performance Cu/lowk interconnects under steadystate and transient stress conditions. Chapter 3 discusses the impact of Joule heating on the scaling trends of advanced VLSI interconnects. as efficient heat transfer paths to alleviate the temperature rise in the metal wires.
4 Dissertation Organization Chapter 4 describes a compact 3D electrothermal simulation methodology to evaluate interconnect design options from a thermal point of view. Furthermore. this simulation methodology provides an efficient diagnosis tool for transient thermal stress analysis. performance. tradeoffs among power. mostly unsolvable. which is usually extremely cumbersome. Use of dummy thermal vias as additional heat sinking paths is evaluated to alleviate the thermal impact on interconnect reliability. Furthermore. in analytical forms.Section: 1. 11 . A complete thermal model including power consumption due to both transistors and interconnect joule heating from multiple strata is presented. Chapter 5 presents detailed thermal analysis of high performance three dimensional (3D) ICs under various integration schemes. Chapter 6 summarizes the main findings and concludes this dissertation with suggestion on possible future work. chip real estate and thermal impact for 3D ICs are evaluated.
Chapter 1: Introduction 12 .
lower wire resistance and higher current density requirements have prompted an industrywide shift from aluminum(Al) to copper(Cu). heat dissipation in the interconnect level has become more difficult. Since these lowk dielectrics also have lower thermal conductivity than silicon dioxide. VLSI scaling trends. such as increasing number of interconnect metal levels and increasing current density. have caused everprofound thermal effects in the interconnects.Chapter 2 Analytical Interconnect Thermal Model 2.1 Motivation The conventional aluminum(Al)/silicon dioxide (SiO2) interconnect scheme reached its performance limitation a decade ago [3]. Similarly. Furthermore. Hence. In a stateof –art chip. the interconnect temperature can rise more than 100 degrees above the ambient 13 . the minimization of interconnect capacitance is driving the introduction of a variety of lowpermittivity dielectric materials.
2. the temperature profile of a single interconnect is difficult enough to be obtained experimentally for its miniature scale [25]. In general.Chapter 2: Analytical Interconnect Thermal Model temperature due to Joule heating and the heat dissipation from transistors in the substrate level. Although knowing the interconnect temperature is the first step toward any kind of further thermal effect analysis. I am Hot ! M8 via M7 M6 Substrate Heat sink Fig.1: Typical multilevel interconnect structure with heat sink attached to substrate. interconnect thermometry based on temperature dependent electrical resistivity of the wire is used to calculate a spatially averaged 14 . not to mention multilevel measurements.
31] have proposed simplified formule.1 Motivation temperature rise of the interconnects [26]. where kILD is the 15 . On the other hand. However. However. most of them only consider a single heat source without considering the effect of vias. The impact of vias has been modeled using i) a characteristic thermal length representing the distance along a metal line through which vias effectively remove the heat. which is not the realistic case for most interconnects. In addition. Analytical thermal models are desirable to facilitate quick estimation of temperature rise in order to provide better insight and thermal design guidelines for advanced interconnect structures.eff. the predicted temperature rise of interconnects can be much higher than actually observed. with kILD. However. the effect of vias. and ii) an effective thermal conductivity of interlayer dielectrics (ILD). kILD. the complicated nature of these expressions makes it very difficult to easily apply them to multilevel metal layers. while Chen [23] and Rzepka [32] applied 3D finite element analysis to estimate temperature rise on multilevel interconnects.eff = kILD/η. This chapter presents compact analytical thermal models for estimating the temperature rise of multilevel VLSI interconnect lines incorporating the effect of vias as efficient paths for conduction of heat. Consequently. various papers have been devoted to deriving analytical expressions for the temperature distribution in IC chips [2729]. it can provide neither the local temperature rise profile on the wire nor the temperature distribution along the multilevel interconnects.Section: 2. Recently. Im [24] and Hunter [30. as an efficient path for conduction of heat has not been included in those simplified thermal model and simulation.
2. To provide a baseline. The main results are summarized in Section 2. we apply the model to multilevel interconnects and discuss the temperature distribution along metal layers.. It is shown that in multilevel interconnect arrays. In Section 2.2 Preliminary Observations The fact that the interconnect temperature does not increase inversely proportional to the nominal thermal conductivity of dielectrics suggests that the vias. Both the spatial temperature profile along the metal lines and their average temperature rise can be easily obtained using these models. providing more comprehensive and accurate results. which have 16 .6. Sections 2. via density along the lines can significantly affect the temperature rise of such interconnect structures.. The remainder of this chapter is organized as follows.3 and 2.5.2.4 describe the derivation of our analytical model and discuss the impact of via effect.Chapter 2: Analytical Interconnect Thermal Model nominal thermal conductivity of the ILD and η is a physical correction factor representing the heat removal by vias. The predicted temperature profiles are shown to be in excellent agreement with the 3D finite element thermal simulation results. The model is then applied to estimate the temperature rise of densely packed multilevel interconnects. The influence of via self heating and thermal impedance of vias is considered here. in Section 2. we validate that it is a legitimate approach to consider via effect.
The metal width is taken to be 0. it may raise a concern that the resultant higher current density in the vias may generate significant heat. However.2 Preliminary Observations much higher thermal conductivity than ILD.Section: 2. The configuration simulated. lower metal levels have much higher via density than the higher levels. deteriorating their effectiveness in heat conduction. 2. In this work. the predicted temperature will be significantly higher than the temperature in practical situations. the via effect is evaluated in details and is incorporated into our multilevel temperature distribution model. A simulation validation was obtained by using ANSYS. 2.2. as shown in Fig. The electrical resistivity and thermal conductivity of Cu are assumed to be 2. The RMS current density of 2. serve as efficient heat dissipation paths.1 First glance at via effect Since the cross sectional area of vias is generally smaller than that of the wires. in general. varied as 250 µm.3 W/mK. L.2. as shown in [33]. Polymer is chosen as the lowk insulator with nominal thermal conductivity of 0. due to the interconnect architecture that. there is danger of significant errors in using overly simplistic assumptions to assess thermal problems in advanced interconnect structures with lowk materials. a three dimensional finite element thermal simulation package.2 µΩcm and 400 W/mK. the temperature rise among metal layers is expected to be quite different from what is predicted by the simplistic model. Furthermore.3 µm and both of the metal height and via height are 0.1e6 A/cm2 is flowing in 17 . As a result. 100 µm and 5 µm. If the via effect is not properly considered. the temperature rise in the via is not as high as that in the wire.6 µm. is Cu wire/via embedded in a lowk dielectric with via separation.
where Tm(x) is the wire temperature. Rth. condition (a). T0 . 2. Adiabatic boundary conditions are applied to the top and four sides of the block where symmetry boundary conditions are satisfied. A uniform temperature. only the thermal impedance caused by vias. L represents the separation between vias.2: Heat generated by Joule–heating is conducted downwards through both ILD and vias. is assigned to the underlying layer of the block.3. L= via separation wire Via via MN M via Tm(x) ILD via Heat conduction T0 MN1 1 MN Fig. Therefore. v. The temperature rise of wires on top of substrate temperature is represented by θm(x) ≡ Tm(x)T0. Two extreme simulation conditions have been carried out in Fig. while the extra heat generated in the via due to current 18 . and with all the current flowing through via. with no current flowing through the via. is included in the analysis in condition (a). condition (b).Chapter 2: Analytical Interconnect Thermal Model the wire. 2.
2. 19 . in both cases. It should be noted that.2 Preliminary Observations x=0 x=L/2 current (a) x=0 x=L/2 current (b) Fig. the via helps transfer heat to underlying layer and reduce the temperature rise on the wire. so there is no heat generation in the via.Section: 2. (b) all the current flowing through the wire will also flow through the via and cause heat generation.3: Two extreme simulation conditions: (a) no current flows through via.
in the following sections. the simulations show that the temperature rise in the vias is minimal. Furthermore. as long as the vias are connected to lowertemperature underlying layers. Even though percentage wise. In addition. is also taken into account in condition (b). Jv. It appears that within a certain range from via. all the vias are not conducting current at the same time but they always help dissipate heat. The comparison of temperature distributions along the metal wires are shown in Fig. they will help dissipate heat no matter how far they are away from substrate. 100 µm and 5 µm. it is legitimate to consider vias as efficient thermal sink paths. we will quantify via effect and evaluate if vias are always helpful. With the above preliminary observations in mind.Chapter 2: Analytical Interconnect Thermal Model density flowing in the via. Finally. the absolute value is quite small. The vias are simply too short to produce much heat. 2.4 (a). under normal operating conditions.4 (a) and (b). the temperature reduction with the help of vias is particularly consumptious. with via effect completely ignored are also plotted in Fig. the average temperature as well as the peak temperature in the wire is reduced. 20 . 2. 2. The temperature profiles of the simplified 1D thermal model. It has the same value as well for the case of 5 µm via separation. (b) and (c) for via separation of 250 µm. with shorter via separation. all the vias do not carry the same amount of current as in the wire. although it is not plotted in Fig.4 (c) to keep other curves more readable. respectively. Consequently. constant temperature value. It should be noted that. the via contact exhibits a major contribution to the temperature rise in the wire for short via separation. In practice.
100 µm (middle) and 5 µm (bottom) for case (a) and (b) of Fig.2 Preliminary Observations Temperature Rise.05 0 0 0.v included 20 30 40 50 (b) Location along Interconnect [um] Temperature Rise. It can be observed that there is no significant difference between the two simulation conditions and temperature rise in the via is minimum. ∆T [°C] 7 6 5 4 3 2 1 0 0 25 50 75 100 125 only Rth.2 0.v included (a) Location along Interconnect [um] Temperature Rise. 21 .1 0.v (c) Location along Interconnect [um] Fig.Section: 2.5 2 2.3.15 0.v only R included th. ∆T [°C] 0.5 both Jv and R included th.5 1 1. 2.v included via effect ignored both Jv and Rth. 2.4: Temperature distribution along the metal wire with via separation of 250 µm (top).v included only Rth. ∆T [°C] 7 6 5 4 3 2 1 0 0 10 via effect ignored both Jv and Rth.
qx = km ∂T Ac ∂x (2.3 Interconnect Thermal Model and Assumptions 2. respectively. Consider a rectangular metal wire with crosssectional area Ac. leaving. and generated in the control volume are Ėin. The temperature gradient along y and z directions is negligible for a long metal wire with relatively small cross section and high thermal conductivity. . neglecting higher order terms. and periphery p embedded in a dielectric medium as shown in Fig.5. The conduction heat rate perpendicular to the control surface at x is indicated by the term qx.2) where km is the thermal conductivity of the metal wire and Ac is the crosssectional area of the wire. . 22 . Ėout.1) where the thermal energy entering. The rate of change of energy stored within the control volume is designated as dEst/dt. 2. the first principle of energy conservation law must be satisfied at any instance.Chapter 2: Analytical Interconnect Thermal Model 2. The conduction heat rate at the opposite surface can then be expressed as a Taylor series expansion where.3. The temperature variation can be modeled as a onedimensional situation which refers to the fact that only one coordinate (x) is needed to describe the spatial temperature variation. and Ėgen. E in + E gen − E out = . Garden has proven that the error due to this simplification is less than 1% even for larger cross sections [34].1 Energy conservation law In either steady state or transient condition. dEst dt (2.
shows energy balance. from a conducting metal wire embedded in a dielectric medium.3) dqcon represents the heat conducted to the surrounding medium. q x + dx = q x + ∂ qx dx ∂x (2.5: A differential control volume.4) 23 .Section: 2. It is proportional to the difference between the surface and the surrounding temperatures and the contiguous surface area.3 Interconnect Thermal Model and Assumptions y z x dx w tILD dAs=pdx Ac qx q”’ dx qx+dx dqcon Fig. 2. dqcon = h( S 'm dAs )(Tm ( x) − T∞ ) (2.
heat convection to the ambient air is ignored by the application of the adiabatic boundary condition on the four side walls and top of the chip. measures the ability of a material to store thermal energy. commonly termed the volumetric heat capacity. Therefore. S’m is the spreading factor to model the multidimensional heat spreading phenomena. heat radiation is simply too small to be taken into account.Chapter 2: Analytical Interconnect Thermal Model where the proportionality constant. is termed as heat transfer coefficient.2 Conduction dominant heat transfer Since.3. In addition. The product. Tm(x) is the metal temperature at x and T∞ is the temperature of the medium far from the metal surface. the governing heat diffusion equation with temperature variation dominating in x direction can be concluded as ∂T ∂ ∂T [km Ac ] − hS 'm p (T − T∞ ) + q ' ' ' Ac = ρ c p Ac ∂t ∂x ∂x (2. Therefore. only heat conduction will be considered as 24 . ρ cp [J/m3K]. thermally insulated package materials encapsulate the IC chip. the heat entering the differential control volume (qxqx+dx) plus the volumetric rate of thermal energy generation in the control volume (q’’’) minus the heat dissipated to the medium (dqcon) must be equal to the rate of change of thermal energy stored within the differential control volume. in general.1. h. 2. As required by energy conservation law Eq. 2.5) where ρ is the metal density and cp is the metal specific heat. dAs is the surface area where heat in the control volume can be transferred into the medium.
8) 25 .3 Interconnect Thermal Model and Assumptions the prevailing heat transfer mechanism.6) where q” [W/m2] is the heat flux and k [W/mK] is the thermal conductivity. the substrate. 2.5 results in ∂T ∂ ∂T k [km Ac ] − ( D )( S 'm w)(T − T∞ ) + q ' ' ' Ac = ρ c p Ac ∂t ∂x ∂x t ILD (2. Furthermore.Section: 2. Eq. the majority of heat will transfer through the substrate [35]. 2. i. is assumed to be the sole heat dissipation path to the outside ambient. kd / tILD acting as the proportionality constant and can be treated as an equivalent conduction heat transfer coefficient. the net heat transfer processes can be quantified by the rate equation known as Fourier’s law.e. dqcon = ( kd )( S 'm wdx )(Tm ( x) − T0 ) t ILD (2. with heat sink attached.7) where kd is the thermal conductivity of the dielectric medium and tILD is the underlying dielectric thickness.6.4 can then be further simplified if we consider heat is only transferred downwards and only heat conduction mechanism takes place. Accordingly. w is the width of the metal wire and T0 is the temperature of underlying layer. Therefore. q " = − k∇ T (2. to which a heat sink is usually attached. hence only heat conduction toward the substrate will be considered in our interconnect thermal modeling. 2. Eq. This assumption is validated in Fig. as Zhou has shown that.
the majority of heat will conduct to outside ambient through substrate rather than to printed circuit board (PCB) [35]. 26 . to PCB thru. substrate Fig. particularly with fan.Chapter 2: Analytical Interconnect Thermal Model Flip Chip Assembly PCB Heat sink 100 % heat dissipated 80 60 40 20 0 no heat sink fin heat sink fan heat sink cond. 2.6: With the help of heat sink.
such as under electrostatic discharge (ESD). electrical resistivity ρm and thermal conductivity km. will be discussed in chapter 4.7). Transient stress analysis. representing parallel metal wire array. separated from the underlying layer by interlayer dielectrics (ILD) of thickness tILD and thermal conductivity kILD. The two ends. at x=±L/2. width w. of the wire are connected to the underlying layer through vias (Fig. The vias are modeled L W H tILD Tv(y) D d x=0 Tm(x) T0 y=0 Fig. length L. consider a rectangular metal wire with thickness H.Section: 2. The thermal environment on the interconnects is considered to be under steadystate condition. 27 .7: Geometry used to derive interconnect thermal model. we will develop the interconnect thermal modeling under normal chip operation circumstances. 2.3 Interconnect Thermal Model and Assumptions 2. To begin with. 2.3.3 Steady state thermal modeling In this section.
under steady state conditions. respectively.9) and d 2θ v ( y ) θ v ( y ) q ''' − 2 =− v .12) where Sm and Sv are the shape factor per unit length of the wire and the via. jrms. It is assumed that heat only flows downwards toward silicon substrate which is usually attached to a heat sink.v are the thermal healing length of the wire and the via.m k A ≡ m m k ILD S m k A ≡ v v k ILD S v 1 2 1 2 k wH 2 = m .Chapter 2: Analytical Interconnect Thermal Model as round pillars. the governing heat equations for the wire and the via are '' d 2θ m ( x) θ m ( x) qm' − 2 =− . (2.11) and L H . thus.10) where qm’’’ (=j2rms. The underlying layer temperature set to be T0. k ILD S m 1 2 1 (2.v k (πD 2 4) = v k ILD S v . flowing in the conductors. LH. dx 2 LH .mρm) and qv’’’ (=j2rms. The electrical resistivity and thermal conductivity of vias are ρv and kv.m and LH. with diameter D and height tILD. and therefore temperature at the via bottom Tv(0)=T0. 2.vρv) are the volumetric heat generation in the wire and the via due to Joule heating. defined as L H . dy 2 LH . with uniform rootmeansquare current. The temperature rise of wires and vias is represented by θm(x) ≡ Tm(x)T0 and θv(y) ≡ Tv(y)T0. to represent heat spreading and. Following Eq. m km (2. v kv (2. respectively.8. to correct the deviation from onedimensional heat 28 .
LH. from vias. the two boundary conditions to solve Eq. x =0 θm(x=±L/2)= θJ.m ) qm' L2 . As it will be seen later.m ) 1 − . sinh t ILD L H .m.v ] 1 − .10 as θ m (x ) = θ J '' cosh (x LH . θv(y=0)=0. heat will flow through ILD and the via effect is diminished.m is that within the range of thermal healing length. dθ m ( x ) dx = 0.Section: 2. and the junction temperature θJ at the two ends of the wire (x=±L/2).9 and Eq. (2. (2. shape factor Sm and spreading factor S’m are interchangeable and both describe the multidimension heat conduction flow from metal surface to surrounding medium.v ( ) (2.13) In addition.v ) + sinh[(t ILD − y ) / LH .15) θv ( y ) = θ J sinh( y LH .10 are the temperature rise θJ at the top of the via (y=tILD) and zero at the bottom of the via (y=0). The two boundary conditions used to solve Eq.m (2. v ) q ''' L2 + v H . 2. cosh (L 2 L ) H . The physical meaning of LH.v sinh(tILD LH .16) 29 .m.m ) km cosh (x LH .3 Interconnect Thermal Model and Assumptions flow.m H + cosh (L 2 LH . 2.v ) kv sinh ( y LH .9 are the adiabatic condition at the middle of the wire (x=0) due to symmetry. 2. It should be noted that the shape factor is only a function of geometric parameters. heat generated will flow through vias to the underlying layer. with θJ=θm(x=±L/2)=θv(y=tILD). Beyond LH. 2.14) The temperature rise along the wire and within the via can then be solved from Eq. θv(y=tILD)= θJ.
θJ can then be solved as θJ = '' Am LH . (2.m2/km = jrms. v + ' Av LH . 2. v ) − csc h(t ILD LH . Thereafter. It should be noted that we should certainly include the temperature coefficient of resistivity (TCR) for electromigration tests where very high current densities are employed.15 as ∆T1D.17. k m Am kv Av tanh( L / 2 LH . 2.m2/km (= jrms. The ratio of θJ/∆T1D has other important indications as will be explained in a later section. from Eq. m ) + v v coth(t ILD / LH .Chapter 2: Analytical Interconnect Thermal Model with L/2≤ x≤ L/2 and 0≤ y≤ tILD.m2ρmHw/kILDSm) in the second term of Eq.15 can be interpreted as the combined contribution from via selfheating and the thermal impedance caused by the via (first term) and the Joule heating in the wire corrected with the help of via effect (second term). m ) k m Am k A tanh( L / 2 LH . v qv''[coth(t ILD LH . 2.17) Substituting Eq.18) Effect of variation in electrical resistivity ρ with temperature was found to be small for practical situations and is ignored in this model. v (2. It is simple to do so because 30 . 2.m2ρmLH.16 into Eq. m LH . which can be shown as the continuity equation − km Am dθ m ( x) dx x =± L / 2 = kv Av dθ v ( y ) dy y = t ILD . v ) LH . the wire temperature rise.15 and Eq. θm(x). we need to realize the continuous flow of heat at the junction of the wire and the via. It is instructive to note that we can define qm’’’LH. v )] . 2. m LH . To determine θJ. m qm' tanh( L 2 LH . v ) LH . m ) + coth(t ILD / LH . which represents the temperature rise in the wire obtained from simplified 1D heat conduction model where via effect is not included.
there are multiple heat sources due to parallel metal wire array. in typical IC layout. A new expression of heat spreading factor. This is due to the fact that it assumes a single heat source. The last pieces of the puzzle in this model are the shape factors Sm and Sv. Several papers have been devoted to study the case under electromigration test condition [36. however. Therefore. As shown in Fig. Assuming the lateral spreading to increase linearly with vertical coordinate. whereas. For the worst case scenario. w + 2 y 2k ILD w (2. The commonly used Bilotti’s equation [29] is not adopted in this work to account for the deviations from 1D heat flow.8. can be derived as Rspr = ∫ d 2 1 k ILD 0 1 dy w+d = ln . Rspr. is therefore derived here for uniformly separated infinite number of parallel wires and will be used in the following analysis throughout this work.19) 31 .Section: 2. all metal wires are assumed to carry the maximum RMS current density and separated by spacing d. is to incorporate the via effect and to predict temperature distributions along multilevel interconnects under normal operation conditions. the Joule heat transfers downward as well as spreads laterally in the ILD. the resistivity used in this work is assumed to be a constant at ρ(TDie) to a first order approximation. 37]. which accommodate 3D heat spreading and different boundary conditions. The focus of our work.3 Interconnect Thermal Model and Assumptions electromigration tests are usually conducted on isolated wires and not multilevel structures. 2. S’m. the spreading thermal resistance.
Rth .8: Geometry used for calculating Rspr. Cross sectional view is shown on the left hand side.ILD. 2. (2. Then. Quasi 2D heat conduction is used to correct 1D heat conduction model.Chapter 2: Analytical Interconnect Thermal Model where d is the spacing between wires and is shared by the two wires for heat dissipation. can be calculated by combing the spreading resistance with the volume resistance. the total thermal resistance of ILD. 32 . ILD d 1 w + d 1 t ILD − 2 ln = + 2k ILD d k ILD w + d . Space between any two wires is shared for heat dissipation. Rth. Rrect and the spreading factor S’m.20) 1D Model Quasi 2D d w d/2 d/2 heat spreading d Rspr Rrect Metal Wires w tILD d/2 tILD heat dissipation path Fig.
ILD (2. Sm = ( w t ILD ) S 'm (2. Rth. the conversion between Sm and S’m can be readily obtained as.ILD can be manipulated as. k ILD weffective k ILD wS 'm (2. ILD = t ILD t ILD = . 2.23) where q(x) is the heat transfer rate from the metal wire to the dielectric medium.ILD can be also expressed as Rth .Section: 2. Rth.21 and Eq.21) By comparing Eq. ILD = 1 S m k ILD (2.24.3 Interconnect Thermal Model and Assumptions On the other hand.20 and Eq. 2. S’m can be obtained as d w 1 w+d w t ILD − 2 .26) 33 . q( x) = S m k ILDθ m ( x) = θ m ( x) Rth .24) Again. Sm 1 w + d t ILD − d / 2 = [ ln ] + w+d 2 w −1 (2. 2. By the definition of shape factor [38].25) and thus. .22) The next step is to convert spreading factor S’m to shape factor Sm. . 2.21. S 'm = + t 2 w t w+d ILD ILD −1 (2. Rth . Hence. by comparing Eq.
(2.27) After the shape factors are installed in Eq.m = 3. w = d = 0.11 and Eq. all the equations derived here can still be valid as long as the appropriate Sm is determined by either an appropriate analytical expression or extracted from simulation. Sm could have different expressions than Eq. ANSYS Oxide. S v = 2π /[ln( 4t ILD / D )] .7E6 A/cm . depending on different layout and operating conditions.Chapter 2: Analytical Interconnect Thermal Model It should be noted that. ANSYS Fig.12.26. 2. 2 34 .3 µm. and jrms.8 µm. The Sv is adopted from [38] for a vertical cylinder in a semiinfinite medium attached to a constant temperature surface.3 W/mK and koxide =1. 2. Analytical Oxide.2 W/mK are assumed here. Analytical Polymer. the effect of the via separa Polymer. 2. kpolymer = 0. However. 2.9: Temperature profile along the Cu wires with 100 µm via separation. H = tILD = 0.
∆Tave is defined as the average temperature rise in one metal layer and it can be expressed as 2 2 ∆Tave = qm Rth.4 Impact of Via Effect on Effective kILD tion and heat spreading on the temperature profile along metal wire can be captured completely by Eq. 2. 2.eff. koxide =1.4 Impact of Via Effect on Effective kILD 2. 2. LH.28) On the other hand. By defining a via correction factor (η). the via effect is more important for lowk insulators. the thermal healing length.15.1 Via correction factor It should be noted that as predicted from Eq.15 as ∆Tave = θ m. ∆Tave can also be obtained from Eq. kILD.Section: 2.9.s. which can then be used in place of the nominal kILD in the conventional thermal equations.9. the via effect can be incorporated into the effective thermal conductivity of ILD.2 W/mK).eff Sm (2.eff incorporating the via effect is now derived here. ILD = ( I m Relec. 2. ILD = jrms. kILD.3 W/mK v. An analytical expression for kILD. (kpolymer = 0. Consequently.11 and validated from Fig. 2. in a wire is longer if the ILD has lower thermal conductivity. As can be observed form Fig. m ) Rth. 2.m. the result from analytical expression is shown to be within 3% agreement with the 3D finite element thermal simulation using ANSYS.4. m ρm Hw .ave = 1 L2 1 L2 ∫− L 2 (T (x ) − T0 )dx = L ∫− L 2θ m ( x)dx L 35 .
2. η. 2.m ρ m Hw ( ) LH .m LH .m H . as indicated by Eq.v ρ v S m Av LH . 2. hot spots can occur within the vias if η>1.29.30) k ILD .31) (2.<1.m 2 = jrms . therefore.v (2.34) 36 . L / 2 LH .31 and Eq. η>1 may occur and via effect will be detrimental to the wire.28 and Eq.m 2 t ) cosh( ILD ) < 1 . This critical condition can be further shown to be qm''' k v LH . if via self heating and thermal impedance caused by the via is excessive. alleviate interconnect temperature rise if η<1.30.2 Hot spot location As implied by Eq.m ) .m (2. With θJ/∆T1D.29) Comparing Eq.m ρ m S v Am t ILD ) <1. via correction factor.v (2. via effect will help increase the effective thermal conductivity of the ILD and. The essential criterion to differentiate these two domains is to evaluate the ratio of θJ/∆T1D.4. 2.33) or 2 jrms . 1 − ( ''' )( )( q v k m L H . 2.m k ILD S m (2.Chapter 2: Analytical Interconnect Thermal Model L2 L/2 tanh tanh L L θJ Hw H .eff = k ILD / η and (2. 2.32) ∆Tave = ∆T1− D η .32. 1 − ( 2 )( )( )( ) cosh( jrms . can be deduced as η = 1 − (1 − which yields two important results.m ρ m + (1 − ) 2 L2 L/2 k ILD S m jrms .v L H . On the other hand. θJ ∆T1− D ) tanh(L / 2 LH . In addition. a beneficial via effect is guaranteed.
the longer the LH. Dielectrics with lower nominal kILD experience lower η even with longer via separation. 37 . and hence.4 Impact of Via Effect on Effective kILD Fig. with two different D/W ratios. η. 2. stronger via effect. the stronger is the via effect.10: Via correction factor. 2.Section: 2. Wires with shorter via separation feel Fig.11: Effective ILD thermal conductivity increases with decreasing via separation. The lower the nominal kILD.m. are plotted for three types of dielectrics.
D/w.Chapter 2: Analytical Interconnect Thermal Model From Eq. η and kILD. 2.eff are plotted against via separation with two via diametertowire width ratios. This fact can explain why the interconnect temperature is not as high as commonly assumed when lowk ILD is implemented in the advanced interconnect structure. To illustrate the effect of the via. various interconnect parameters can be quickly evaluated to determine the hot spot location (in the wire or within via) and the nature of via effect (alleviate or enhance wire temperature rise). polymer and air are about 5µm. 2.33 and Eq.11. It can be observed that. Geometries of the interconnect structure are taken from ITRS [5] 65nm technology node for global wires. 2. 2. as shown in Fig. so that the hot spot locates in the middle of the wire and the via effect is always beneficial. In this case. under the condition η<1. short via separation suffers more heat backflow from the vias. 2. in the typical VLSI interconnect structure.1. 2.5. 2. 10um and 30um as calculated from Eq.12. respectively. in Fig. for the case of D/w=0. D/w is generally greater than 0. However. incorporation of the via effect results in increased kILD. 2.10 and Fig.10. Fig. The temperature profile along the interconnect is shown in Fig.34. the LH.11.11 also shows that the via effect diminishes rapidly for the portion of wire beyond thermal healing length from each end. the cross section of the via is much smaller than that of the wire which results in substantial via heating. 38 .m’s for oxide.eff especially for ILD materials with lower nominal thermal conductivity. On the other hand.
This assumption is fairly legitimate due to the fact that the chip is usually encapsulated with insulation materials.Section: 2.5.7 1 10 20 30 40 50 Location along Interconnect [um ] Fig. The hot spots usually locate in the wire except for extremely small via dimensions.5 Temperature in Multilevel Metal Layers 2.5 0 0 via D/W D/W D/W D/W D/W D/W = = = = = = 0. Therefore.5 2 1.5 1 0. 2.1 Multilevel interconnect formula Following the previous argument. i defined as the average 39 .5 Temperature in Multilevel Metal Layers Temperature Rise [°C] 3 2.2 0.12: The figure shows the temperature profile along the wire from the middle of the interconnect to the via at the right end. With ∆Ti1. all the heat generated in the upper metal levels has to transfer through the lower metal levels to the substrate.5 0.1 0. 2.3 0. Joule heat generated in metal wires is considered to dissipate only through the heat sink attached to the Si substrate.
the temperature rise at the top layer for an Nlevel metal interconnect can be obtained as ∆TN = TN − Tsubstrate = ∑ ∆Ti −1.13. is the accumulated heat dissipated through the ith level interconnects and αi is the metal coverage of the ith level metal layer. there is hardly any temperature rise in the lower levels even when 40 . there is more heat flowing through the lower levels since Qi represents the sum of all the heat generated from ith layer to Nth layer. with the via effect considered. In addition.1e6 A/cm2 is assumed for all wires and 50% metal coverage is assumed for each metal layer. the overall temperature rise is much lower with the help of the vias. As a result. 2.Chapter 2: Analytical Interconnect Thermal Model temperature rise between metal layers i1 and i. For the case when the via effect is neglected. jrms. Second.36) where Qi.35) ≅∑ i =1 N N wiηi 2 ∑ jrms.i Qi i =1 i =1 N N (2. First. i = ∑ Rth . As can been seen from Eq. it can be observed that the temperature distribution among metal layers is quite different in these two cases. On the other hand.i Sm . of 2. ηi is set to be 1. 2. jα j H j k ILD . some reasonable values of via separation are assigned to each of the 10 metal layers with polymer used as the ILD in a 65 nm technology node structure. a worst case current density.iα i j =i (2.35. as shown in Fig. j ρm. For the following demonstration of the importance of the via effect. Ignoring the via effect results in large temperature jump in the lower layers and then the temperature rise levels off. a substantial temperature rise will occur in local wires if the effect of the dense via population is not taken into account.
Section: 2. 2. based on 65nm technology node.respectively. In the case of via effect included.14. Therefore. This phenomena can 41 . global interconnects are more problematic . kpolymer is only one fourth of koxide. 100. Most of the temperature rise is attributed to the upper metal layers with long via separation. the via separations assigned to the metal layers. 30.5 Temperature in Multilevel Metal Layers 45 40 35 30 25 20 15 10 5 0 1 2 Temperature Rise ∆T [°C] via effect ignored via effect included 3 4 5 6 7 8 9 10 Metal Layer Number Fig. 10. curve (a). 2.5. 2.2 Interconnect temperature rise trend The trend of interconnect temperature rise due to Joule heating is investigated in Fig. The concern of RC delay in the global wires may get worse with this additional temperature effect.13: Temperature rise distribution along metal layers from substrate to top metal level. 200. 150. It is interesting to see that even with increasing current density and lower thermal conductivity from the lowk ILDs used for the more advanced technology nodes. 50. the interconnect temperature rise will reach a plateau and then drop. from 1st to 10th levels. are 5. 300. 500 and 1000 metal pitches. from the thermal design point of view.
(b) j scales with ITRS. The same functionality can be reached within a shorter distance.Chapter 2: Analytical Interconnect Thermal Model Temperature Rise ∆T [°C] 30 25 20 15 10 5 0 130 scenario (a) scenario (b) scenario (c) scenario (d) 90 65 45 22 Technology Node [nm] Fig. while all the other aspects of scaling continue.14: Trend of interconnect temperature rise along technology nodes for different scenarios. Overall. while kILD continues scaling. but kILD stops scaling at 65nm node. (a) both j and kILD scale according to ITRS. Several scenarios are studied here to identify their impact on interconnect temperature.14. 2. (d) both j and kILD stops scaling at 65 nm node be understood by realizing that the via separation keeps reducing due to scaling. (c) j stops scaling at 65nm node. the interconnect temperature rise will be less if either the trend of increasing current density or lowerk materials stops at 65nm nodes. the magnitude of current density and the dielectric constant of the lowk insulators are believed to encounter a bottleneck due to reliability concern at 65nm technology node. 42 . as shown in Fig. On the other hand. 2.
It is observed that global interconnects would suffer much higher temperature rises than local interconnects due to the much longer via separation.6 Summary 2. It will be essential for high performance chip designer being able to comprehend the thermal impact on the circuit performance and reliability implication in the early design stage. a closed form thermal model incorporating the via effect has been formulated to estimate the temperature rise of interconnects in multilevel metal arrays. the thermal problem associated with lowk insulators is not as bad as it appears.6 Summary In conclusion. Beyond the 45nm node closer packing of vias will alleviate the temperature rise problem. Additionally. 43 . Both the via effect and heat spreading have been taken into account to ensure accurate predictions. The interconnect temperature can be substantially lower than that predicted from overly simplified 1D thermal model. It is shown that via effect must be considered in the thermal analysis of interconnect structures. Therefore. the effective thermal conductivity of the ILD materials can be significantly higher than their nominal values if via separation is comparable to the characteristic thermal length.Section: 2. a compact analytical thermal model has been presented to evaluate the spatial thermal distribution and the average interconnect temperature rise under the influence of vias. It has been shown that with the help of vias as efficient thermal paths. This model provides a quick and accurate interconnect temperature rise estimation as well as a comprehensive thermal design guideline.
EM lifetime. most of interconnect failure mechanisms are temperature related including electromigration (EM). Thermal effects impact the interconnect design in the following ways. Accurate integrity estimation of onchip interconnect temperature rises caused by joule heating is essential for high performance chip design because the wire currentcarrying capability is stringently limited by interconnect temperature. For the temperature coefficient of resistivity (TCR) of Cu given in [39]. the wire slows down with higher temperature due to increasing metal electrical resistivity. Second. will be discounted to some extent. the delay reduction expected from the introduction of lowk dielectrics. wire resistance goes up by 510% as interconnect temperature rises by 1020°C.1 Introduction Thermal effects are very important in determining both reliability and performance of interconnects. meantime tofail 45 . As a result.Chapter 3 Thermal Impact on Interconnect Design 3. First. which have invariably poor thermal conductivity.
The temperature and consequent performance/reliability predictions with those overly simplified thermal analysis will deviate even more profoundly with more advanced technology nodes as the via effect is more effective for lowerk dielectrics and the via separation is shortened with scaling [45]. traditional wisdom requires the onchip interconnect temperature to be no more than 5°C above the silicon temperature (100°C for a typical high performance microprocessor) for EM lifetime consideration. decreased exponentially with inverse interconnect temperature according to Black’s equation [19]. the impact of Joule heating on the scaling of deep submicron Cu/lowk interconnects will be investigated in detail.5°C [32].Chapter 3: Thermal Impact on Interconnect Design (MTF). 44]. 41]. The interconnect temperature is commonly limited to ~105°C for electromigration life time considerations [40. the estimated interconnect temperature rise can be much higher than twice as high compared to the realistic case [43. In this chapter. 31] and propose “selfconsistent” analysis for allowed current density [30. [46]. the current density design limits developed by reliability engineers for circuit designers are very conservative. Therefore. without considering via effect properly. our multilevel interconnect thermal model incorporating the via effect enables more realistic circuit timing simulation and reliability assessment without being excessively conservative. Having realized the importance of the temperature rise on interconnects. It is reported that MTF could be reduced by 90% when interconnect temperature rises from 25°C to 52. many efforts have been devoted to estimate the interconnect temperature [24. traditionally. Extending the work from Chapter 2. However. 42]. In addition. However. the prospects of rising 46 .
Section: Thermal Effect on Interconnect Metrics interconnect temperature and the need for greater current carrying capability are not compatible.6 investigates the trend of interconnect temperature rise under various scaling scenarios. Section 3. The remainder of this chapter is organized as follows. we perform a coupled assessment of EM reliability and current drivability for global interconnects for 22130 nm technology nodes. In a later section of this chapter.3 addresses the foreseeable impact from interconnect scaling trend on wire temperature due to Joule heating. a comprehensive EM evaluation methodology is proposed from the view of coupled performance and reliability analysis [45].4 describes the definition of a ‘reasonable’ worstcase scenario of interconnect thermal analysis and applies it to evaluate the thermal effect on Cu/lowk interconnects.7. Finally.2. In Section 3. This 47 .2 Thermal Effect on Interconnect Metrics It suffices to mention here that a nonnegligible and increasing resistance of the wires leads to a j2ρ (j is the current density) power dissipation in the form of heat.5. combined with our new interconnect thermal model. we summarize and conclude in Section 3. the lack of realistic interconnect temperature estimation makes the proposals fruitless. In Section 3. 3. 42]. 30. we discuss how the interconnect metrics are influenced by thermal effect. Section 3. Section 3. Although it has been proposed to consider EM lifetime and wire selfheating simultaneously in generating EM guidelines [23.
CILD. especially if lower dielectric constant materials. 48 .2. 3. Also showing aspect ratio (AR).1 RC delay The delay of the wires can be well approximated by the product of resistance (R) and the capacitance of the wire (C). (Fig. which are also invariably poorer heat conductors. The wire capacitance typically has three components: interlevel. CILD H CIMD AR=H/W W Fig.Chapter 3: Thermal Impact on Interconnect Design raises the temperature of the interconnects above the device temperature. intermetal (within the same level between metal lines). are used. if inductive effects are not important. CIMD.1: Schematic showing the intermetal and the interlevel components of capacitance. 3. To model the RC delay of the wires it is imperative to accurately model both the resistance and the capacitance accurately. 3.1) and the fringe component.
heterogeneous dielectric schemes (with different dielectric materials for ILD and IMD). In terms of interconnect capacitance.3) with AR denotes wire aspect ratio. and can be modeled as. ε represents the respective premittivity for IMD and ILD. CIMD.1) here β being the temperature coefficient of resistivity (TCR). where H is the metal thickness (height) and w is the width of the wire. let us ignore any fringe capacitance for this moment. to gain better insight of the role of CILD and CIMD. we notice that.2) (3. (3. Thickness of the ILD is approximately the same as the height of the metal wire.5e3/K for Al. 49 .3e3/K for Cu and 4. and has the value of 4. Therefore. The total interconnect capacitance per unit length can then be simply expressed as: Ctotal=2(CILD+CIMD). due to the high wire AR. ρm. ρ m (T ) = ρ m (T0 )[1 + β (T − T0 )] (3. the ILD assumes the major contribution to the thermal impedance and IMD being a minimal factor. should be exploited to optimize RC delay and subside thermal effect. The factor of 2 in the denominator for CILD accounts for the overlap with orthogonal wires on adjacent levels. First. such as use lowk dielectrics for IMD and keep silicon dioxide for ILD. the intralevel (linetoline) capacitance. The length of overlap is taken to be half the length of the interconnect based on the assumption that wire width is half the pitch. defined by H/w. where CILD=εILD/2AR and CIMD=εIMDAR. is dominant. Second.Section: Thermal Effect on Interconnect Metrics It is well known that the electrical resistivity. since most of the heat is flowing downwards toward the heat sink.is a function of wire temperature.
leads to depletion of enough material so as to initially increase the wire resistance and finally cause an open circuit [49]. an heterogeneous dielectric approach will be appropriate to lower coupling noise while contain interconnect temperature rise. at a given technology node. The current through metal wires leads to metal atom migration due to momentum exchange between electrons and metal atoms [19]. 48]. Cross talk is proportional to the ratio of the intermetal to the total capacitance [47] of the wire.4 Electromigration reliablity Electromigration is wildly regarded a major failure mechanism of VLSI interconnects[28. Sw is the switching activity factor representing the probability of a particular interconnect switching during a clock cycle. 3. Hence. from the cross talk perspective it is more beneficial to lower just the intermetal capacitance. over time.Chapter 3: Thermal Impact on Interconnect Design 3.4) Here. Thus. V is the voltage to which the interconnect charges and f is clock frequency. Cint is the total interconnect capacitance. Thus. the power dissipation due to interconnects.2. it also causes excess metal atoms to 50 .2. On the other hand. This migration.2. again. is a result of charging and discharging its capacitance and is given by the dynamic power dissipation formula Pint = S w C intV 2 f (3. the interconnect power is heavily dependent on its total capacitance.3 Cross talk noise . 3.2 Dynamic power consumption The second metric of importance.
3. the exponent n is typically 2 under normal operation conductions. temperature and the current density play the most important role in dictating electromigration.5) where MTF is the meattimetofailure (typically for 0. MTF will drop by 50% when interconnect temperature rises from 100°C to 110°C. MTF = A j − n exp(Q / k BT ) (3. The activation energy Q for Cu has been reported in the range of 0. Usually. kB is the Boltzmann’s constant. which influence electromigration can be divided into those related to the physical structure of the metal wire and those related to the conditions of operation. 51 .1% of accumulative failure or 10 years lifetime). the maximum allowable current density is limited by the goal that the electromigration lifetime will achieve 10 years with interconnect temperature maintained at or below 105°C. Among the conditions of operation. For a middle of the range Q value of 0.51eV [5052]. which in extreme cases can cause a short to the adjacent wire through metal hillocks.75eV. and Tm is the metal wire temperature. as modeled by the well known Black’s equation. given by Eq. j is the DC or average current density.5. Certain crystal orientations of thin films (metal wire) are more conducive to preventing electromigration.Section: Thermal Effect on Interconnect Metrics accumulate at a different location along the wire. A is a constant that is dependent on the geometry and microstructure of the interconnect. The primary factors.
ILD Node 130 90 65 45 22 3.1).6 1.7 0. Aggressive interconnect scaling has resulted in increasing current density. So explosive has the growth been.0 2.7 1.8 1.0 1.1 2.45 0.3 Scaling Trend of Joule Heating As VLSI technology advances.6 1. n is the number of layers in each tier.25 0.5 2.3 2.8 2.4 265 1.0 4 4 4 4 4 Semiglobal Tier Wire Wire Via Pitch AR AR [nm] 450 1.7 1.8 [W/mK] n kILD 0.1 1. The growing demand of higher current driving capability and the aggressive shrinking of metal pitch has resulted in significant current density rise in the wires (Table 3. 3.5 Via AR 1.8 1. εr.3 100 2.0 Via AR n 1.6 210 1.8 2.17 2 2 2 2 2 Local Tier Wire Wire Pitch AR [nm] 350 1.1 290 2.6 135 1. more metal levels.1: Interconnect parameters for 130 nm to 22nm technology nodes based on ITRS[5].7 150 1.5 195 1. 52 . the rise of interconnect current density will outpace the average chip power density by a factor of two for high performance microprocessor throughout the technology nodes (Fig.7 105 1.ILD and kILD are the relative dielectric constant and thermal conductivity of the interlevel dielectrics (ILD) for each technology node.2).9 2. interconnects have become the limiting factor to IC chip performance [3].3 Table 3.6 65 2.8 1.36 0.7 1. Wire width is half the wire pitch for all cases if not defined specifically. according to ITRS [5]. Tech.8 50 2. and introduction of low dielectric constant (lowk) materials.Chapter 3: Thermal Impact on Interconnect Design 3.2 205 2.8 n 2 3 4 4 5 Global Tier Wire Wire Pitch AR [nm] 670 2. εr.0 460 2.
Tref. The work presented in this chapter is for the 65nm technology node with all parameters quoted from the newly updated ITRS ’01[5]. The combination of greater heat generation and thermal impedance is leading to a continuous increasing interconnect temperature and consequent impact on wire delay and reliability wires is fast emerging as an urgent issue. The substrate temperature. 3.3 Scaling Trend of Joule Heating 30 5 4 3 Power Density [W/cm2] 25 20 15 Jmax [ MA/cm2] Jmax [ MA/cm2] 2 10 5 0 22 45 65 90 130 180 1 0 Technology Node [nm] Fig. Therefore. interconnect joule heating is becoming nonnegligible. is 53 . the impact of the interconnect “thermal”scaling trend on wire temperature due to Joule heating requires immediate attention. 54]. Chip power density is calculated by total power of the chip divided by chip size.1). thermal conductivity of lowk dielectrics is decreasing rapidly with the reduction of dielectric constant (Table 3. In addition. surface scattering and skin effect [53. the following four layers as semiglobal tier and the remaining layers as global tier. We assign the first two metal layers as local tier. Cu resistivity will increase due to barriers. As a result. Furthermore.Section: 3.2: Trends of chip power density and interconnect Jmax along technology nodes suggested by ITRS [5].
200.4. Case (2) represents an isolated global wire. this work includes the via effect in the analysis to obtain more realistic results. This is especially true for dielectrics with lower thermal conductivities [44. 46]. Reasonable via separations are assigned for each metal level from level one to ten: 5.e. case (1) has the highest temperature rise (∆Tjoule heating). 54 ..4 Temperature Effect on Cu/lowk Interconnects 3. 1000 and 2500 interconnect pitches. 100. and the temperature of the top global wire is Tm= Tref+∆Tjoule heating. Therefore. it is important to identify a reasonable worst case scenario. Although there is no additional heat source between this wire and the substrate. Previous work has attempted to evaluate the thermal characteristics of interconnects neglecting vias as an effective heat conduction path [4]. 300 . 3. 3. Case (1) in Fig. 150. i. 20. Case (3) has the same current condition as case (1) but with vias taken into account for all metal layers. As shown in Fig.4. ignoring vias in heat transfer predicts unrealistically high temperature rises even with moderate current densities.1 Definition of worst case condition To provide robust thermal analysis for interconnects. 3. However.500. the thermal impedance is higher in the absence of lower level metal. all wires flowing the same current density and the via effect ignored.Chapter 3: Thermal Impact on Interconnect Design assumed to be fixed at 100°C (which will be true if twophase microchannel cooler is employed in the future [55] ).3 shows the widely used condition. 50.
3. 180 Temperature Rise.Section: Temperature Effect on Cu/lowk Interconnects But we will disregard this condition as unrealistic as vias are ignored. Current flows through all wires at all metal layers in cases (1) and (3).4: Temperature on the toplevel interconnects rises rapidly with increasing current density. 3. vias Case (1) Case (2) Case (3) Fig. ∆T [°C] 160 140 120 100 80 60 40 20 0 0 c a s e (1 ) c a s e (2 ) c a s e (3 ) C u rren t D en sity . J [M A /cm ] 1 2 3 2 4 Fig.3: Configurations of the three thermally worst case scenarios. Case (3) is worse than case (2) and we will use this condition for the following analysis. 55 .
k=0. k: thermal conductivity [W/Km].2  Conductor Al W 240 180 3. thus higher Parameter k [W/mK] ρ [µΩcm] εr Cu 400 2.54 2.5 compares the thermal performance of several lowk materials.2: Materials properties of various conductors and dielectrics used in this study. 0. εr: relative dielectric constant.5 4 C urrent D ensity .4. k= 0. k [W/mK] is thermal conductivity.7 Oxide 1. 2.03 1.17 1. interconnects with lower dielectric constant materials exhibit significantly higher temperature rises. J [M A /cm ] Fig. 7. k= 1.7 Dielectrics Polyimide HSQ 0.5 3 2 3. The materials properties are shown in Table 3. k =0 .05 3.2.1 Thermal impact on Cu/lowk Interconnects Fig.5 3. k= 1. 17) HS Q (!r =3 .5: Temperature of top global interconnects rises sharply for Lowk dielectrics. 2) 25 0 po l yim ide (!r= 2.0 Table 3.2 4. 3.54) FSG (!r=3 .1 FSG 1. 2. k= 0. ρ [µΩcm] is electrical resistivity near 100ºC and εr is the relative dielectric constant 30 0 Metal Temperature Tm [°C] a i r (!r= 1.Chapter 3: Thermal Impact on Interconnect Design 3.6 10  Air 0. 05) airga p (!r= 2. 56 . 3) 20 0 15 0 10 0 0.3.3 0.75. As the figure shows.5 2 2 . 0) a e roge l ( !r=1.1.0 Aerogel 0.5 1 1.
5 1 1 . It should be noted that for the case of airgap scheme (ILD: SiO2 and IMD: Air).5 2 2 .6 as a function of current density. J [ M A /c m 2 ] Fig.Section: Temperature Effect on Cu/lowk Interconnects 140 RmetCtotal [psec/mm] 120 100 80 60 40 20 0 . 3.5 4 F SG H SQ ai r gap pol y i mi de aer ogel ai r C u r re n t D e n s ity . The product of RmetCtotal per unit length for these dielectrics is shown in Fig. The lower the dielectric constant. 3. the lowk advantage of reduced capacitance can be offset by the increased temperature rise due to poor thermal conductivity. This is because airgaps reduce the dominating linetoline capacitance while leaving the SiO2 ILD intact for better thermal conductivity. As the figure shows.5 3 3 . Since electromigration mean time to failure (MTF) is exponentially de 57 . the stronger the Joule heating and greater RC degradation. the RmetCtotal is relatively constant through the range of current density and it is even better than polyimide. Tm. interconnect temperature.6: RC delay is strong function of current density on the wires because of Joule heating.
3. The constraint is more stringent for lowk dielectrics.7: Maximum allowed current density is limited by the maximum allowed ∆T on the metal wires. in Fig. the maximum allowable current density for a given temperature rise is evaluated in Fig. 3. ILD thickness should be evaluated to assess the tradeoffs of thermal impedance and capacitance.8 show the optimization of the wire aspect ratio (AR) and the ILD 58 . 3.3. In addition. it is important to know the thermal impact of using low aspect ratio lines on the expected performance and reliability.Chapter 3: Thermal Impact on Interconnect Design pendent on wire temperature. Current Density [MA/cm2] 6 5 4 3 2 1 0 1 d elt a T=50 C d elt a T=3 0 C d elt a T=2 0 C d elt a T=10 C d elt a T=5 C Dielectric Constant ε r 2 3 4 Fig. The contours of constant RmetCtotal.5 Delay and Reliability Optimization With the possibility that much fatter wires may be used in the global tier to reduce IR drop.7 for different lowk materials.
3 0.9 shows constant MTF contours. with the temperature effect included.32µm/0.5 2 1.5 0 0.1 0. 3.5 1 0. Global Wire Aspect Ratio (AR) 3 2.15 µm). 3. 3. The value of H/S (=0. Fig.15µm) is fixed.9 that although fat wires can provide better speed performance. (H/S=0. where H is metal height and S is wire spacing.2 0. but they are also more subject to electromigration failure since the temperature is higher. Solid curves represent the case thermal effect neglected.8 and 3. The MTF can never reach 50% of MTF at 105°C which is the wire temperature specified in ITRS.5 Delay and Reliability Optimization thickness for delay consideration. 59 .4 0.Section: 3.5 60ps 40ps 20ps 100ps 80ps ILD thickness [µm] Fig.32/0. It demonstrates that Joule heating can increase delay of the interconnect by as much as 15%. We can conclude from Fig. Dashline curves include the influence of thermal effect.8: Constant RmetCtotal contour plots of global level wiring as functions of wire aspect ratio (AR) and ILD thickness.
Column 2 shows the maximum current density specified in ITRS and column 3 shows the resultant temperature of the top global wires. The MTF(Tm) is compared to the MTF(105°C). 3.9: Constant normalized electromigration MTF.35 0.2 0. but drops sharply for the following technology nodes. 3.2 0. (H/S=0.5 2 1.Chapter 3: Thermal Impact on Interconnect Design Global Wire Aspect Ratio (AR) 3 2.25 0.5 ILD thickness [µm] Fig.15 µm).4 0. The RmetCtotal delay is shown in column 4 and the corresponding MTF is shown in Fig.3 0.45 MTF/MTF(Tm=105°C) 0.3 shows the coupled analysis of delay and electromigration MTF for various technology nodes.5 1 0.1 0. The achievable MTF is about 90% at 130nm node.5 0 0. 3.32/0. MTF(Tm)/MTF(105°C). contour plots of global level wiring as functions of wire aspect ratio (AR) and ILD thickness. the current density has to be reduced as shown in column 60 .6 Impact of Joule Heating on Scaling Trend Table 3.4 0.10 as the solid curve.3 0. To confine the wire temperature at 105°C.
drops drastically as can be observed in column 6. as modeled by the wellknown Black’s equation. Table 3. The optimal current density is much closer to the value specified in ITRS. We notice that in both cases trying to match certain specific wire temperatures (105°C and 110°C). compared to the maximum current density specified in ITRS. which are related by Joule heating.3. 61 .Section: 3. as shown in the flat line in Fig. including substrate and Joule heating from all metal levels underneath.6 Impact of Joule Heating on Scaling Trend 5 and resultant current drive. It implies that this approach is overly conservative. To relax the temperature rise to 10°C above substrate. Tm is the top global wire temperature with all the heat sources. as shown in Fig. we can achieve the expected MTF by optimizing both the current density and wire temperature. the current drive can be improved as shown in column 7.3: Coupled evaluation of electromigration reliability and performance for global interconnects for 22130 nm technology nodes.10.3. The result is shown in the last three columns of the table. taken into account. although the wire temperature is higher than 105°C. Since the MTF of electromigration is also inversely proportional to the square of current density.10. On the other hand. the MTF is on target. as shown in the last column. the corresponding MTF rises above the required MTF.
M T F=1 0 0 % 45 65 90 130 Technology Node Fig. Global wires will be more problematic with higher operating temperature and careful consideration is imperative. Optimization with various interconnect parameters is provided. Thermal effects can severely degrade both reliability and speed performance.7 Summary In conclusion. T m=1 1 0 C 60 40 20 0 22 4 3 2 1 0 J ma x . 3. T m=1 0 5 C J ma x .Chapter 3: Thermal Impact on Interconnect Design 100 MTF achieved (%) 6 5 J ma x . 3. a detailed analysis of the impact of Joule heating on the characteristics of future Cu/lowk interconnects is presented using a realistic full chip model with via effect included. Joule heating will limit scaling of current density and use of lowk materials. I T R S J ma x . 62 MTF achieved (x) 80 .10: The solid curve shows the MTF (%) that can be achieved under the current density specified in ITRS. The brokenline curves show the MTF under various wire temperature (Tm) criterions and the values refer to the right axis.
However. the electrothermal simulation approach is desirable to facilitate quick estimation of temperature rises and to investigate thermal coupling effects between wires in order to provide thermal design guidelines for advanced interconnect structures [56. without solving complex differential equations. Nevertheless. we analyzed the thermal effects in interconnects using analytical thermal models and demonstrated the strong influence of vias on the temperature distribution in metal lines. 57]. This chapter presents a fast SPICE based 3D electrothermal 63 .Chapter 4 SPICEBased ElectroThermal Simulation Methodology 4. without including thermal capacitance.1 Motivation In Chapter 2. Teng [58] has used lumped thermal circuit model to predict interconnect temperature. the model is not capable of transient analysis.
The remainder of this chapter is organized as follows. respectively. Furthermore. it is very desirable to have an efficient 3D simulation methodology to 64 .2. as efficient heat sinking paths to alleviate the temperature rise in the metal wires. To serve as a starting point. Finally. characteristics is discussed.2 SPICEBased Thermal Modeling In order to provide robust thermal analysis for deep submicron Cu/lowk interconnects. The effect of vias. is included in our analysis to provide more accurate and realistic thermal diagnosis.3 & 4. All the dimensions of the interconnect structure in this paper were taken from the 100 nm technology node based on the ITRS [59].Chapter 4: SPICEBased ElectroThermal Simulation Methodology simulation methodology to characterize thermal effects due to Joule heating in high performance Cu/lowk interconnects under steadystate and transient stress conditions. The simulation methodology has also been applied to quantify the use of dummy thermal vias as additional heat transfer paths to lower the temperature rise in the metal wires for the first time. thermal coupling between wires is evaluated and found to be Finally. in Section 4.4 describe the impact of via effect of interconnect thermal characteristics under steadystate and transient conditions. we summarize and conclude in Section 4. The results demonstrate excellent agreement with experimental data and those using Finite Element (FE) thermal simulations (ANSYS). the impact of metal wire aspect ratio on interconnect thermal significant.5. Sections 4. 4. we give an overview of our thermal simulation methodology.
This thermal network can be easily implemented and simulated using SPICE in the same manner as an electrical circuit network by simply employing the proper counterparts as illustrated in Fig. as shown in Fig. Rm = ∆l k M wH (4. Ri. Based on the thermalelectrical analogy (Fig. 4.1.1). Rm. per segment can be computed as. the thermal resistance along the metal wire per unit length. and the thermal resistance of the insulator. 65 . 4.Section: 4.2. 4.1) Thermal Temperature Heat T [K] Q [J] Electrical Voltage V [V] Charge Q [C] Current i [A] Electrical resistance R [V/A] Heat transfer rate q [W] Thermal resistance RT [K/W] Thermal capacitance CT [J/K] Electrical capacitance C [C/V] Governing equations SteadyState condition Temperature Rise ∆T = q RT Heat diffusion ∂T 2 ∇ T = RTCT ∂t Voltage Difference ∆V = i R Transient condition RC transmission line ∂V 2 ∇ V = RC ∂t Fig.1: ThermalElectrical analogous parameters. For example. a 3D RC distributed thermal circuit model has been developed.2 SPICEBased Thermal Modeling estimate the temperature profiles in the metal wires and evaluate the thermal coupling between them. 4.
and ∆w is the width of the chosen insulator segment. Furthermore.3) (4. tI is the thickness of insulator. Cm. kM and kI are the thermal conductivity of metal wire and insulator.Chapter 4: SPICEBased ElectroThermal Simulation Methodology Mn+1 IMD ILD Heated Metal Wire Via Neighbor Metal Wires Mn Mn1 Fig. M ρ M ( wH∆l) (4. H is metal height. respectively.2) where ∆ℓ is the unit length of metal wire. C m = c p . w is the width of metal wire.2: 3D thermal circuit RC transmission line model for transient thermal analysis of interconnect structures. 4.4) and C I = c p . and Ri = tI k I ∆l∆w (4. can be shown as. Ci. I ρ I (t I ∆l∆w) 66 . and the thermal capacitance of the insulator per segment. the thermal capacitance along the metal wire per unit length.
q = ( j rms wH ) 2 ( ρ (T ) ∆l ) wH (4.5) where jrms is the rootmeansquare current density flowing through the wire and ρ(T)(=ρ0(1+βT)) is the temperature dependent metal resistivity.I. 4. This technique was validated by comparing with both experimental data [60] and simulation results carried out by ANSYS. RC transmission lines are used to model the heat diffusion due to the similarity of the governing equations as illustrated in Fig. ρ0 is the metal resistivity at 0°C and β is the temperature coefficient of the resistivity (TCR). Inclusion of the lateral thermal impedance in the model captures the heat spreading effect and thermal coupling from nearby interconnects. there is no need to add any datafitting modification in the circuit model as in [58]. To account for the temperature dependence of the metal resistivity. The main advantage of this SPICEbased methodology is that once layout data is available. In addition. a finite element (FE) simulation package. HSPICE was used for the 3D simulations of the distributed thermal RC circuits in this work. also be conveniently analyzed.Section: 4. ρM and cp. It will be shown in the following sections that it exhibits excellent accuracy within 5% agreement. respectively. 67 . thus. Therefore.M.1 and transient thermal effects in interconnects can. ρI are the specific heat and density of metal wire and insulator. the heat generation (q) in each segment of the wire has been modeled as a voltage (temperature) controlled current source. the thermal analysis can be quickly done.2 SPICEBased Thermal Modeling where cp.
1 Impact of Via Separation on Effective kILD In the steadystate analysis.4. current is not flowing in the nearest.3.7x106 A/cm2 and via separation of 100 µm. third M n Se V p a ia ra tio n S M n 1 W 1 V ia (a) W 2 (b) Fig. The resultant simplified model is shown in Fig. Current is flowing in alternate wires. Simulated temperature profile along Cu interconnect with current density J=3. The temperature profiles along the embedded Cu wires with 100µm via separation for two different dielectrics are shown in Fig.3: (a) Interconnect configuration. i. 4.3 SteadyState Analysis 4.3. (b) Correspondent distributed thermal circuitry. 68 .e. 4. 4.Chapter 4: SPICEBased ElectroThermal Simulation Methodology 4. thermal capacitance can be removed from our RC thermal circuit model for the obvious reason.
HSPICE 40 60 80 100 Location along metal wire [µm] Fig. wire width (w)=wire spacing (S)=0.3 4. The interconnect geometries used in the simulations are stated in the caption and the materials properties are listed in Table 4.5 Dielectrics Polymer 0. nearest. fifth nearest and so on neighboring wires but current is flowing in the second nearest. and ILD thickness tILD=0. Since the Parameters k [W/mK] ρ [kg/m3] Cp [J/kgK] ρ [µΩcm] @ 0°C TCR × 103/K εr Conductor Cu Al 400 220 8933 2720 385 900 1.0 Table 4. ANSYS SiO2.3 SteadyState Analysis Temperature Rise.87 1014 1. 4.1: Materials properties used in this work 69 . forth nearest and so on neighboring wires.8 µm.8 µm.Section: 4.66 4.5 Oxide 1.67 2. ∆T [°C] 18 16 14 12 10 8 6 4 2 0 0 20 Polymer. The diameter of the vias is assumed to be the same as the width of the wire for the damascene process. HSPICE SiO2.1.03 0.3µm. Geometries used in the simulations are metal height H=0.4: Validation of HSPICE thermal simulation with Finite Element Analysis (ANSYS) data. ANSYS Polymer.3 1380 1195 2.0 Air 0.2 2220 745 4.
This is due to the larger ratio between the thermal conductivity of Cu and the dielectric and hence a stronger influence from vias is expected. Owing to the lack of available experimental data for such small spatial resolution under steadystate current. separated from the underlying layer by ILD of thickness tILD and thermal conductivity kILD and s is the heat spreading factor to accommodate the deviation from 1D heat flow between a metal wire and the underlying layer. 4.Chapter 4: SPICEBased ElectroThermal Simulation Methodology substrate. to which a heat sink is usually attached.2 W/mK) is more curved. It can be observed in Fig. The distance over which via effect is important can be roughly estimated by thermal healing length LH. It can be thought that within 70 . the chip is enclosed by thermally insulated package materials. in general. which is 14 µm and 30 µm for SiO2 and polymer respectively. This is a reasonable assumption because.3 W/mK compared to kSiO2 =1. 4. as compared to traditional finite element simulation. H is the metal wire thickness. is the metal thermal conductivity. and it demonstrates excellent agreement (Fig. only heat conduction downwards is considered in this interconnect thermal modeling. the results from SPICE simulations are compared against with those using finite element thermal simulation (ANSYS). Heat convection to the ambient air is ignored by the application of adiabatic boundary condition on the four side walls and top of the chip. this approach. is assumed to be the sole heat dissipation path to the outside ambient. However.4 that the temperature profiles along the wire for dielectrics with lower thermal conductivity (kpolymer = 0. because of the wide familiarity of SPICE to the circuit design community.4). can be adopted more easily and it requires much less effort to set up as well as less CPU time to run. (=kMtILDH /kILDs)1/2 [45]. where kM.
the longer the LH.3 SteadyState Analysis 100 kILD.eff is obtained by taking the average temperature of the wire and adjusting from the nominal 71 . the kILD.1 0. Therefore.01 0 100 200 300 ILD: SiO2 ILD: polym er ILD: air kSiO2 kpolymer kair 400 500 Via Separation [µm] Fig. The lower the nominal kILD. the range of LH from vias. Fig.eff is plotted against via separation for three different ILD materials. and hence. 4. 4. large portions of the wires are still under the influence of vias. Knowing the temperature profile (e. the effective thermal conductivity of lowk dielectrics is considerably higher than the nominal values with the help of vias. heat generated will flow through the vias to the underlying layer. Even at the global level. where the kILD. the longer the LH and the stronger the via effect.4). the stronger is the via effect. Since the via separation for the local interconnects is generally much shorter than LH. as shown in Fig. heat will flow through the ILD and the via effect is diminished. Beyond LH. via effect can not be ignored in local interconnects. eff [W/mK] 10 1 0.Section: 4.g.5.5: Effective ILD thermal conductivity increases with decreasing via separation. It can be observed that the lower the nominal kILD. 4.
5. significant difference in temperature profiles along the wires and in the maximum temperature rise can arise between the realistic situation of heat dissipation in the presence of vias and the overly simplified case that ignores via effect.Chapter 4: SPICEBased ElectroThermal Simulation Methodology kILD accordingly. As concluded from Fig. 4.4.1 µm to 2 µm. The interconnect structure used here is the same as that used for Fig. which is the maximum current density specified for the 100 nm technology node from ITRS [58]. the configuration shown in Fig. 4.4x106 A/cm2).3. 4. The tempera 72 . our SPICEbased electrothermal methodology provides a very convenient solution. 4.4 and 4.3 µm. It is very difficult to solve the coupling effect analytically.6(a). It can be observed that incorporation of via effect results in increased kILD. In Fig. To investigate the parallel thermal coupling effect between wires. This fact can explain why the interconnect temperature is not as high as commonly assumed when lowk ILD is implemented in advanced interconnect structures and. but with wire width fixed at 0. In addition. Furthermore.2 Thermal Coupling Effects Thermal coupling between wires in the same layer and different layers has become an important issue due to the scaling of wire spacing. polymer is used as the insulation material. via effect has to be included for accurate interconnect temperature estimation. However. and wire spacing varied from 0.3(a) is examined and symmetrical boundary conditions applied.eff especially for ILD materials with lower nominal thermal conductivity. it should be noted that the steeper temperature profile for the lowk dielectrics can exacerbate electromigration due to the larger temperature gradient. the middle wire is carrying a current density of J0(=1. 4. thus.
1u m S=0.0u m (TT0)/T0 (%) 1 50 1 00 50 0 50 0 0 .3µm. used in this simulation and current flow patterns showing current flowing on alternate wires.8um S=1 .0um S=2 . J [x J 0 ] (b ) n o c u rre n t c u rre n t flo w in g flo w in g w ire S1 (i) S2 (ii) (c ) Fig. The current density in the middle wire is J0.4 0. Line width is 0.3 SteadyState Analysis 2 00 S=0. The temperature of wires carrying lower current density is strongly affected by the heat coupled from the neighboring wires.8 1 C u rr e n t D e n s it y in N e ig h b o r in g W ire .Section: 4. J [x J 0 ] (a ) 1400 1200 1000 800 600 400 200 0 200 0 (TT0)/T0 (%) S=0 . Line width is kept at 0.3µm in this simulation. 4.8 1 C u r re n t D e n s ity in N e ig h b o r in g W ire .0u m S=2.4x106 A/cm2. 1.8u m S=1. with different spacing. (b) The current density in the middle wire is one third of J0.4 0.6: (a) T is defined as the maximum temperature in the middle wire.6 0.6um S=0 .2 0.6u m S=0.0um 0. 73 .2 0 .3um S=0 . (c) Top view of interconnect structures. Reference temperature T0 is the maximum temperature in the middle wire with no current flowing in the nearest neighboring wires and with equal line width and spacing (S).3u m S=0.6 0.1um S=0 .
4.6(c)(i) for a given area. The strong thermal coupling may cause resistance variations and further enhance crosstalk problem.3µm).4.Chapter 4: SPICEBased ElectroThermal Simulation Methodology ture variation due to thermal coupling is normalized to T0. In the case of equal line width and spacing (0. 4. Transient Stress Analysis Apart from normal circuit operating conditions. S (=0. which causes accelerates thermal failures [61].3µm in this simulation). only one third of J0 is flowing in the middle wire and the thermal coupling effect is shown in Fig. The coupling effect can be even more drastic when the wire is carrying a low current density. which represents a typical signal line. Semiconductor industry survey has indicated that ESD 74 . ICs also experience highcurrent stress events. With S1 < S2. In this case. 4. 4. which is with S equal to w. the temperature of the middle wire can be increased by nearly 100%. more current is flowing in Fig. which is the maximum temperature in the middle wire when there is no current flowing in the nearest neighboring wires and with equal wire width and wire spacing. thus causing higher thermal coupling effect to the middle wire. The negative values are due to the symmetric boundary condition chosen in this simulation and the temperature variation is normalized to T0. the most crucial of them being electrostatic discharge (ESD).6(c) depicts the top view of two current patterns and can be used to understand the results. Fig. It can be seen that both the current density of the neighboring wire and the line spacing greatly affect the temperature of the middle wire.6(b).
The governing heat diffusion equation of temperature rise during brief transients can be written down based on energy conservation law as [60. Sv and Sl are shape factors in vertical and lateral directions to account for threedimensional heat spreading. ILD and IMD materials respectively.4.Section: Transient Stress Analysis is one of the major causes of the failure of ICs [62]. Again. Due to the nature of heat 75 . cIMD and ρM. Therefore. 4. 64]. only heat conduction is considered in this analysis.7) and cM. Ceff can be understood as the effective thermal capacitance per unit length to absorb the heat. ρIMD are the specific heat and density of metal. the maximum temperature rise on the wire. cILD. a highcurrent shortpulse (J > 10MA/cm2. Furthermore. ∂ ∂T ∂ ∂T ∂ ∂T i 2 ρ (T ) ∂T (k M ) + ( k ILD ) + ( k IMD ) + rms = Ceff ∂x ∂x ∂y ∂y ∂z ∂z wH ∂t (4. ρILD. it is important to be able to analyze interconnects heating under transient stress conditions. and pulse width tpulse < 200 ns) usually causes much higher ∆Tmax.6) where C eff ≅ [c M ρ M wH + 2c ILD ρ ILD sV w α ILD t + 2c IMD ρ IMD sl H α IMD t ]−1 (4. I/O circuitry also exposes to high current events since it interfaces between multiple power supplies [63]. αILD and αIMD are the thermal diffusion coefficients of ILD and IMD materials respectively. SPICEBased Simulation Methodology In general. due to more severe selfheating than under normal operating conditions and the heat diffusion is limited to the immediate materials in contact with the metal line.1 Analytical Model vs.
It should be noted that. After substituting these thermal quantities into the correspondent electrical identities in the thermal circuit model. In contrast with the unsolvable nature of the analytical analysis. the temperature dependence of ρ(T) absolutely can not be ignored because of the large temperature rise.1Eq. 4. 4. Ceff is increasing with time as heat diffuses through larger volume. calculated by Eq. RC transmission lines are used to model the heat diffusion (Fig. to form thermal RC network. ∆T = E C eff (4. unlike in the case of normal steadystate operation condition.9) where E is the pulse energy. can be easily 76 . 4. Together with the timedependent Ceff. 4.5. temperature rise on the wire under current pulse irms with duration tpulse can be shown as. 4. manifested as node voltage.Chapter 4: SPICEBased ElectroThermal Simulation Methodology diffusion process.6.2. 4. Heat generation.8) ≅ 2 irms ρ (T )t pulse wH [cM ρ M wH + 2c ILD ρ ILD sV w α ILD t + 2c IMD ρ IMD sl H α IMD t ]−1 (4.2) due to the similarity of the respective governing equations as illustrated in Fig. is realized as voltage (temperature) controlled current sources as shown in Fig. using Eq 4. there is no easy analytical solution available for Eq.4.1. an efficient simulation methodology is desirable and will be discussed in details in the following. Therefore. SPICEbased simulation methodology provides a convenient and efficient transient thermal analysis. Therefore. The distributed thermal resistance and thermal capacitance components are constructed. the temperature along the interconnect.
3 um 0.15um TiN Al 0.6um. 4.7(b) and for simulations in Fig. t. 4. data t=1. ∆Tmax [°C] 800 600 400 200 0 0 t=0. simulation t=1. exp.7(a) is used. (b) Validation of HSPICE thermal simulation with experimental data from [60]. are under examination.05um 0.01 um SiO 2 SiO2 (a) 1000 Temperature Rise. 4. Two different wire thickness .8. exp. Al interconnect test structure shown in Fig. where it shows excellent agreement.7.Section: Transient Stress Analysis obtained through SPICE simulation.6um. simulation 2 4 6 8 Current Density. 77 . 4.data t=0.2um.7: (a) Schematic cross section of Al and Cu interconnects with cladding layers used in model validation Fig.6um 3 um TaN Cu 0. This technique was validated by comparing with experimental data from [60] as shown in Fig. 4. J [x10 7 A/cm 2] (b) Fig.6um 0.2um.
2um Current Density. Cu Interconnects With the interconnect structure of Al and Cu wires described in Fig.1. ∆Tmax [°C] 1200 1000 800 600 400 200 0 0 2 4 6 8 10 Al. 4. 4.8. J [x107 A/cm2] Fig. t=0. It should be noted that the simulation is done for a 3D parallel metal array.7(a). t=0. The materials properties again are listed in Table 4. due to their higher Temperature Rise. 4. As can be observed from Fig. Al vs. Symbols represent simulation data.Chapter 4: SPICEBased ElectroThermal Simulation Methodology 4. 4. Cu wire shows much lower temperature rise compared to Al wire under a 200 ns pulse stress.6um Cu. for the same cross section. 78 .4. temperature rise is simulated and the comparison is shown in Fig. but only the wire under examination has a current pulse flowing through it because event like ESD usually strikes very few wires at a time.8. t=1. and surrounding dielectric. Cu wires would experience lower ∆Tmax than Al wires.2um Cu. 4. current density.2.7(a). t=1. shown in Fig.6um Al.8: For the same wire cross sections.
and ILD thickness tILD=0. subjected to transient current pulses of 200 ns and 2 µs duration.3. it is prudent to study the effect of vias on the thermal characteristics of Cu wires to make reliable use of them in deep submicron designs. the interconnect dimensions used throughout the following sections are taken from global interconnect levels from the 100 nm technology node based on the ITRS [58] with metal height H=0.26 µm. However. 4.Section: Transient Stress Analysis thermal conductivity and thermal capacity. wire width w=wire spacing S=0. Thermal diffusion length LD can be interpreted as the distance over which via effect is prominent. The heat diffusion length. 4. wire width w=wire spacing S=0. which is a function of time and should not be confused with the thermal healing length LH under steadystate condition as mentioned in Chapter 2 and Section 4.8 µm.9 compares the normalized spatial temperature distribution along an interconnect line with polymer as insulation material and a via separation of 100 µm.13µm. and ILD thickness tILD=0. we would like to remind the readers again that. due to their lower resistivity. then the dimensions are metal height H=0. which results in longer diffusion lengths. is 5 µm and 16 µm for the 200 ns and 2 µs pulse durations respectively.3 Impact of Via Separation and Lowk Dielectrics Before continuing our discussion.8 µm. Fig. and most importantly. ∆Tmax can still be high when metal dimensions are scaled down and lowk dielectrics are incorporated. Therefore.4.3µm. where α is the thermal diffusivity of the interconnect materials. If a local interconnect is simulated. This along with higher melting point (~1100 0C) than that of Al (~660 0C) would provide more thermal margin to Cu interconnects. 79 . LD ∝ (αt)1/2. if it is not otherwise mentioned. It can be observed that the temperature rise profile for the 2 µs pulse is more gradual due to the increased influence of vias for longer diffusion time.32 µm.
1000 Temperature Rise. The decay is facilitated by the presence of vias.8 0.6 200ns pulse 0.9: Simulated normalized temperature (∆T/∆Tmax) profile along global Cu/lowk (polymer) interconnect for two pulse durations with current density J=4x107A/cm2.: 10um sep.: 35um sep.:100um 10 1 0. 4.4 0.1 0 2 4 6 8 10 Time after ESD pulse passed [us] Fig.2 1 0. 4.: 20um sep.2 0 0 10 20 30 40 50 60 70 80 90 100 2us pulse Location along Interconnect [um] Fig. ∆Tmax [°C] 100 via via via via via sep. for Cu/lowk (polymer) global wires.: 50um sep.Chapter 4: SPICEBased ElectroThermal Simulation Methodology Normalized Temperature Rise 1.10: The temperature decay after a 200 ns current pulse with J=8x107 A/cm2. 80 .
which results in greater Joule heating. This is dictated by thermal diffusion length (LD). It can be observed that while the temperature rises sharply and lowk dielectrics show worse situation for a typical global line with via separation of 100 µm. the temperature rise is significantly alleviated for local interconnects due to a smaller via separation of 1 µm. For larger via separation ∆Tmax saturates since the effect of vias diminishes. This suggests that they are all within LD and via effect dominates the thermal characteristics. both Fig. 4. 4. 81 . 4. 4.11(a) and Fig. which shows that the spatial temperature distributions for various dielectrics are similar and ∆Tmax is nearly independent of the dielectric material. 13 with 1 µm via separation. at the end of a 200 ns pulse. 12(b)) for different dielectrics as a function of the current density. Even if air is used as both the ILD and the IMD dielectric (worst case thermal scenario). The maximum temperature rise. This shortens the high temperature span that the interconnect would experience and thus reduces thermal problems. However. is shown for global (Fig.Section: Transient Stress Analysis LD is a function of time under transient condition and LH is only a function of materials properties and is not varying with time. 12(a)) and local interconnects (Fig.11(b) indicate that ∆Tmax would be reduced dramatically for smaller via separation.10 plots the temperature decay after a 200 ns pulse. Fig. 4. ∆Tmax is higher for the 200 ns pulse due to higher pulse energy. 4. This is due to the fact that heat does not have sufficient time to diffuse through the surrounding dielectrics. It can also be observed in Fig. via separation for different dielectrics is shown for high current pulse duration of 200 ns in Fig. ∆Tmax. which can be interpreted as the distance over which heat generated in the wire flows through the via.11(b) that for shorter pulse duration the differences between ∆Tmax for various dielectrics is much smaller. ∆Tmax vs. no significantly higher ∆Tmax is observed for the local interconnects.11(b). It can be observed that the temperature decays more rapidly with vias placed closer.11(a) and 100 ns in Fig. This observation is further validated in Fig.
IMD: A ir 0 0 25 50 75 100 125 150 175 200 V ia Separation [um ] (a) 300 Temperature Rise. ∆Tmax [°C] 600 400 ILD & IMD: SiO2 200 ILD & IMD: Poly mer ILD & IMD: A ir ILD: SiO2. ∆Tmax [°C] 250 200 150 100 50 0 0 50 100 150 200 ILD & IMD: SiO2 ILD & IMD: Polymer ILD & IMD: Air ILD: SiO2.11: (a) ∆Tmax of global interconnects for a 200 ns current pulse with J=6x107 A/cm2. 82 .11(a). with the same current density as in Fig. 4. 4. ∆Tmax increases with via separation and saturates when via effect has diminished.Chapter 4: SPICEBased ElectroThermal Simulation Methodology 800 Temperature Rise. IMD: Air V ia Separation [um ] (b) Fig. (b) For a short pulse of 100 ns. ∆Tmax of the global Cu interconnect is much less dependent on the dielectric materials.
J [ A/cm2] (b) Fig.E +00 ILD & IM SiO2 D: ILD & IM P er D: olym ILD & IM Air D: ILD: SiO2.Section: Transient Stress Analysis Temperature Rise. Temperature rises sharply with current density. (b) ∆Tmax of Cu local interconnect with 1 µm via separation under a 200 ns current pulse. 83 .E +08 2.E +07 8. IM D:Air Current Density.E +08 2.E +07 1. ∆Tmax [°C] 1400 1200 1000 800 600 400 200 0 0 2 4 6 8 ILD & IM SiO2 D: ILD & IM P D: olym er ILD & IM Air D: ILD: SiO2. 4. ∆Tmax [°C] 100 80 60 40 20 0 0.E +08 Current Density.12: (a) ∆Tmax of Cu global interconnect with 100 µm via separation under a 200 ns current pulse. IM D:Air 4. Temperature rises are much lower and nearly independent of the surrounding dielectric materials. J [x107 A/cm2] (a) Temperature Rise.
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
Temperature Rise, ∆Tmax [°C]
18 16 14 12 10 8 6 4 2 0 0 0.2 0.4 0.6 0.8 1 ILD & IM SiO D: 2 ILD & IM Polymer D: ILD & IM A D: ir
Location along Interconnect [um]
Fig. 4.13: Temperature profiles along local Cu interconnect under a 200 ns pulse are gradual for all the dielectric materials and ∆Tmax are about the same with J=1x108 A/cm2.
4.4.4 Impact of Dummy Thermal Vias
Dummy thermal vias, which conduct heat but are electrically isolated, can be installed in Cu/lowk structure to lower the temperature rise. Since a highcurrent shortpulse (J > 10MA/cm2, and tpulse < 200 ns) usually causes much higher ∆Tmax due to more severe selfheating than under normal operating conditions and the heat diffusion is limited to the immediate materials in contact with the metal line, the transient condition will define the most stringent via separation requirement for lowering the temperature of the metal wires. It should be noted that the simulation methodology developed here is quite general, and can be easily extended to study steadystate stress conditions by using longer pulses. The advantage of the dummy via effect can be demonstrated by using the
84
Section: Transient Stress Analysis
800
Temperature Rise, ∆Tmax [°C]
600 400 200 0 0 20 40
(a)
ILD & IMD: Air ILD & IMD: Polymer ILD & IMD: SiO2
60
80
100
Via Separation [um]
Temperature Rise, ∆Tmax [°C]
600 500 400 300 200 100 0 0 20 40 60 80 100 ILD & IMD: SiO2 ILD & IMD: Air
Via Separation [um]
(b)
Fig. 4.14: (a) ∆Tmax of Cu/air and Cu/polymer can match ∆Tmax of Cu/oxide global interconnect if dummy thermal vias are added every 20 µm and 30 µm, respectively, with tpulse = 200 ns and J=6x107 A/cm2. (b) Cu/air with thermal vias every 20 um shows nearly the same ∆Tmax as that of Cu/oxide global interconnect with 100 µm via separation under a 200 ns and J=6x107 A/cm2 current pulse.
85
Chapter 4: SPICEBased ElectroThermal Simulation Methodology
lowestk dielectric, air, in ULSI interconnects despite its poor thermal properties. In the case of global interconnect, Fig. 4.14(a) shows that thermal vias would be required approximately every 20 µm in Cu/air and every 30 µm in Cu/polymer interconnect structures to match the temperature rise of Cu/SiO2. For the purpose of comparison, under normal steady state operating condition with the Jmax specified in the ITRS [17], the thermal via separation can be much more relaxed based on a simple analytical evaluation. The temperature profiles for Cu/SiO2 with via separation of 100 µm, and for Cu/air with via separation of 20 µm, experiencing the same ∆Tmax under a 200 ns pulse are shown in Fig. 4.14(b). It can be observed that the temperature rise along the Cu/air wire varies spatially and that the average temperature is much lower than that of Cu/SiO2 wire, resulting in reduced thermal problems and may relax the requirement for thermal via separation.
4.4.5. Impact of Interconnect Aspect Ratio
Finally, the effect of wire aspect ratio on the thermal characteristics is evaluated. For the same metal thickness, wires with smaller aspect ratio would suffer higher ∆Tmax because of the smaller surface areatovolume ratio (Fig. 4.15(a)). For the same cross section area, indicating same current capability, a larger perimeter would result in lower
∆Tmax by offering larger area for heat to diffuse out of metal wires, as shown in Fig.
4.15(b). However, for embedded air gap (ILD: SiO2 and IMD: Air), ∆Tmax increases slightly with aspect ratio. This can be explained by the fact that with higher aspect ratio, embedded air gap interconnect structure would have increasing area contacted by air. Fig. 15(a) and (b) can be used to provide thermal design guidelines for interconnect.
86
5 2 ILD&IM : SiO D 2 ILD&IM : P er D olym ILD&IM : Air D ILD SiO IM : Air : 2.5 3 3. tpulse = 200 ns and J=6x107 A/cm2. D 2. (b) For the same metal wire cross section area (5 µm2).5 3 3.5 Metal Wire Aspect Ration (AR) (b) Fig.5 ILD& IM : SiO D 2 ILD& IM : P er D olym ILD& IM : Air D ILD SiO IM : Air : 2.15: (a) For the same metal thickness (2.5 1 1. ∆Tx [°C] 750 700 650 600 550 500 450 400 350 300 0 0. 87 .5 2 2. global interconnects with lower aspect ratio (AR). ∆Tmax peaks at aspect ratio=1 due to the smallest perimeters. the impact of aspect ratio (AR) and dielectric structure strategy is shown.5 1 1. D Metal Wire Aspect Ration (AR) (a) Temperature Rise. shows higher ∆Tmax due to lower surface areatovolume ratio. tpulse = 200 ns and J=6x107 A/cm2. 4.5 µm).Section: 750 Temperature Rise. ∆Tmax [°C] 700 650 600 550 500 450 400 350 300 0 0.
Finally. the effective thermal conductivity of ILDs can be significantly higher than the nominal value if via separation is comparable to the thermal characteristic length. The impact of wire aspect ratio (AR) on the thermal characteristics has also been shown to be important for the thermal design of deep submicron interconnect structures. it has been demonstrated that thermal coupling can be significant for wires in densely packed structure. thermal effects should be carefully evaluated not only to address reliability concerns.5 Conclusions In this chapter. to accurately estimate the temperature rise and the thermal characteristics of interconnect. In fact.Chapter 4: SPICEBased ElectroThermal Simulation Methodology 4. The SPICE simulation results show excellent agreement with rigorous finite element simulations and experimental data and can be employed for accurate reliability and performance analysis. the impact of dummy thermal vias on the thermal characteristics of interconnects has been presented. It is demonstrated that via density strongly affects the spatial temperature distribution as well as the maximum temperature rise in interconnects and should be considered in interconnect design. Additionally. It has been shown that. as technology keeps scaling. the effect of via as efficient thermal paths must be considered properly. Furthermore. 88 . It is shown that by optimal spacing of dummy thermal vias the Cu/lowk structures can have the same thermal capability as Cu/SiO2. a simple SPICEbased 3D thermal circuit simulation methodology for steadystate and transient stress conditions is presented which allows quick evaluation of various Cu/lowk interconnect structures. but also for accurate interconnect performance analysis.
g. RF ICs and optical interconnects [6872]. highperformance logic.. heading to dramatic increase in chip area and the use of numerous digital blocks with many long global wires. analog. Additionally. 5. e. as schematically illustrated in Fig. The three dimensional (3D) IC architecture has emerged as a unique solution to alleviate the interconnect delay problem [6567].Chapter 5 Thermal Analysis of 3D ICs 5. to incorporate such diverse combinations as embedded DRAM. SystemonaChip (SoC) designs are often driven by the evergrowing demand for increased system functionality. Furthermore. 3D ICs provide a vehicle for heterogeneous integration to realize SoC designs.1 3D Integration: Background and Motivation Interconnect RC delay is increasingly becoming the dominant factor determining the performance of advanced ICs.1. 89 . On the other hand. the noise generated by the interference between different embedded circuit blocks containing digital and analog circuits can be minimized.
Each Si layer in the 3D structure may have its own dedicated or shared interconnect network. Each of these layers are connected together through Vertical InterLayer Interconnects (VILICs) and common global interconnects as shown in Fig. Souri [73] has demonstrated the basic concept for this 3D analysis.3. as depicted in Fig. 5. By migrating to a 3D structure it is assumed that logic blocks can be rearranged in some fashion so as to occupy any number 90 . 5. 5. 5. increase transistor packing density.2) to create the vertical stacking of multilevel active layers of Si ICs is a concept that may significantly improve deep submicron interconnect performance.Chapter 5: Thermal Analysis of 3D ICs Fig.2.1: Schematic of a 3D chip showing integrated heterogeneous technologies. 3D integration (schematically illustrated in Fig. A general representation of a wirepitch limited 2D IC is considered as consisting of a number of logic blocks. and reduce chip area and power dissipation.
T1: first active layer device.2: Schematic representation of 3D integration with multilevel wiring network and VILICs. M3 and M4 are shared by T1. M1 and M2 are for T2. and the I/O de of active layers of Si.Section: 5. Such an arrangement makes the vertical dimension available for logic block interconnectivity. which can be arranged vertically stacked on top of another. This VILIC is characterized by its much shorter length and smaller contribution to chip size as compared to the original global wire. By performing such 91 . a global wire in the 2D IC connecting two logic blocks across the chip and contributing to the chip size can now be replaced with a VILIC connecting the same two blocks. 5. For instance. M’1 and M’2 are for T1. T2: second active layer device. T2. Optical I/O device: third active layer I/O device.1 3D Integration: Background and Motivation Fig.
a significant fraction of lateral wires can thus be replaced with VILICs which ultimately reduces the horizontal wiring requirement and chip size and prevents the interconnect delay problem from dominating IC performance. there is concern of poor heat dissipation and consequent chip temperature rise in 3D ICs due to increased power density and lower thermal conductivity of inter layer dielectrics (ILD) [24].Chapter 5: Thermal Analysis of 3D ICs replacements across the entire interconnect network. However. reducing wiring requirement. 1 3 2 5 3 2 5 4 4 1 2D 3D 2 active layers Fig. chip area and interconnect delays. The majority of the thermal power dissipated in This heat is typically conducted integrated circuits arises due to signal switching.3: Horizontal interconnects are replaced with VILICs. 5. 92 .
Section 5. chip area and thermal impact are evaluated. Assuming low thermal conductivity of the ILDs. The vertical links (VILICs) and vias have much higher thermal conductivity and hence can effectively reduce the thermal resistance caused by the ILD layers. The analysis presented here is the culmination of collaborative work with Shukri Souri at Stanford University. devices in the upper layers will also generate a significant fraction of the heat. and reduction in chip reliability due to increased junction leakage.) which typically have much lower thermal conductivity than does Si [74. all the active layers will be insulated from each other by layers of dielectrics (LTO. With multilayer device designs. Tradeoffs between power.5. 93 . a complete thermal model including power consumption due to both transistors and interconnect joule heating from multiple strata is presented. Furthermore. 75]. a detailed thermal analysis of high performance 3D ICs is presented under various integration schemes [77]. and by accelerating other failure mechanisms [76]. Section 5. the heat dissipation issue may become even more acute for 3D ICs and may cause degradation in device performance.4 discusses the thermal impact on heterogeneous 3D integration and the main results are summarized in Section 5. Ignoring the effect of these structures can result in overly pessimistic estimations predicting unacceptably high 3D chip temperatures. In Section 5.2. electromigration failures.1 3D Integration: Background and Motivation through the silicon substrate to the package and then to the ambient by a heat sink. In this chapter.Section: 5. polyimide etc.3 shows schematically the power analysis of 3D ICs. HSQ. performance.
5. surface scattering and skin effect. interconnect Si _3 interconnect Si _2 interconnect Si Substrate Bonding Interface Package + Heat Sink Fig. due to barriers.4: Schematic of multilevel 3D IC with a heat sink attached to Si substrate. together with increasing Cu resistivity. 5. Exacerbated by the lowk dielectrics with poor thermal conductivity. 94 .4).Chapter 5: Thermal Analysis of 3D ICs 5.2 3D Thermal Modeling The significant increase in device switching speed and current density in the interconnect layers. thermal effects will not only lead to higher interconnect temperature in 2D ICs but also impact the device temperature in various active layers in 3D ICs (Fig. have augmented heat generation in IC chips.
even for the case of two active layers. Required Rpkg [cm2C/W] 3.5 3 2. technology nodes. The required Rpkg is the maximum allowed value which gives the maximum junction temperature specified in the ITRS [58].5 0 35 50 70 100 130 180 35 20 15 10 5 0 Technology Node [nm] Fig. Rpkg. RILD will become the dominant factor to determine temperature rise in 3D ICs.Section: 5. heat sink) thermal resistance (Rpkg) increases rapidly for future technology nodes.2 3D Thermal Modeling As seen in Fig. 5. to achieve the maximum junction temperature specified in ITRS and the ratio of RILD and Rpkg vs.5 2 1.5.5 1 0.5: The required package thermal resistance. With multiple active layers. 95 RILD / Rpkg (%) 30 25 . 5. the ratio of thermal resistance caused by the ILD layers (RILD) to the required package (including glue layers.
Si sub. mn smn η mn (∑ j i =n Nm 2 rms .Chapter 5: Thermal Analysis of 3D ICs The analytical expression derived based on first principles. mn ρH mn + j = m +1 ∑Φ M j ) ] + Rm ( ∑ Φ k )} k =m M Temperature rise caused by ILDs Temp. s : heat spreading factor [45].eff is the effective thermal conductivity of ILD with the 96 . mn k ILD . Φ : total power consumption of mth stratum. η : via correction factor. tILD : thickness of ILD. Via effect is incorporated in the expression by the via correction factor η (0≤η≤1). glue layer. ρ : electrical resistivity of metal wires.eff = kILD/η. where: • • • • • • • • • • • • • Tamb : ambient temperature. Rise caused by PKG. R : thermal resistance of glue layer and Si layer for each of the stratum. heat sink and Si substrate. including power consumed by active layer and interconnect joule heating. with R1 represents the total thermal resistance of package. kILD : thermal conductivity of ILD materials. 0≤ η≤ 1 [45]. mn : the nth interconnect level in the mth stratum. M : number of strata. with kILD. where kILD. Nm : number of metal levels in the mth stratum. to evaluate temperature rise in 3D structure is given below [77]: TSi _ N = Tamb + ∑ {[ ∑ m =1 n =1 N Nm t ILD . jrms : rootmeansquare value of current density flowing in the wires. H : thickness of metal wires.
5. Power consumption due to both active (device) layers and interconnect joule heating are included.6: ∆T3 = Q3R3+(Q3+Q2)R2+(Q3+Q2+Q1)R ∆T2 = (Q3+Q2)R2+(Q3+Q2+Q1)R1 ∆T1 = (Q3+Q2+Q1)R1 ElmoreDelay model – electrical analogy for the thermal mode employed [45]. This expression can be better understood by comparing it with the Elmoredelay model following an electricalthermal analogy (Fig. 5. 5.85x105 A/cm2 and power 97 . ElmoreDelay Analogy Q3 Q2 R3 Q1 R2 R1 Fig. 5. The results obtained from analytical expressions show excellent agreement with ANSYS (Fig. The model has been validated by comparing it with full chip thermal simulations done using ANSYS [24].7.8).6). where Jrms is 4. the validation is done for the case where via effect is ignored (η = 1). Since it is too complicated to construct the 3D structure with thousands of vias in ANSYS.Section: 5. The twolayer 3D structure with wafer bonding technique used for the validation is shown in Fig.2 3D Thermal Modeling help of via effect and kILD is the nominal thermal conductivity with via effect ignored [45].
it is crucial to include via effect for thermal analysis of 3D ICs. Therefore. 200 and 500 µm. 5. For the case where via effect is included. as shown in Fig. the via separations assigned to the metal layers. 67. via effect greatly helps with heat dissipation and the resultant temperatures are much lower as compared to previous works [24.7: Schematic of 3D structure fabricated by wafer bonding using Cu pad thermocompression.615 W/cm2. 15. However.8. respectively. 30. 98 .Chapter 5: Thermal Analysis of 3D ICs density of each active layer is assumed to be 0. 5. 50. 5. from 1st to 8th levels. Metal wires and ILD Si_2 Cu Air Cu Interstratum interconnect Vias Si_1 Fig. 78]. are 1. the analytical model takes much less computation time and provides key design insight. Interconnect parameters are quoted from the ITRS 100 nm technology node. Furthermore. 80.
Comparing power dissipation and temperature rises as a result of migration from 2D to 3D.8: Temperature distribution along the vertical layers from the Si substrate surface (Si_1) to top metal level of the second stratum. For instance. then the operating frequencies are necessarily different. if the chip areas are assumed invariant. vias included Fig.3 Power Analysis of 3D ICs In general. performing any comparison between 2D and 3D configurations of the same IC is problematic. yet the chip area and hence power densities must change. etc. Such an assumption serves well for a performance comparison. 5. vias ignored Analytical.3 Power Analysis of 3D ICs Temperature Rise above Si Substrate (C) 100 80 60 40 20 0 Si_1M11 M12 M13 M14M15 M16M17M18 Si_2M21M22 M23 M24 M25 M26 M27 M28 0 2 4 6 8 10 12 14 16 18 ANSYS. on the other 99 . the operating frequency can be maintained constant. At issue is the lack of common ground on which to perform an objective comparison. vias ignored Analytical. The 3D structure is shown in Fig. 6. 5.Section: 5. Conversely. M11 represents the first metal level in stratum 1.
3D. Their characteristics are summarized below: • • • • 3D.6V from ITRS projections for 2D ICs at the 50 nm node. and compared to 2D. The resulting chip area. where Ac is the chip area. To address this thermal comparative issue. 3D. All the data in this table are calculated based on the 50nm technology node and the thermal resistance of the package is assumed to be 2. Ac. Case 4: 2fc and equal Ac. several 3D integration cases and scenarios are explored to cover an adequate space for the comparison. A summary of this wirepitch limited 2D case and different 3D integration cases used in this comparative study is listed in Table 5. Case 3: Equal fc and Ac. Case 5: Equal Ac with fc determined by maintaining 2D PTotal. is a different matter. is determined by the larger logic area and power dissipation is unchanged relative to the 2D case.1. The remaining four 3D cases are obtained. The different characteristics of these 3D integration cases give rise to different 100 . is a special 3D integration case in that memory and logic from the 2D are each dedicated to separate active layers without any modifications to the wiring.15cm2 °C/W for a constant supply voltage of 0. fc is the operating frequency and Ptotal is the total power dissipation. Case 2: Equal fc and decreased Ac. by modifying the chip wiring and assuming the memory is interspaced with logic. However. Case 1.Chapter 5: Thermal Analysis of 3D ICs hand. The data in the 2D column represents the standard 2D IC. 3D. the analysis in this section is again focused on the ITRS projection at the 50nm technology node. 3D.
such as leakage and shortcircuit currents.78 6.1: Comparison of power dissipation due to logic. Case 3 3D.97 27.93 174 174 136.17 3.8 69. interconnect (local. 3D IC cases are presented for comparison by varying the chip area.96 4.19 14.92 3.1.92 4.17 20.4 17.25 4.65 Active Layers fc (MHz) Feature Size (nm) Chip Area (cm2) Memory Area (cm2) Logic Area (cm2) PLogic (W) PLocal (W) PSemiGlobal (W) PGlobal (W) PClock (W) PRepeaters (W) POther (W) PTotal (W) Power Density Per Active Layer (Wcm2) Table 5. fc is the operating frequency and C is the capacitance.4 17.82 293.44 20. are all combined under POther.96 6. Vdd is the supply voltage obtained from ITRS. Case 4 3D.24 29.6 43.17 20.19 73.17 8.66 10.92 3. Case 5 1 2 2 2 2 2 3000 3000 3000 3000 6000 3559 50 50 50 50 50 50 8.47 15. Other power dissipating components include memory.37 174 21.63 6.63 11.15 151. 2D 3D. interconnect.95 10.Section: 5.92 3.92 3.25 4. clock distribution and repeaters and are calculated using PDynamic=1/2αCVdd2fc where α is the activity factor (assumed to be 0.30 20.17 8.42 34.76 45. and represent the same 2D IC (conserving feature size.68 18.65 20.24 45.1).7 35.89 8.17 40.3 Power Analysis of 3D ICs power dissipation results as summarized in Table 5.29 17.09 9.17 4.1 12.8 22. clock distribution and repeaters for 2D and 3D ICs with 2 active layers for ITRS 50nm technology node.8 34.63 14.51 8.2 6. I/O pads and static components. Case 2 3D. fc.17 20.34 23.28 17.92 3.8 34.18 5.25 5.8 34.6 41. Ac.93 33. number of transistors and functionality) converted to 3D with 2 active layers 101 . and operating frequency.16 30. Case 1 3D.21 56.44 6.99 34.42 12. The dynamic power dissipation components considered are due to logic.8 34.42 12. semiglobal and global).
In 3D Case 2. Clock distribution capacitances are calculated using the BACPAC model proposed in [79] by considering a buffered HTree model. it is considered linearly dependent on the operating frequency for all 3D cases. 3D Case 4 shows a dramatic increase in the power dissipated primarily due to the significant increase in operating frequency. Since this component is assumed dominated by dynamic dissipation. Interconnect capacitances for the local. Power dissipated by repeaters is calculated based on the driver capacitances and the number of repeaters. lower power density is achieved due to larger chip area. reducing number of required repeaters and reducing the clock distribution network. is calculated for each component to determine the associated power dissipated.Chapter 5: Thermal Analysis of 3D ICs The capacitance. overlap capacitance and junction capacitance all of which are calculated from ITRS data [58]. For PLogic. C. POther is determined in the 2D case to be the sum of remaining components to achieve the ITRS projected total power dissipation for this generation. a larger number of repeaters and clockdistribution network all of which increase the power dissipation. the total power dissipation is seen to decrease primarily due to the reduction in the wiring requirement thus reducing the interconnect power dissipation. 3D Case 3 is associated with a larger chip area which requires longer interconnect lines. the device capacitance is calculated by considering gate oxide capacitance. However. 3D Case 5 illustrates the increase in the operating 102 . semiglobal and global tiers are found from the wirelength distribution and the dimensions of the wire pitches for each tier [73].
Section: 5.3 Power Analysis of 3D ICs
Maximum Temperature (C)
160 140 120 100 80 60 40 20 0 2D 3D 3D 3D 3D 3D , , , , , case1 case2 case3 case4 case5
Fig. 5.9:
Comparison of temperature performance among 2D ICs and five different twoactivelayer 3D ICs scenarios
frequency if the chip area and the power dissipation requirements are maintained constant to 2D. Although the total power consumption as shown in Table 5.1 is reduced by going from 2D to 3D ICs due to the reduction in the interconnect and the clock network related capacitance, the heat removal capability could deteriorate as the upper active layers experience a longer heat dissipation path to the heat sink. Fig. 5.9 compares the temperature rise for different 3D integration scenarios. The lowest temperature achieved is that for 3D Case 3. Here, although the wiring has been reduced by migrating to 3D, the operating frequency and the chip area have been maintained constant as compared to 2D. The power density, then, is significantly lower than in the 2D case giving rise to a lower die temperature. It should be noted that, in most of the cases, 3D ICs have similar 103
Chapter 5: Thermal Analysis of 3D ICs
temperature rise but have the advantage of either reduced chip area (Case 2) or increased operating frequency (Case 5). In the case of equal chip area and operating frequency (Case 3), lower temperature than 2D ICs can be achieved. Note that to double the operating frequency (Case 4), temperature will invariably increase. Nevertheless. the temperature is still much lower than that estimated with via effect ignored [24, 67, 78]. Furthermore, it is desirable to put memory in close proximity to the logic circuitry in order to reduce latency in high performance microprocessors. 3D IC technology provides an excellent opportunity to stack memory and logic. The power consumption in onchip memory is generally less than 10% of total power consumption and the area occupied by memory and logic are comparable at 50 nm node. With these assumptions, several schemes are developed, and applied in the following analysis to 3D Case 4. These schemes are illustrated in Fig. 5.10 where four 3D stack schemes are shown, each with a different configuration of memory and logic. Fig. 5.10 shows four 3D stack schemes along with their temperature performance. It can be observed that with logic in the bottom active layer and memory in the upper layer (Scheme 1), the resultant temperature rise is the lowest. On the other hand, stacking logic parts back to back will
experience much higher temperature rise.
104
Section: 5.3 Power Analysis of 3D ICs
Memory Logic
(1)
Maximum Temperature (C)
250 200 150 100 50 0 Scheme 1 Scheme 2 Scheme 3 Scheme 4 M Temp. @ Si_1 ax. M Temp. @ Si_2 ax.
L M
(2)
L L
(3)
M M
M L
(4)
L M
Fig. 5.10: Thermal capability of high performance 3D ICs (Case 4) for four different 3D logicmemory integration schemes.
4
Noise Figure, NF (dB)
3.5 3 2.5 2 1.5 1 0 100 200 300 400 500
Temperature rise above 290 K, ∆T [K]
Fig. 5.11: The noise performance of a typical low noise amplifier (LNA) under the temperature effect is evaluated. NF is assumed to be 2dB @ 290K.
105
In a dense optical interconnect. power and area for clock distribution [80]. Even the dark current (noise) increases rapidly with temperature (Fig. Employing strained multiple quantum well (MQW) active regions with effective facet coatings [81]. e. Evaluated in Fig.. photodetectors are responsible for opticaltoelectrical signal conversion. 5. imposing more stringent design requirements for 3D integration. In an onchip optical interconnect system. Employing optical interconnects in the global signaling or clocking could eliminate many problems associated with large multiGHz chips like reducing timing skew. 106 . which is enough for reliable communications. the detector photocurrent is ~100µA.12). 5. Usually.g. but indeed gainlimited aiming at low power.11 is the noise figure. the three major temperatureinduced changes in performance are drop in quantum efficiency. that deteriorates rapidly with temperature rise. with a microprocessor to facilitate wireless communication. degradation of laser threshold current. Focusing on the transmitting components like lasers and modulator diodes.Chapter 5: Thermal Analysis of 3D ICs 5. and shift in transmitter wavelength. the receivers are not noiselimited. An initial investigation to look at the thermal aspects of the required optical components is shown below. has attracted major interest with the intent to isolate substratecoupled noise between digital and analog components [67]. indexing RF performance. a phototodark current ratio of 100 is still retained.4 Thermal Impact on Heterogeneous 3D Integration Heterogeneous integration of RF circuit in 3D ICs.
ambience temperature is shown here. Even up to 170°C. From the above analysis. a negligible change in absorption coefficient (an index for fiber loss) has been demonstrated from about 20 to 120°C [86]. A typical example of TiSi metalsemiconductormetal (MSM) PD dark current vs.E11 25 125 225 325 Temperature [C] Fig. the absorption loss is still tolerable. optical attenuators. the incorporation 107 .Section: 5. For a waveguide (attenuator) with an InGaAsP MQW active layer and InP claddings. The last major components are the waveguiderelated devices like modulators. 5. and electric field sensors. [83] and a bimetallic heatsink [84] were explored to get around the detrimental wavelength shift. a quantumdot laser with a maximum operating temperature of 160°C in pulse mode [85] has been demonstrated..E07 1.E08 1.E05 1. Recently.E09 1.3 Power Analysis of 3D ICs 1. etc.E10 1.E06 Idark (A) 1.12: Temperature effect on optical receiver performance is evaluated.
5.5 Conclusions In conclusion. 108 . Temperature performance of various 3D integration schemes has been examined thoroughly. It is shown that with careful thermal designs. a compact analytical thermal model to evaluate temperature distribution in 3D ICs including via effect is presented.Chapter 5: Thermal Analysis of 3D ICs of optical interconnect into the 3D architecture may require some more device design efforts for higher temperature operation. better circuit design and advanced package solution will be necessary. It is demonstrated that via effect must be considered to evaluate the thermal capability of 3D ICs. examples of 3D heterogeneous integration of RF circuits and optical interconnects have been explored from the thermal view point. 3D ICs can have the same thermal capability as that of 2D ICs. For the high performance case where higher temperature is not avoidable. Finally.
The results demonstrate excellent agreement with experimental data and with Finite Element (FE) thermal simulations (ANSYS).1 Summary Due to scaling trends of VLSI technology. as efficient heat transfer paths. Joule heating at the interconnect levels is becoming nonnegligible. not only will thermal effects be a major reliability concern. In this thesis. more realistic and accurate interconnect thermal analysis is finally possible. With the effects of vias. As a result. being taken into account.Chapter 6 Conclusion 6. but also the increase of wire electrical resistivity with temperature can degrade the expected speed performance. both a compact analytical formula and a convenient SPICEbased simulation methodology are developed to evaluate the temperature rise on interconnects under steadystate and transient stress conditions. Significant differences in temperature distribution and maximum temperature rise are observed between the realistic situation of heat 109 .
Furthermore. Coupled analysis of delay and electromigration MTF for various technology nodes suggests that Joule heating will limit scaling of current density and use of lowk materials. in turn. potential bottlenecks and opportunities of future heterogeneous three dimensional (3D) ICs with various integration scenarios are identified from the thermal 110 . the thermal problem associated with lowk insulators is not as bad as it might appear. which may offers a solution to alleviate hot interconnect phenomena. The impact of Joule heating on the scaling trends of advanced VLSI interconnects has been evaluated in detail. it shows that the effectiveness of vias in reducing the temperature rise is highly dependent on the via separation and the dielectric materials used. greatly degrade the expected speed improvement from the use of lowk dielectrics. The effective thermal conductivity of ILDs can be significantly higher than the nominal value if via separation is comparable to the thermal characteristic length. the thermal advantage gained by using dummy thermal vias in advanced Cu/lowk interconnects is quantified. thus. It shows that the interconnect Joule heating can strongly affect the maximum operating temperature of the global wires which will. constrain the scaling of current density to mitigate electromigration and. Therefore. It should be noted that global interconnects would suffer much higher temperature rises than local interconnects due to the much longer via separation and further away from heat sink. Additionally.Chapter 6: Conclusion dissipation with vias and the overly simplified case that ignores the via effect. Finally. Optimization with various interconnect parameters is performed to provide thermal design guideline/insight in the early design phase.
These assumptions are fairly accurate with wire bonding package and still quite reasonable with the current flipchip technology. On the other hand. This will lead to a nonclosed form formula. with the reasoning that the chip is encapulated by some insulating materials. and require a few iterations. It is shown that under certain scenarios. However. 111 . carrier scattering and skin effect etc. we assume the wire resistivity is at a constant value. The effect of this additional heat transfer path should be investigated to ensure valid thermal analysis. more heat will be dissipated through these thousands of metal balls. performance.2 Future Work In our analytical thermal model under steadystate normal operation condition. However. 6. chip real estate and thermal impact for 3D ICs is evaluated.Section: Future Work point of view. we consider the heat sink as the sole thermal dissipation path to the outside ambient and assume adiabatic boundary conditions on the top and on all the four side walls of the chip. with the help of computer. this process should not be too difficult. 3D ICs can actually lead to better thermal performance than planar (2D) ICs. Although errors introduced because of this simplification are minimal. more accurate results can be obtained by including the temperature dependency. Tradeoffs among power. with the increasing number of bumps for each generation.
Chapter 6: Conclusion
Lastly, the thermal modeling and insight developed in this work should not be limited to interconnect level; it can be easily extended to device level and chip level with certain modification. The thermal effects will continue to be a major issue for years to come in chip design, hopefully, more accurate and convenient methodologies can be built upon this work to facilitate the continued advancement of semiconductor industry
112
Bibliography
[1]
G. E. Moore, “Cramming More Components onto Integrated Circuits,” Electronics, vol.38, no.8, April 19th 1965.
[2]
C. Hu, “MOSFET Scaling in the Next Decade and Beyond,” Semiconductor International, pp. 105114, 1994.
[3]
M. T. Bohr, “Interconnect ScalingThe Real Limiter to High Performance ULSI,” IEDM Tech. Dig., pp. 241244, 1995.
[4]
K. C. Saraswat and F. Mohammadi, “Effect of Interconnection Scaling on Time Delay of VLSI Circuits,” IEEE Trans. On Electron Devices, vol. ED29, pp. 645650,1982.
[5] [6]
The International Technology Roadmap for Semiconductors (ITRS), 2001 M. T. Bohr and Y. A. ElMansy, "Technology for Advanced HighPerformance Microprocessors," IEEE Trans. Electron Devices, vol 45, no. 3, pp. 620625, 1998.
[7]
T. N. Theis, “The Future of Interconnection Technology”, IBM Journal of R&D, Vol. 44, 2000.
[8]
R. Ho, K. Mai, and M. Horowitz, “The Future of Wires,” Proceedings of the IEEE, vol.89, no.4, pp. 490504, April 2001.
[9]
S. Venkatesan, A. V. Gelatos, S. Hisra, B. Smith, R. Islam, J. Cope, B. Wilson, D. Tuttle, R. Cardwell, S. Anderson, M. Angyal, R. Bajaj, C. Capasso, P. Crabtree, S. Das, J. Farkas, S. Filipiak, B. Fiordalice, M. Freeman, P. V. Gilbert, “A High Performance 1.8V, 0.20 µm CMOS Technology with Copper Metallization,” IEDM Tech. Dig., pp. 769772, 1997.
[10] N. Rohrer, C. Akrout, M. Canada, D. Cawthron, B. Davari, R. Floyd, S. Geissler, R. Goldblatt, R. Houle, P. Kartschoke, D. Kramer, P. McCormick, G. Salem, R. Schulz, L. Su, L. Whitney, “A 480MHz RISC microprocessor in a 0.12 µm Leff 113
Error! No text of specified style in document.
CMOS Technology with Copper Interconnects,” Int. SolidState Circuits Conf., Tech. Digest, pp. 240241,1998. [11] E. M. Zielinski, S. W. Russell, R. S. List, A. M. Wilson, C. Jin, K. J. Newton, J. P. Lu, T. Hurd, W. Y. Hsu, V. Cordasco, M. Gopikanth, V. Korthuis, W. Lee, G. Cerny, N. M. Russell, P. B. Smith, S. O'Brien, R. H. Havemann, “Damascene Integration of Copper and UltraLowk Xerogel for High Performance Interconnects,” Tech. Dig. IEDM, pp. 936938,1997. [12] B. Zhao, D. Feiler, V. Ramanathan, Q. Z. Liu, M. Brongo, J. Wu, H. Zhang, J. C. Kuei, D. Young, J. Brown, C. Vo, W. Xia, C. Chu, J. Zhou, C. Nguyen, L. Tsau, D. Dornish, L. Camilletti, P. Ding, G. Lai, B. Chin, M. Johnson, J. Turner, “A Cu/lowk Dual Damascene Interconnect for High Performance and Low Cost Integrated Circuits,” Symp. VLSI Technology, Tech. Digest, pp. 2829,1998. [13] F. Pollack, “New Challenges in Microarchitecture and Compiler Design”, PACT 2000, Oct 2000. [14] P. K. Ko, “HotElectron Effects in MOSFET’s,” Ph.D. dissertation, Univ. of California, Berkeley, 1982. [15] C. Fiegna, H. Iwai, T. Wada, T. Saito, E. Sangiorgi, and B. Ricco, “Scaling the MOS Transistor below 0.1 µm: Methodology, Device Structure, and Technology Requirements,” IEEE Trans. Electron Devices, vol. 41, no. 6, p. 941, June 1994. [16] J. Ida, M. Yoshimaru, T. Usami, A. Ohtomo, K.Shimokawa, A. Kita and M. Ino, “Reduction of Wiring Capacitance with New Low Dielectric SiOF Interlayer Film for High Speed/Low Power Subhalf Micron CMOS,” VLSI Technology Symposium, Dig. Tech. Papers, pp. 5960,1994. [17] B. P. Shieh, B. Shieh, K. C. Saraswat, J.P. McVittie, S. List, S. Nag, M. Islamraja, R. H. Havemann, "AirGap Formation During ILD Deposition to Lower Interconnect Capacitance," IEEE Electron Device Lett., vol. 19, no. 1, January 1998, pp. 16~18.
114
[21] K. IEDM. SangiovanniVincentelli. Dig. Y." IEEE 2000 1st International Symposium on Quality Electronic Design. Seung. M. pp. Banerjee. [18] D. 6771. vol. 2000. ED16. V. pp. S. Ramaswamy.” Tech. and Signal Propagation Delay for 0. H. Kang. Eo. Amerasekera. 115 . “Interconnect Thermal Modeling for Determining Design Limits on Current Density. [27] R. “Full Chip Thermal Analysis of Planar (2D) and Vertically Integrated (3D) High Performance ICs. Black. No. issue 2. 1955.E. 4550. G.1999. Ju and K. 2000.1996. Kwon. pp. Cho. K.1969. 19.35 µm CMOS Technology. “Interconnect Capacitance. 6568. 2. Yoh and F. pp. [24] S. 51. N.1996. A. Im and K. 727730. vol.” ComputerAided Design of Integrated Circuits and Systems. O. Mehrotra. R. H. Najm.1997. K. [19] J.S. pp. 619622. and C. Dixit. “The Effect of Interconnect Scaling and Lowk Dielectric on the Thermal Characteristics of the IC Metal. [20] G. 320324. “Shorttimescale Thermal Mapping of Interconnects. Andrews. Park. IEEE IRPS.” Chemical Engineering Process. Wee.Error! No text of specified style in document. San Jose.R. March 2022.S. 197205 Feb. S. C. [25] Y. "A Statistical Model for Electromigration Failures.” 36th ACM Design Automation Conference. IEEE transactions on. Banerjee. Hu. pp. “EOS/ESD Reliability of Deep Submicron NMOS Protection Devices. pp. J. H. Kim.” Tech. 338347..” Proc. Banerjee. A. [26] S. “Electromigration – A Brief Survey and Some Recent Results. Dig. IEDM. “On Thermal Effects in Deep Submicron VLSI Interconnects.” IEDM Tech. [23] Danqing Chen. pp. [22] K. N. Hu. pp. Erhong Li. Dig. Crosstalk. and C. 885891. A. Goodson.” IEEE Transactions on Electron Devices. IEEE I. Duvvury. and H. 284291. and S.” Proc. 2000.P. CA. “Solving Conductive Heat Transfer Problems with Electricalanalogue Shape Factor. Elyse Rosenbaum and SungMo Kang. 1995. vol. pp.
“ IEEE Trans.” Trans. Incropera and D. 1945. 40. Low Cost Flip Chip Technologies. 310316. and B. Katto. Hunter. 621631. and Y. [36] H. 314. ED44. Bilotti. P. Higuchi. 1993. “Selfconsistent solutions for allowed interconnect current density – Part I: Implications for technology evolution. 1971.” IEEE Trans. P. Hunter. 1997. Schafft. Mecrow. Trattles.3. Inc. 13441347.” IEEE J. No. [37] H. SolidState Circuits. A. “Analysis of electrothermal integrated circuits. G. 1997. 406411. 116 . McGrawHill. Banerjee. Electron Devices. pp. Harada.Error! No text of specified style in document. Electron Devices. [33] J. New York: John Wiley & Sons. 1991.C. “Characterization of selfheating in advanced VLSI interconnect lines based on thermal finite element simulation. pp. vol. pp. “Static temperature distribution in IC chips with isothermal heat sources. [34] K. [35] John H. pp. p298305. 192194. [38] F. ED21. pp. Lau. O’Neill. 814. Hu. ED44. ASME. “ IEEE Trans. Rzepka. NY.” IEEE Trans. 67. 2000. T. A. Electron Devices. Introduction of heat transfer. vol. M.” IEEE Trans. K.” IEEE Trans. 7. R. [30] W. 1987. Gray and D. 217226. on Components. R. vol. [28] P. [32] S. “Thermal analysis of electromigration test structure. “Efficiency of extended surfaces. 1998. Hamilton. R. Electron Devices. 664672. Meusel and C. 21. of the IRPS. No. J. ED34. pp. vol. 304309. [29] A. p. Feb.” Proc. “Selfconsistent solutions for allowed interconnect current density – Part II: Application to design guidelines. “Waferlevel Jramp and Jconstant electromigration testing of conventional & SWEAT patterns assisted by a thermal and electrical simulator. pp. Packaging and Manufacturing Technology – part A . vol. A. vol. 2002. A.. [31] W. “Threedimensional finiteelement investigation of current crowding and peak temperatures in VLSI multilevel interconnections. DeWitt. E. Electron Devices. 1974. pp. pp. Gardner. SC6. vol. Mar. vol.
No. January. 158164. “Implications and solutions for joule heating in high performance interconnects incorporating lowk dielectrics. "Analytical Thermal Model for Multilevel VLSI Interconnects Incorporating Via Effect. Poulton. pp. 2002. pp. W. and K. [39] Virgil A." IEEE Electron Device Letters. Vol. Havemann. pp. pp. of Tech.” ICCAD. Colorado. 2003. B. pp. "Closedform Analytical Thermal Model for Accurate Temperature Estimatin of Multilevel ULSI Interconnects " Digest of 2003 Symposium on VLSI Circuit. Chang. [42] K. [44] T. Yoh and F. Saraswat. 2001. 511. 1998. Y.Error! No text of specified style in document. [41] W. 2001 [43] T. Kyoto. Papers. 8384. C. Shieh and K. Saraswat. Fort Collins.” Digest of 1996 Symposium on VLSI Technology. K. CA. Banerjee and A. Najm. Y. Chiang. 2000. and K. [45] T. Banerjee. Shin. 1997. Hawaii. C. N. 1972. Metrology Press. 3133. 2002. [40] W. Sandborn. pp. March 2022. "Impact of Joule Heating on Scaling of Deep SubMicron Cu/LowK Interconnects. 23. Chiang. Y. Chiang." IEEE 2000 1st International Symposium on Quality Electronic Design. 117 . June. Y. 1. J. Honolulu. "Coupled Analysis of Electromigration Reliability and Performance in ULSI Signal Nets. Japan. CA. [46] T. Chiang. 275278. R. Y. "A Statistical Model for Electromigration Failures. pp. [47] W. Saraswat. 4550. K. and J. Resistance temperature transducers. Cambridge University Press. June.H. “The implications of selfconsistent current density design guidelines: comprehending electromigration and joule heating for interconnect technology evolution. Hunter. [48] G. “Compact modeling and SPICEbased simulation for electrothermal analysis of multilevel ULSI." Digest of 2002 Symposium on VLSI Technology. San Jose. June." IEEE International Conference on ComputerAided Design (ICCAD). and K. 483486.” IEDM Dig. NY. pp. Digital Systems Engineering. San Jose. Banerjee.C. Levine. 1995. 141142. Dally and J. M. Mehrotra. R. Saraswat. pp 165172. November 48.
vol. pp. Hu.. Kapur. and K. Banerjee. 2000. Cheung. [54] P. 8492.. “Impact of vias on the thermal characteristics of deep submicron Cu/lowk interconnects.153 – 157.E. Symp. Electron Devices. 2002.G. 37. [52] Jiang Tao. Uzoh.M. Mikkelsen. vol. [51] C. dissertation. Y. “Electromigration characteristics of copper interconnects. pp. Maveety.Error! No text of specified style in document. D. and C.B. N. March 2001 [56] T. Dig. C. June 1997. Nathan W. J. “Copper interconnect: fabrication and reliability. F. Seventeenth Annual IEEE Symposium. no. 13431351. pp.” Thin solid Films. Saraswat. May 1993. [49] B. Lian Zhang. IEDM. Pearson.” Semiconductor Thermal Measurement and Management. and C.” IEEE Trans. Stanford. 141142. [53] P. pp.” IEEE Trans. April 2002. Luther. JaeMo Koo. pp. 1995. Harper. pp.W. 2001 118 . Saraswat. 1997 International Symposium on. Kenny.” Tech. “Effect of via separation and lowk dielectric materials on the thermal characteristics of Cu interconnects.18 – 22. Q. no. J. and K. Proceedings of Technical Papers.” VLSI Technology. 4.E.C. T. 1997.. Kapur. Y. and Applications. 5.” Ph. K. Kaufman. Goodson. J. Chiang. J. K. Chiang. B. C. Shulin Zeng. 49. vol. pp.D. K. K. Peng Zhou. [57] T. pp. [50] C... 262. “Projecting interconnect electromogration lifetime for arbrtrary current waveforms. Hu. 261264. [55] Linan Jiang.G. K. J. “Scaling induced performance challenges/limitations of onchip metal interconnects and comparisons with optical interconnects.” Tech. May 1990.” IEEE Electron Device Lett. Hu.C. “Technology and reliability constrained future copper interconnectsPart I: Resistance Modeling. Systems. Liew. Chandra and K. “Twophase microchannel heat sinks for an electrokinetic VLSI chip cooling system. Hu and J. G. 14. Saraswat. “Copper interconnection integration and reliability.A. Santiago. 249251.and Tran. C. 590597. Hummel. VLSI Technology. Dig. Electron Devices... Cheung.
690702.” Int. 1997. DeWitt. 714. and C. [68] Akasaki. 18. New York: John Wiley & Sons. Pelgrom and E. 2426. pp. 6165. P. “Interconnect performance modeling for 3D integrated circuits with multiple Si layers. C. C. pp. J. 30. Banerjee. [62] T. Cheung. vol. Interconnect Technology Conf. vol. 37th ACM Design Automation Conf. 602633. [60] K. 1997. Hu. vol. J. P. “Multiple Si layer ICs: motivation. 89.” Proc. A. "ThreeDimensional IC Trends". Saraswat. vol.” IEEE J. pp. M.. [59] The International Technology Roadmap for Semiconductors. Saraswat. 873880. Green. [67] K. No. Proceedings. pp. Dijkmans.” in Proc. Y. 823825. [66] S. May 2001. pp. pp. “iTEM: a temperature dependent electromigration reliability diagnosis tool. ComputerAided Design.Error! No text of specified style in document. A." Proceedings of the IEEE. Souri and K. Introduction to heat transfer.” IEEE Electron Device Lett. EOS/EDS symp. 9. K. Special Issue. Banerjee. Aug. 7. 405407. 8. 1999. E. [65] S. pp. Amerasekera. No. “ESD: A pervasive reliability concern for IC technologies. 2002. SolidStatae Circuits. 2000. Interconnections. J.” Proc. Souri. Mehrotra. Saraswat. [63] M. S. Banerjee. and K. Duvvury and A. 1999. J. performance analysis. 882893. Kang. 1998. C. Dec. 1986 119 . “Highcurrent failure model for VLSI interconnects under shortpulse stress conditions. and design implications. Amerasekera.K.” IEEE Trans. IEEE. “A 3/5 V compatible I/O buffer. Vol. Proceedings of the IEEE. [64] F. Incropera and D. [61] C. 5. No. pp. and K. [58] C. Souri.. 16. C. Cheng. C. 81. N. P. Rosenhaum and S. Teng. "3D ICs: A Novel Chip Design for Improving Deep Submicron Interconnect Performance and SystemsonChip Integration..Addressing The Next Challenge of IC Technology. 1995. pp. Kapur. 10th. May 1998. “A review of EOS/ESD field failures in military equipment. no.
M. Lentine. pp. B. Ramm. Goossen. [77] TingYen Chiang. V. Shimada. S. Horiba. A.487490. [76] K. 1995. Chenming Hu. University of California at Berkeley. D. Senba. 7. Morisaki. A.” Annu. 1999. Ph.” in IEDM Technical Digest. D. S. M. B. Dahringer. W. Walker. E. vol. “Thermal analysis of vertically integrated circuits. Leibenguth. SangiovanniVincentelli. E. Akasaka.” IEDM Tech.” Proc. Takahashi. Dig. pp. A. Rev.” Proc. “3D ICs interconnect performance modeling and analysis. Kumamoto. “On thermal effects in deep submicron VLSI interconnects. Sci.1986. P. “Threedimensional memory module”. May 2001. D'Asaro..D. 111114. Kuhn. 1995 [79] D. Souri. Ju. S. 36th ACM Design Automation Conference. [71] T. T. Chi On Chui. [72] A. Sugahara. IEEE Trans. 1998 [70] Y. 29: pp. 1987. 1999. 885891. dissertation. R. pp. on Advanced Packaging. Stanford. 1999. Sylvester. Ford. Weber. Woodward. Nishimura. Dissertation.Error! No text of specified style in document. B. Kusunoki. Hu.. “Analytical modeling and characterization of deepsubmicrometer interconnect. Kossives. IEDM Technical Digest. K. “Three dimensional IC for high performance image signal processor. pp. L. Nishimura. 634664. Y. K. 2001 [78] S.” Ph. Nakagawa. D. Akasaka and T. A. G. P. Tokuno. Souri. Banerjee.D. Tseng. “3D integration of MQW modulators over active submicron CMOS circuits: 375 Mb/s transimpedance receivertransmitter circuit. IEEE. no. [73] S. Y. Feb. Hui. K. pp. “Heat conduction in novel electronic films. N. J.” Electron Devices Meeting. 488491. Aplin. Nakaya. Goodson and Y. 120 . [74] K. and Y. 2003. S. 261293. Krishnamoorthy. T.” IEDM Tech. A. Chirovsky. “Concept and basic technologies for 3D IC structure. A. Dig. and C. [69] N. J. 11. 12881290. Kleiner. F. Y. Miller. pp. [75] K.. and W. L. K. Mehrotra. A. Banerjee. F. I. S. “Thermal analysis of heterogeneous 3D ICs with various integration scenarios. 2001. Saraswat. J. M. Mater. Inoue. K. C.” IEEE Photonics Technology Letters.
” Selected Topics in Quantum Electronics. R. Daniel A. Favire.” IEEE JLT. A. M. pp. Heieh. IEEE Journal of. pp. Miller. pp. Krebs. 18. Hwang.Error! No text of specified style in document. Coldrene.3µm quantumdot lasers for telecommunication. [80] D. 2001. Cohen. “Reduced temperature sensitivity of the wavelength of a diode laser in a stressengineered hydrostatic package. TeinPei Lee. 1994. Cohen et al. pp. 8. Jun 1995.3µm AlxGayIn 1xyAs/InP strainedlayerquantumwell lasers for subscriber loop applications. Bhat. Boudreau. [85] F. P. F. Masut. Mark E. 4. N. D. M. N. and Larry A. Volume: 30. [81] A.Nov. no.Dec. “Highperformance uncooled 1. B. N.3µm GaInAsPInP strainedlayer quantumwell lasers by use of temperature dependent reflectivity (TDR) mirror. 121 . 293300. “Temperature determination in optoelectronic waveguide modulators. B. pp. Issue: 2. 813818. “Very high characteristic temperature and constant differential quantum efficiency:1. Volume: 6.. J. vol.. [82] CE Zah. Yokouchi.”Selected Topics in Quantum Electronics. [86] M. “Optical interconnects to silicon. Koza. vol. A. Kasukawa. Darby.1312 – 1317. Flanders. Cohen. Dig. vol. IEEE Journal of. Issue: 2. Klopf.” APL. Forchel. N.” IEEE PTL. IEEE Journal of.” Quantum Electronics. pp.C.A. Wei Lin Wang. Feb. Iwai. M. Volume: 1. “Hightemperature operating 1. 2000. Reithmaier and A. 511523. J. 455457. R. R. Allard. Issue: 6 .C. [83] D. Pathak. Andreadakis. 2000. A. 303304. no. A. 69. 6. N. Zheng Wang D. Heimbuch. 13. Yamanaka.M. 764766. no. OFC ’97 Tech. D. 1997. 1996 [84] D. pp. J.