You are on page 1of 13

628 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 36, NO.

4, APRIL 2017

Probabilistic Wire Resistance Degradation
Due to Electromigration in Power Grids
Vivek Mishra, Student Member, IEEE, and Sachin S. Sapatnekar, Fellow, IEEE

Abstract—Electromigration (EM) is a growing concern in While there has been intensive research on understanding
on-chip interconnects, particularly in the computing and auto- the physics and modeling of EM at the level of a single
motive domains. EM can cause wire resistances in a circuit to interconnect, there is a large gap between what is known
increase, which may result in circuit performance failure within
the lifetime of a product. Classical circuit-level EM models have at this physical level and the knowledge that is used at the
two drawbacks: first, they do not accurately capture the physics circuit level. Traditional EM analysis is based on failure cri-
of degradation in modern copper dual-damascene (Cu DD) met- teria measured under accelerated aging conditions through
allization and second, they fail to model the inherent resilience in the application of temperature and current stress in a set of
a circuit that may allow it to continue to function even after some interconnect test structures. Under such conditions, failure is
wires fail. This paper overcomes both limitations, and develops a
method to analyze the effect of EM on the resistance of wires in a deemed to occur when the interconnect resistance crosses a
circuit. For a single wire, our probabilistic analysis encapsulates predetermined threshold, and the time at which this occurs is
known realities about Cu DD wires, e.g., that some regions of defined as its time-to-failure (TTF). The TTF parameters from
these wires are more susceptible to EM than others, and that accelerated aging conditions (of the order of hours) are extrap-
void evolution shows statistical behavior. We develop a new cri- olated to normal operating conditions (of the order of years)
terion for identifying mortal wires based on this analysis. One
part of the criterion incorporates the achievement of steady state using Black’s equation [6].
and provides a result that is slightly tighter than the traditional For a specific circuit, EM failure criteria may be derived by
Blech criterion, while another part is tied to the lifetime of the a circuit designer. In the traditional design flow, some wires
system. We apply these ideas to the analysis of on-chip power in a circuit may be considered immortal, i.e., immune to EM.
grids and demonstrate the inherent robustness of these grids that The root cause of immortality is related to a back-stress that is
maintains supply integrity under some EM failures.
created after the movement of atoms due to EM, which induces
Index Terms—Copper (Cu), electromigration (EM), lognormal, a back-flow of atoms to counteract the forward EM-induced
power grid, probabilistic, redundancy, resiliency, voids. atomic flow: when these two balance, the wire is immortal.
The Blech criterion [7] states that for a wire with current den-
sity j and length L, if the product ( jL) is below a threshold,
I. I NTRODUCTION then it is immortal. For mortal wires that do not satisfy the
LECTROMIGRATION (EM) in interconnects occurs due Blech criterion, Black’s equation is applied to derive maxi-
E to the movement of metal atoms, activated by momen-
tum transfer from collisions with free electrons [2]. When
mum current density limit rules, declaring a wire as having
failed if it does not adhere to the limit.
bounded by a blocking boundary, such as an encapsulat- There are several problems with such an approach. First, in a
ing barrier layer that prevents metal diffusion in copper real circuit, the impact of such failures is context-dependent.
dual-damascene (Cu DD) wires, this movement causes a deple- In some cases, small changes in resistance may cause per-
tion of atoms at the cathode, which eventually leads to void formance failures in the circuit even before the resistance
nucleation and subsequent growth [3], [4]. Cu DD structures crosses a large threshold, e.g., on critical paths in a clock and
are particularly susceptible to EM since the critical stress for data networks in circuits [8]. In other cases, a large failure
void nucleation in such wires is observed to be very small [5], may be tolerated due to the inherent resilience in the circuit,
and voids may form early in the lifetime of a design. e.g., due to redundancy in a power grid, where the failure of
one wire may be compensated by current flow through other
Manuscript received September 23, 2015; revised January 22, 2016 and paths, which results in a more resilient performance of the
April 18, 2016; accepted June 5, 2016. Date of publication June 23, 2016; circuit [1], [9]. Therefore, the use of a single threshold for the
date of current version March 17, 2017. This work was supported in part resistance change may either be excessively conservative, or
by the National Science Foundation under Award CCF-1421606 and Award
CCF-1162267, in part by the SRC under Contract 2012-TJ-2234, and in part not conservative enough, depending on how the threshold is
by the Doctoral Dissertation Fellowship, University of Minnesota. A prelimi- chosen and on the sensitivity of circuit failure to a resistance
nary version of this paper was presented in [1]. This paper was recommended change in a specific interconnect. Second, for Cu DD intercon-
by Associate Editor T. Kim.
The authors are with the Department of Electrical and Computer nects, complexities in the manufacturing process imply that a
Engineering, University of Minnesota, Minneapolis, MN 55455 USA (e-mail: straightforward extension of prior EM approaches is untenable.
vivek@umn.edu). A previous study has indicated that the Blech criterion-based
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. approach, which was originally derived for aluminum intercon-
Digital Object Identifier 10.1109/TCAD.2016.2584054 nects, is not valid for copper interconnects, and that some lines
0278-0070 c 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

and then conclude this and |(∂σH /∂x)| represents the stress gradient along the line paper in Section VI. the EM-induced vd = eZeff ρ j (5) resistance change in the wire. we overview the physics of EM in Section II. The probabilistic viewpoint   reflects both the fact that EM mechanisms are stochastic. kB T ∂x  cuit simulation. from the anode (site of nect microstructure parameters that influence void dynamics.   Deff    we calculate the circuit performance degradation by observing N eZeff ρ j . move in the direction of electron flow through one or more cies has been utilized in the industry. as before. and (4). j is the current density through the wire. our Using (1). i. resulting in tensile stress. (6) the IR drop variation in standard power grid benchmarks. The magnitude of this reverse atomic flux is given as Our new mortality criterion is presented here. This stress gradient results in described in Section III. as discussed in Section IV. is caused by transfer of momentum from moving electrons to These peculiarities for Cu DD wires indicate the need to the atoms as current flows [6]. the net atomic flux due to the combined effect of electron wind and back-stress is formulated as II. and has been pointed diffusion pathways. the drift velocity can be rewritten as framework uses an analytical model to predict the distribution Deff    of EM-induced void growth and consequently.. physics-based statistical method. and environmental parameters ode. wherein The diffusion of atoms causes metal depletion at the cath- we discuss the material. we have probabilistic EM-induced degradation is not captured. Our EM model. σH is the hydrostatic stress. resulting in compressive stress.e. as a result of mass transfer from cathode to anode. circuit. The atomic flux due stress. The drift velocity. A new mortality criterion is kB T developed. The authors indicate that the variation in EM lifetimes to the electron wind force from the cathode to the anode is is caused due to the variation in the structure of the metal given by [2] interconnect at the microscopic level. The idea of accounting for vd . such as activation energy and grain size.e. EM M ODELS    Deff  ∂σH  The fundamental phenomenon responsible for EM in metal Jflux = N EM   eZeff ρ j −   . microstructure. The work in [11] built up on [12] to consider and. Deff We propose a probabilistic EM analysis framework that μe = (4) incorporates the impact of variation in the material properties kB T of the interconnect on EM performance of a circuit. (2). and to enable probabilistic circuit analy. The metal atoms to be displaced from their lattice sites and to resilience of power grids to EM due to structural redundan. In Section V. kB T Next. and Fe = eZeff E = eZeff ρj (1) that the number of interconnects on a chip is large enough that where e is the elementary electron charge. The electron wind force. Nvd = Nμe Fe (3) Their consideration of the statistical nature of EM focuses on evaluating the mean of the EM-induced circuit performance where N is the atomic density. combining (6) and (7). and T is the temperature. depletion and accumulation. and thereby predicts the probabilistic evolution of EM-induced in a direction that opposes the flux due to the electron wind [7]. Further. (8) interconnects consists of the electron wind force that drives k T B ∂x  . incorporates the variation in intercon.. i. using the shift. Therefore. is develop EM models that incorporate the physics specific to a function of the electric field and material properties of the such interconnects. Starting where Deff is the effective diffusivity for EM. we numerically evaluate our approach. The impact of     Deff  ∂σH  EM-induced wire resistance change on circuit performance N   (7) parameters is observed by performing Monte Carlo (MC) cir. voids and the consequent probabilistic change in resistance. its microstructure. based on the underlying statistics that drive EM. electron wind force makes it possible for thermally-activated but it was partly based on the traditional Blech criterion. Fe . and accumulation at the anode. where ρ is the resistivity of Cu these directions. and the atomic flux due to electron wind in (3) becomes Based on the distribution of the change in the wire resistance. interconnect and has the form sis in a context-sensitive manner.e. which is based on a centration. The some EM issues beyond the conventional Black’s equation. and There have been few prior publications that have explored E = ρ j is the electric field. Zeff  is a con- these effects may be manifested over the chip. where  is the atomic volume. but these methods were vd = μe Fe (2) primarily based on Black’s equation. a stress-induced backflow of atoms. The work in [9] uses a physics-based model that incorporates the role of mechanical where μe is the mobility of metal atoms.MISHRA AND SAPATNEKAR: PROBABILISTIC WIRE RESISTANCE DEGRADATION DUE TO EM IN POWER GRIDS 629 fail probabilistically even if they satisfy the Blech criterion on atoms from the cathode end of the wire to the anode.. i. of atoms moving due to the electron wind is redundancy in power grids while analyzing EM was subse- quently also used in [13] and [14]. compressive stress) to the cathode (site of tensile stress). and the variation in the circuit performance shift due to Nernst–Einstein relation as discussed in [15]. kB from the known distributions of parameters related to the is Boltzmann’s constant. [10]. producing regions of uneven atomic con- out in the precursor to this paper [1]. that influence the EM-induced void dynamics. stant that represents the apparent effective charge number. This force the value of their jL product [5].

At present technology nodes. we assume surface diffusion is the dominant component since connect. for our power grid simulations in els of magnification. regions of those simulations are performed on wires with a bamboo-like similar physical properties that are visible at microscopic lev. for advanced technologies. interplay between electron wind and back-stress forces is DS . Deff . resulting paths. process enhancements such as the introduction We begin by outlining a typical Cu DD process. 4. Recent For a line of height h and width w.     we will use the model from [3]. and the Cu used to construct the inter. These void growth. 36. δI . Next. a trench is first etched the EM diffusion along the capping surface [18]–[20]. For previous technology nodes. showing grains within dominant component in order to be consistent with the pro- the metal wire and the regions of intersection of the grains. and DB are diffusion constants. The dynamic where the subscripts correspond to the diffusion pathway. given as    the effective thicknesses. The role of the The correct expression to be used for EM modeling depends barrier is to prevent Cu from diffusing into the ILD. into the interlayer dielectric (ILD). Fig. we compare our results against the 3-D EM The structure of a Cu DD interconnect is key to under. However. commonly referred to as 2 1 1 Deff = DI δI + + DS δS the Korhonen model.. ∂σH ∂ ∂σH dinal direction (Fig. G = eff . The most popular approach • Through the capping surface (S). is based on the work from Korhonen et al. an important parameter forms statistical EM analysis for single Cu DD interconnects. DGB . of CoWP as an alternative material to SiX NY . depending on the properties of the interconnect equations under appropriate boundary conditions. a Ta-based barrier Therefore. At the posi. available for the migration of metal atoms are as follows: • Along the grain boundaries. where as noted known as the grain boundaries (GBs). In this paper. additional terms can be omitted. ation. have retarded tion where the wire is to be located. ing the physics of EM in circuits. 1 is a schematic that illustrates the Section V-B. defined previously in Section II. For comparison with the FEA results in [4]. B. A. in a largely polycrystalline structure [17]. on the context with respect to the microstructure of the wire. There are multiple approaches that can be used for model- • Through the nonideal Ta barrier interface (I). supplemented with a compact modeling approach for tions of atomic transport along various diffusion paths. the first one .g. APRIL 2017 microstructure.630 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS. This polycrystalline structure introduces a large number of grain boundaries. recently it is observed that the Cu microstructure has become increasingly polygranular. δS . the microstructure is observed to be polygranular. the Cu microstructure was composed of bamboo-like grains. we assume the grain boundary diffusion as the cross section of a Cu DD interconnect. where κ = DeffkBBT  . grain structure. connect is then deposited within the trench. w h h   The Korhonen model uses the atomic flux relation in (8) 1 in conjunction with the continuity equation to represent the + DGB δGB + nBV DBV (9) d dynamics of EM atomic flow in interconnects. sion along the capping surface and grain boundaries [15]. For instance. and B is the effective bulk Typically. VOL. EM Diffusion in Cu DD Additionally. in (8) that affects EM-induced performance degradation. Cu DD interconnect cross section showing possible atomic diffusion the Cu microstructure is dominated by fine grains. model-based simulation work in [4] in Section V-A that per- standing the effective diffusivity. The work in [3] also proposes solutions to this differential Furthermore. and nBV is the fraction of atoms diffusing through eZ  ρ j the bulk volume. the primary diffusion is deposited therein. d is the grain size along the longitu. other terms have been diffusion contribution can be neglected compared to diffu. e. i. This using finite element analysis (FEA) for simulating void nucle- parameter can be considered as the sum of the contribu. 1. NO. path for EM is along the grain boundaries.e. it has been observed that the bulk and interface modulus for the Cu and ILD system. DI . The diffusion pathways above. 1 shows di which is the grain size of the = κ +G (10) ∂t ∂x ∂x ith grain). the grain boundary component can be neglected in intercon- nects that have a bamboo-like grain structure wherein the grain boundaries are perpendicular to the width of the wire [15] and there is no continuous path for EM-induced atomic diffusion. In contrast. EM-Induced Void Evolution • Through the bulk volume (BV). and δGB denote described by a differential equation. Finally. and can be modeled as [16] concentrations. In this paper. rendering surface diffu- sivity as the dominant EM diffusion mechanism. which is the arrangement of grains. [3]. adding more potential paths for grain boundary diffusion. the lines are capped above by a silicon nitride (SiX NY ) layer. cess conditions in state-of-the-art technologies. Deff can be considered works in power grid analysis have proposed EM modeling as a sum of contributions of atomic transport along these four approaches based on evolution of vacancy and plated atom pathways. the model used in [9]. Fig.. diffusion paths are a function of the microstructure of the inter. with a mixture of bamboo-like and polycrystalline microstructure.

as shown (i. Although. when Deff = D0 exp − (14) the void spans the entire interconnect. For spanning Observing the form of (14). Mechanisms of void evolution for the (a) via-above and (b) via-below ing in increased wire resistance since the void effectively cases (microstructure not shown for simplicity). in fact. from its lattice site and migrate in the direction of elec- mented during the wire routing stage. and (13). in power grid circuits the interconnects After nucleation. drift velocity. to −tn . electron flow is downward. [21]. 2) Void Growth: After nucleation. it is easy to diffusion. In Section III-D. tn . Fig. tn . metal atom needs to overcome in order to dislocate and move gate the possibility of slit void formation. is guaranteed-pessimistic. the wire. the void begins to grow. result. 1) Void Nucleation: Under EM. depending the impact of current flowing through neighboring wires can on the direction of the current [12]. [24]. the depletion of atoms at the cathode creates a tensile stress. there is an exponential increase in . as given by of micrometer. The analytical solution from [3] states that the   Deff  time. 2. Lvoid . as dis- and necessitates different modeling approaches. once a void is formed. We make this choice in the void size is due to drift under an electron wind force. the via is located above the void). shown in Fig.e. For void nucleation. such as the Ea represents the energy barrier that a thermally activated introduction of new barrier technologies [23] completely miti. these equations are valid for single where σc is the effective critical stress for void nucleation. 2(a) corresponds to the via-above case (i. voids inside the via trench trends. in the upper wire. Various void growth are connected through vias. isolated interconnects. typically at a corner of the wire or within while vd and Lvoid reduce as temperature increases. tn = (11) Deff The set of equations for EM dynamics are computationally π σc2  kB T efficient. and the stress predicted is pessimistic [3]. in agree- ment with the Korhonen model [4]. we illustrate that scenarios have been observed in Cu DD structures. However. and constitute the foundation for EM physical models in older Al and current-day Cu interconnect metallization technolo- gies [4].e. our attention on the impact of spanning voids. vd . voids evolve in two phases. and void length. because the semi-infinite solution has a simple form which with a constant drift velocity vd [4]. [12]. and the other corresponding to a finite length wire. accelerated temperature EM experiments indicate the opposite Through process enhancements. the nucleation time. the void has been growing for a length of time. is easy to compute.. At first sight. we focus in the direction of electron flow. Moreover. (11). Although slit cussed in Section II-A. the primary mechanism of growth temperature increases. For the case where the electron flow is upward. an upstream void is potentially formed in (5). experiments indicate that EM failure in Cu interconnects occurs in two phases. and slit growth. time to . and closely If the void nucleates at time tn . it is clear that as the voids. the void nucleates. this is often imple. we notice that tn appears to increase. then at an observation matches the accurate stress solution in the region of inter. further movement of metal atoms from the void results in void growth. and C. and Ea is the activation energy for EM voids tend to form earlier than spanning voids. est for power grids. if we observe the temperature dependence of as illustrated. at which a void nucleates Lvoid (to ) = vd · (to − tn ) = eZeff ρ j (to − tn ) (13) kB T Ktn where the second equality follows from (5). recent process tron flow along the dominant diffusion path(s). enhancements in Cu DD metallization technology. Effect of Temperature results in potential void formations at the via or in the wire. According to this. where kB T it forms along the via (for the via-above case)—is different where D0 depends on the microstructure of the wire. For both Al and Cu. exponential dependence of EM diffusivity on temperature [12] The mechanics of void formation in each case—spanning   Ea growth (for both the via-above and via-below cases). This apparent anomaly is explained by examining the have been resolved [22] and are not considered here. The result- ing analytical formulas can help to predict the conditions for formation and growth of EM-induced voids in a wire. 2(b). the via is located below the void). Alternatively. The scenario where the be modeled using the idea of current divergence. Once this exceeds a critical stress threshold value.MISHRA AND SAPATNEKAR: PROBABILISTIC WIRE RESISTANCE DEGRADATION DUE TO EM IN POWER GRIDS 631 corresponding to a semi-infinite approximation of a wire. we choose the solution from [3] cor- responding to the semi-infinite wire.. Therefore. reduces the cross section available for current flow. which signifies the amount of energy that a ther- build redundancy into the power grid to guard against slit mally activated metal atom must gain in order to dislocate voids by inserting redundant vias. shown in Fig. where wire lengths can go up to hundreds The length of the void increases due to drift. This has where Ktn =   (12) been the basis of its usage in other circuit analysis works such 4 eZ  ρ j 2 B eff as [11] and [12].

we will flowing through the interconnect. and to accurately analyze performance degradation due to EM. respectively. The parameter. width of the wire. [29]. we work with the idea of the equilibrium in the steady state so that the critical stress σc for effective activation energy for each wire. NO. and showing that EM process. regardless of Joule heating. In the remainder of this section. lines are apt to show probabilistic behavior [5]. Lwire and w denote the length and the energy [4] and/or grain size [27]. In this state. 4. which implies correlation between the experimental EM lifetimes and another that the stress is below the critical stress for void nucleation. the two forces could be in layer [28]. VOL. However. and (13) reported above. Statistical Models for Void Dimensions where Based on the assumptions mentioned in the previous section  ( jL)crit = 2σc /(eZeff ρ). EM activation energy can vary was predicated on the interaction between the electron wind between grains depending on the orientation of the grain force and the back-stress force. denotes the stress at the cathode when steady state is achieved. For a wire of length L with current we use Ea to refer to the effective activation energy for each density j. these formulations TJ depends on the RMS current through the wire [25] consider the transient as well as the steady state for stress. to Joule heating must be accounted for. which are required to calculate the EM-induced probabilistic resistance increase. [30]–[33]. the above argument shows that small perturbations interconnects. For instance. jL ≤ ( jL)crit (18) B. In a circuit context. that the immortality property does not always hold for Cu DD In fact. we will now present a probabilistic framework . immortal to EM effects. before and thereby accelerates EM. Rθ . (19) about Ea and d. that is microstructure property. with with respect to each other and with respect to the interfacial low current and/or small length. implying that these wires are activation energy value for that wire. σpeak [28]. P ROBABILISTIC M ODELING OF EM FAILURE the microstructure. Kins is the nect. since the logical derive the formulas that predict the statistics of EM-induced blocks which drive current are assumed to have steady state void evolution as a result of the possible sources of variation. 36. and we where the stress does not change with respect to time and the model it as an independent Gaussian random variable. a linear relationship where ∂σH /∂x = 2σpeak /L. which causes an increase in the temperature its jL value and it is possible for voids to nucleate early. uncertainties in the microstructure parameters of an intercon- nect and is a function of the metal layer occupied. thus resolving the apparent anomaly Clearly. In effect. 2) Grain Size Variation: Experiments in [27] have shown a A wire is immortal at equilibrium if σpeak ≤ σc . defined as the From (8). further. Therefore. [26] discussed in where Tref is the reference temperature of the interconnect and Section II-B. This nonzero failure probability is our motivation for using We use the model described in [25] to evaluate the inter. characterized where R is the wire resistance and Irms is the RMS current by (11) and (13). In our case. [10].  ρ j ≤ 2σ /L [7]. represents the thermal impedance and is given by tins A. which is an averaged void nucleation is never reached.632 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS. to quantify the perfor. the RMS current flowing through the This analysis will allow us to evaluate the resistance change wire is taken as the corresponding dc current. is a property of III. as a result of nucleation and growth of voids. 1) Activation Energy Variation: As discussed in Section II-C. providing the opportunity for opposing back-stress forces to mance impact of EM accurately. this occurs when eZeff c grain size value averaged over all grains in a wire. dc currents. an analytical model for probabilistic EM. For some interconnects. Therefore. It has been experimentally observed that Ea can vary within the wire depending on the grain boundary The conventional explanation of EM in Al interconnects orientation. Ea . We construct our model by attributing the causes of TJ = Irms 2 R Rθ (16) variation to fundamental material parameters that impact the EM-induced void nucleation and growth. as noted in Section I. the temperature increase due build up. the Blech criterion for immortality can be derived wire. the activation energy.88tins ) Recent works have observed that EM failure is correlated to where tins is the thickness of the dielectric below the intercon. APRIL 2017 effective diffusivity. This parameter has been from (8) by considering the steady state with respect to stress experimentally found to be normally distributed. Void evolution char- connect temperature due to Joule heating given as acteristics. Unlike the Blech criterion. (11). and using an lower value of critical stress. Role of Microstructure in Probabilistic EM Behavior Rθ = (17) Kins Lwire (w + 0. σc for Cu DD compared to Al accurate temperature value is of utmost importance in order interconnects. therefore. This can be attributed to the observation of in T can offset EM calculations significantly. the effective grain size. it is reasonable to assume that the effective activation energy it can be shown from (10) that the stress profile settles to is same for a wire and varies only between the wires [4]. which results in voids being formed easily. it has been observed causes more damage at higher temperatures. current flow in a wire can result in every line has a nonzero probability of failure. ( jL)crit is a property of the material and the fabrication in (5). Here. forward and backward flux are in equilibrium. which relate to the statistical distribution of the activation thermal conductivity. In the rest of this paper. are imported from the nucle- T = Tref + TJ (15) ation and growth model formulas [3]. taken at a macroscopic level.

Neglecting the other mechanisms correlation coefficient r can be evaluated as and considering the dominant term corresponding to the grain   mean(tn )variance Deff.n (22) a spanning void. and such a product results in a lognormal distribution. μD0 = log(D0GB δGB ) − μd and the standard deviation. Specifically. σ ). σ ) is another R ρTa ACu Lvoid (to ) = −1 . Therefore.g ) then we have mean and standard deviation of the underlying Gaussian. σ ) with variable with a scalar results in another lognormal random the mean μ and standard deviation σ . where in Wilkinson approximation. we can mals. where c1 = (eZeff  ρjt )/k T and c = c /t . its length at given observation time. we use the above. since c1 is a deterministic constant and void nucleation and growth. The correlation between the random variables c1 Deff. plays Therefore.g tn δGB D0 = D0GB d where the mean() and variance() for the corresponding lognor- or log D0 = log(D0GB δGB ) − log d (20) mal variables. Due to this. and the fact that Ea for a wire follows a Gaussian distribution. R0 = ρCu Lwire /ACu is the resistance of the deterministic parameters in this equation. The first term. Deff.g + σt2n . 2) Nucleation: The expression for nucleation time tn was provided in (11). the change in resistance. using a moment matching approach.e. Lvoid (to ) = LogN(μLvoid (to ) .g . similar to the widely- we investigate the statistics of Deff . respectively. R0 ρCu ATa Lwire 3) Growth: During void growth. we have r = c1 c2       variance c1 Deff.MISHRA AND SAPATNEKAR: PROBABILISTIC WIRE RESISTANCE DEGRADATION DUE TO EM IN POWER GRIDS 633 that will enable us to obtain the EM-induced void character. Lvoid (to ) is a difference of two correlated log- a key role in the EM degradation of Cu DD wires. and the product of a lognormal random In our discussion below.   tion of a reciprocal of a lognormal LogN(μ. From (21) and (22). then Deff = LogN(μDeff . We refer to the effective diffusivity for the nucleation and growth phases as Deff. For the case of the spanning void. Recall that the Lvoid (to ) = c1 Deff. if μX . Grouping together all per and tantalum. ρCu and ρTa are. used Wilkinson approximation [34]. σY ) are the mean grain size. σEa ). and the mined by this mechanism. σL2void (to ) = log(u2 ) − 2 log(u1 ). where μd and σd represent the for c1 Deff. and affects the physics of void nucleation and normal random variables is approximated by a lognormal growth. if a void nucleates. c2 tn Deff.g is a scaled product of X = eZ as LogN(μ. 1) Effective Diffusivity: The effective diffusivity. D0 = LogN(μD0 . is lognormal. and ATa is the cumulative the form cross-sectional area of the tantalum barrier. From this. d = LogN(μd . σD0 ). one attributable to D0 (via d) and the other attributable use the above set of equations to compute the parameters of the to Ea .g . where the mean.g − c2 tn Deff. approximate the sum of two correlated lognormal random vari- anism in modern Cu DD interconnects is grain boundary ables.g (23) void length at time to . which has been used to As indicated in Section II-A. σX (μY . σtn ).g . σD0 = σd . respectively. Deff .n Specifically.. as discussed normals. and standard deviation of the underlying normal distribution i. the resistivities of cop- evolves with time according to (13). it is clear that tn = LogN(μtn . This shows and variance of the underlying Gaussians. we denote a lognormal variable.g and c2 tn Deff. c1 Deff. σd ). was shown in Section III-B . σDeff ). σLvoid (to ) ). (24) lognormal characterized as LogN(−μ. the parameter D0 in (14) is deter. μY = log c2 + μDeff. we consider the impact of each of these individually. (21) u2 = e 2μX kB T  2  σX +σY2 +2rσX σY μX +μY + 2 The diffusion pathways during nucleation and growth may − 2e be different [4]. C.g boundary (GB) mechanism in (9). the primary diffusion mech. log d. σ ).g (c2 tn Deff.g is required for evaluating the statistics. of void length. it is easy to compute that D0 is a random variable that depends on the effective these quantities. is given as [24] the above result relies on the observation that the distribu. to model the variation in Deff .g is lognormal. distribution by applying the moment matching approach used If Ea = N(μEa . In order to find the analytical expression for the distribution From (14). μX = log c1 + μDeff. to . Statistics of the EM-Induced Resistance Change where We will now use the void length distribution to evaluate the μtn = log(Ktn ) − μDeff.n distribution of resistance change due to EM-induced growth of σtn = σDeff. the length of a void Here. d. and c2 tn Deff. the second term.g + μtn σY = σD2 eff. R. Lvoid (to ). interconnect wire segment. The difference of two log- in Section II-A. it is obvious that Deff is the product of two lognor. o B 2 1 o istics and the corresponding resistance change as a result of c1 Deff. which follows a lognormal distribution [27].g variance c2 Deff.g diffusion. has Lwire and cross-sectional area ACu . μLvoid (to ) = 2 log(u1 ) − log(u2 )/2 and next. which provide the mean where D0GB is a temperature independent constant.n σX = σDeff. lognormals. which is also a lognormal.n and Deff. as indicated in (11) and (13). which is assumed to have length then from (13). This will lead to the following μEa set of equations: μDeff = log(D0GB δGB ) − μd −     kB T μX +σX2 /2 μY +σY2 /2   u1 = e −e     σEa 2 +2σX2 + e 2μY +2σY 2 σDeff = σd + 2 2 . Therefore. respectively. for a distribution Z = N(μ.

effective divergence. A via in a Cu DD intercon- σ R are the mean and standard deviation of the underlying nect structure acts as a blocking layer so that metal atoms are Gaussian. this implies that there is an grow to the point where the wire resistance increases by R. which can eventually lead to void nucle- Using these notions. defined as   ρTa ACu R0 kR = −1 (25) ρCu ATa Lwire we can rewrite (24) as Fig. In other words. equivalent to a current density of 2j on both that a given wire will have a resistance change R as wires. as compared to the case where the left- . VOL. which shows two wires on metal layers then be expressed as R = LogN(μ R . This is variable with a scalar. 4.e. i. (26) Since kR is constant in (26). is away from the via for both segments. R = kR Lvoid (to ). Assuming equal cur- First. (27) In this example. The resistance change distribution can illustrated in Fig. 3. NO. If we introduce a new symbol. The effec- ning void. given by not permitted to migrate through it. for both the via-above and via-below cases. and then this nucleated void must rent densities j on each segment. Therefore. the electron flow direction achieve a resistance change of R at an observation time to . σ R = σLvoid (to ) . 36. and directions of currents in neighboring wires. a void must nucleate. where μ R and Mx and Mx+1 connected by a via. Via-tree structure where the effective current density is more than the density of current flowing through the wire. it can be seen that for a span. kR .. we can now determine the probability ation and growth. the current on both segments of the wire on We summarize the conditions that must be satisfied to layer Mx flows toward the via. 3. similar to the work in [35]. since it is the product of a lognormal random flux-divergence criterion. σ R ). APRIL 2017 to be lognormally distributed after nucleation. any flux that would have gone to the via is transmitted to a neighboring μ R = μLvoid (to ) + log kR wire. R tive current density for a wire is computed in terms of the is lognormal.634 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS.

is predicated for a nucleated void is given by LogN(μ R . tality prediction by the Blech criterion cannot be guaranteed The second case corresponds to the scenario where the void unless this steady state is reached soon. Blech criterion ( jL < ( jL)crit ). we have We model adjacent segments by computing the effective current density on a wire. Mortal Wire Prediction ity of nucleation. a finite time. by showing test circuit where one wire has twice the L2 current density as another. which is based on the magnitude jL > ( jL)crit . we propose an alternate shows fabricated test structures where the failure rates on a void-nucleation-based mortality criterion. to balance the forward EM flux. For wires where the current divergence effect comes into play: for example. [36] steady state may not be achieved. E. for two time elapsed for achievement of back-stress. 1 − Pr(nuc) if R = 0. However. evaluated at time to . based on extrapolating experimen- The conventional approach to estimating EM failure is based tal EM and back-stress results from [37]. The achievement of steady state and Pr(nuc) corresponds to the probability that the nucleation requires significant mass transfer from the cathode to anode time. of the actual current density of j. can be backed D. as in (10). Pr(nuc). tn . with a probability of 1 − Pr(nuc). which may be comparable to circuit lifetime. segment. The probability distribution for R The Blech criterion. should have a shorter mean time to failure. i. the work According to [3]. the one with the larger current density multiple wire lengths. with the mean and standard deviation given by (27). wind and back-stress forces. Incorporating the Effect of Current Divergence up by experimental data. the product of Pr( R | nuc). and strictly speaking. immor- lative distribution function (CDF) of tn . hand segment is missing and the right-hand segment has the Pr( R | nuc) · Pr(nuc) if R > 0 same current density of j. For mortal wires. but experiences consistently later tss = (29) 4κ failures. outlined in Section III. which observe the on a current density-based model. i. The observation that does not nucleate. the approximate time taken to reach steady in [35] demonstrated experimentally that this is not always the state can be written as case. (30) . to . the expected rate of atomic transfer Pr( R) = (28) is doubled at this node.e. in Cu wires of wires of equal length. σ R ) as derived on the achievement of a steady state between the electron above. so that the stress gradient set- The nucleation probability is given by LogN(μtn .. Using the via node vector notion [35]. This is consistent with other reported work where the where κ = (Deff B /kB T).e. the probability of a resistance change given that nucleation has occurred. but also on those on adjacent segments that share Case 1: A wire can be EM-mortal if it does not satisfy the via(s) with this segment. and the probabil. tles to a linear trend in space. is elapsed before the achievement of steady state. is less than the observation time. an effective current density of 2j is used for this wire instead The first case corresponds to a nonzero resistance change.. Under this model. considering three wire segment depends not only on the current density on the cases. the cumu. σtn ) in (22).

3σ ) point of the underlying Gaussian. √ the criteria under all three cases uniformly distributed over the range of x. For the mortal wires.n +3σDeff.n pling (IS). tn < tss = L2 /4κ. from (21).n +3σDeff. 11: Build a circuit sample for the power grid using the Case 3: Let the earliest reasonable void nucleation time of a above set of wire resistances wire. Since 1: Solve power grid to obtain nominal j for each wire κ = (Deff B /kB T). σc π  kB T some values of x have a low probability of occurrence and are j>  . denoted Ep [ f (x)]. this inequality can be written as information. resulting in 10: end for a deterministic criterion. such an approach would see a large num- ber of samples go to waste as they provide little meaningful Using (22). and set of samples would probably see no resistance change in determines the voltage distribution at each node. taken to be the μ − 3σ point of the underlying Gaussian 12: Solve the sample power grid and tabulate the IR drop μ−3σ 13: end for of the nucleation time. of the circuit. given that EM is the span has a uniform probability of 1/K under q(x). given by (19). standard deviation) tively EM-immortal if this earliest nucleation time is beyond the projected lifetime.n can be Ep [ f (x)] = f (x)p(x)dx = f (x)p(x). but “unbiases” it that is Ktn < tlife e . In this paper. M ONTE C ARLO S IMULATION F RAMEWORK a uniform distribution that stretches from 0 to the tail of the We use our probabilistic resistance model to perform MC lognormal distribution of R: the values of this lognormal analysis of power grids in the presence of probabilistic resis. it is EM-mortal if most (and possibly. tnμ−3σ < tlife . tnμ−3σ = eμtn −3σtn = elog Ktn −μDeff. and carrying current j: specified in terms of a distribution p(x). we use the notion of importance sam- μDeff. However. range from R = 0 to the (μ + 3σ ) point of the underlying tance variations. derived in Section III-C builds a sim. The voltages . the EM- adding appropriate weights to f (x).e. which biases the distribution. log( R). and then correcting must be satisfied. all parameters on the right-hand computed as side of the inequality are fundamental parameters of the tech. If K is the span of this distribution from R = 0 to the (μ + facilitating simplified analysis at the circuit level by consid. mean and the standard deviation of the underlying Gaussian. Since π/2 ≈ 0. the for the resistance change. The (and should be) a relatively unlikely event.   nology. feeds them into a our MC analysis to be biased appropriately: a truly random power grid simulator based on dc modified nodal analysis. Note that in the first inequality. the criterion for the error due to sampling from this different distribution by case 1 is automatically satisfied by case 2. IS is an MC method Substituting Ktn from (12) and rearranging a few terms. be denoted as tn . from (11) and (12) 2: Calculate effective current density (Section III-D)   3: Filter immortal wires that are EM-safe (Section III-E) π σc2  kB T L2 kB T  ρ j)2 BD < (31) 4: for each mortal wire do 4 (eZeff eff 4 Deff B  5: Use the resistance evolution model to obtain √ π lognormal PDFs of wire resistance (Section III-C) that is jL > ( jL)crit (32) 2 6: end for where ( jL)crit is the conventional critical jL value from the 7: for each Monte Carlo iteration do Blech criterion. i. In other words. and can be cheaply χ χ precomputed for a technology. Therefore. 8: for each mortal wire do both the left and right hand sides are statistical quantities. that computes the expected value of a function f (x) of a ran- this translates to the following mortality criterion for a wire dom variable x. causing sampling errors. for both tion of f (x) under the distribution p(x). For example. 9: Sample the wire resistance using IS (Section IV) However.n < tlife To overcome this. our framework is shown in Algorithm 1. it is essential for method samples points on this distribution. the expecta- mortality criterion is given by (32) and (35). where μ and σ represent. both depend on Deff . then every point within ering R as a random variable.n not sampled frequently enough. which cancels out. (34) as it interprets the results of sampling. no) wires.n −3σDeff. the statistics of Deff. ple circuit-level abstraction for complex physical phenomena. IS resolves this by sampling according to a function q(x) that is For a wire to be EM-mortal. Our probability distribution function (PDF) Gaussian. without the need for any MC = w(x)q(x)dx = Eq [w(x)] simulations. This method is partic- ularly useful when p(x) is skewed or unevenly distributed. tlife ..89 < 1. The pseudocode for simulations to analyze the impact on the power grid. A wire is effec- 14: Report statistics (mean. Further. we use a sampling distribution q(x).q(x)/q(x)dx seen to be based on the underlying process. during the where w(x) = f (x)p(x)/q(x) and Ep [ f (x)] is the expectation product lifetime. for a small set of samples.e. i. which is IV. (35) eZeff ρ 4 B to eμDeff. in particular. which is defined over a set of values χ . (33) Most importantly. tn .MISHRA AND SAPATNEKAR: PROBABILISTIC WIRE RESISTANCE DEGRADATION DUE TO EM IN POWER GRIDS 635 Case 2: A wire can be EM-mortal if a void can nucleate Algorithm 1 Power Grid MC Simulation before steady-state is achieved. we perform MC circuit of f (x) under the new distribution q(x). is inequalities (32) and (35). respectively.. We apply the mortality criteria (32) and (35) to χ filter out wires that are safe with respect to EM.

we revert to performing available. NO. faster than the FEA approach since it merely involves the eval. the table lists the name. and the growth time. out wires that are immortal under the Blech criterion. obtained by our analytical approximation with MC simulation model. 4 shows the compar- as discussed briefly in Section II-A. number of uation of an analytical expression. This figure demonstrates a close match between our normal for the failure parameters. observation time. the h. Specifically. Lvoid becomes equal to Lvia . we now focus on analyzing of some parameters in [4]. and 19. which computes tn . to = 12 years. Power Grid Simulation discrepancies are attributed to factors such as the unavailability Moving beyond a single wire. tn and tg . Our pre.037eV. but we where the values are imported. allowing the void the value of its jL product that is traditionally used to screen to grow until it spans across the length of the via. Fig. Table I lists the EM parameters along with their our remaining evaluations under normal operating conditions. as observed simulations in [4]. tg . and tg . for the the capping surface diffusion is the dominant EM mechanism. against the corresponding values obtained from FEA. which uses bamboo-like grains where according to our probabilistic formulation in (28). threshold can effectively be considered immortal. Under the assumptions in [4]. The 3-D models used by metal layers occupied. These numbers indicate that time elapsed from nucleation up to the instant at which length every wire has a nonzero probability of failure. μ. The values show a good level of consistency. where it is of the order of several hours. which is erated aging. and the The standard deviation for the normally distributed Ea is TTF is now of the order of several years.6 h. lations. Clearly. of a wire simulation setup in [4]. density value of 1 × 1010 A/m2 . The B. there is very low probability that a specific wire will abilistic framework derived in Section III under accelerated fail in any manufactured part. Here. VOL. for the underlying Gaussian of the log. ison for the PDF of R/R as predicted by our formulation Table II lists the expected value. as against accel- obtained from [28]. failure is defined as the time V. σ . by using the 3σ value in Ea . R ESULTS when Lvoid = Lvia . respectively. and wires that fall below this aging conditions for temperature and current stress. 1) Calibration of Correctness Under Accelerated Aging: 2) Single-Wire Simulation at Chip Operating Conditions: The material parameters for accelerated aging are set to ensure Note that we have compared our statistics with published work a fair comparison.. APRIL 2017 TABLE I TABLE II EM PARAMETERS U SED FOR THE S INGLE W IRE C ASE VOID N UCLEATION T IME (tn ) AND G ROWTH T IME (tg ) C OMPARISON : A NALYTICAL M ETHOD V ERSUS [4] 0 the numerical simulations in the FEA approach give an accu- rate estimate of the EM dynamics. 4. due to the large simulation runtime. For our calibration simulations. of the lognormal and against a plot for the same PDF obtained from MC simu- standard deviation. even for a single wire. we distributions of tn and tg . R/R. and differences in the setup between the impact of EM on a set of power grid benchmark circuits the fully 3-D numerical simulations in the FEA approach described in [39].1 h.11eV. using a current unavailable were extracted from the literature. However. enables this level of scalability. The TTF values for the 0. 15. σ denotes the standard deviation for also perform MC simulations at these conditions to validate Ea . 36. drawing parameter values from [4] where under accelerated aging conditions. and the percentage nominal IR drop of . However. total number of wires. values. 50%ile. and is obtained as 0. We apply the Wilkinson A. As expected. corresponding to the capping surface diffusion to align with the We evaluate the resistance change ratio. 12. despite some differences in the underlying assumptions. For evaluating the corresponding values of experimentally in [5] and [10]. regardless of of the void. and the corresponding reference from We use a similar setup as described in Section V-A1. i.7%ile points under accelerated aging are 12. we use the prob. and the TTF is the sum of the nucleation time. Our work presents an approach that them to the original lognormal using the w(x) factor. and 99. the reduction B = 1GPa was used from [38]. mentioned in [4]. our method is much that were evaluated for our EM analysis. For each power grid. Some parameters that were perform this analysis at room temperature. description. first work under assumptions similar to [4].636 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS. tn . Failures in Single Wire approximation [34] to obtain TTF as the sum of lognormal To calibrate the correctness of our probabilistic model. the usage of such models for the statistical analysis of a multimillion-wire power are then translated back to the original distribution by scaling grid system is unrealistic. Table III lists the power grid benchmarks against our analytical approach.7 h. our approach using 106 samples on a normal distribution of we used the value for effective diffusivity and activation energy activation energy. below a lifetime of time elapsed for void nucleation and growth. Indeed. dicted values are then compared against the published FEA wires that satisfy the Blech criterion will fail. in temperature reduces the rate of EM degradation.e. We specified as 0.7 the time elapsed before a void nucleates in a wire.3%ile.

for convenience. we these benchmarks such that the worst-case IR drop at nominal round this upward to 1000 simulations. we provide mean and variance values that refer to formulation. it was observed that the nom. the mortal wires iden- the underlying Gaussians. a negative mean value tified by our approach is a superset of those that are mortal for the Gaussian is permissible since it translates to a positive under the traditional criterion. i. from the traditional deterministic Blech criterion. number of such wires for each benchmark in Fig. as described in [39]. NMC . We assume listed in Table V. mortal wires is larger than the traditional number. and for convenience. and list the as indicated in [18]. we scale the original values of the current loads for 95%. tlife . As derived in Section III-E. described in the largest value of IR drop among all the nodes expressed as Section IV. However. can be evaluated as was abnormally high (>20% of the supply voltage). 5. However. for power grid simulations. ized power grid simulator were to be used instead.. we have used the For these benchmarks. and therefore the number of mean for the lognormal. TABLE V E XPECTED RUNTIME VALUES FOR MC P OWER G RID E VALUATION BASED ON S INGLE I TERATION RUNTIMES F ROM [42] TABLE III P OWER G RID B ENCHMARKS E VALUATED IN T HIS PAPER We perform MC circuit simulation using samples from the EM-induced resistance increase. we C. R. Mortal Wire Prediction choose the EM process parameters corresponding to the grain We now evaluate our new mortality criterion from boundary diffusion mechanism. This table which consists of only resistive elements. we use the runtime per iteration from [42]. 4. For a confidence of Ptarget that the worst-case IR to EM degradation.e. Section V-A1. It can Table IV lists the values of the process parameters used in be seen that our implementation indicates that a larger num- our simulations. indicates that the MC-based evaluation is computationally For the accelerated aging simulations described in reasonable.7%ile value predicted by our method. defined as the IR drop value of the node with the power grid using a statistical IS approach. 99. evaluated at time 0. In our implementation. drop predicted from our MC simulation does not cross the For power grids PG1 and PG2. which provides NMC = 997. Therefore. that the number of simulations. we analyze the interconnects for EM matrix solver from MATLAB.MISHRA AND SAPATNEKAR: PROBABILISTIC WIRE RESISTANCE DEGRADATION DUE TO EM IN POWER GRIDS 637 TABLE IV EM PARAMETERS U SED IN P OWER G RID S IMULATION 0 Fig. We identify the number of mortal wires the grain boundaries is indicated as the primary EM diffusion predicted from our proposed mortality criterion with those mechanism for advanced technology Cu DD interconnects. at every node for a given circuit lifetime. to demonstrate that it risk in the power grid and simulate the distribution of the IR is realistic to run an MC simulation with 1000 simulations on drop by solving the power grid to obtain the node voltage these power grids. and briefly discussed in Section II-A. ber of wires must be considered mortal under our probabilistic normals. prior MATLAB. since EM diffusion through Section III-E. PDF of R/R for a wire under normal operating conditions.997)). the capping surface diffusion was considered as the primary EM diffusion mechanism to align with the FEA simulations from [4]. conditions is below 15% of the supply voltage. . to estimate realistic runtimes if a special- constant current sources drawing current from the power grid. it can be shown inal IR drop for the original benchmarks. We choose Ptarget = to [9]. We realize our implementation using C++ and a percentage of supply voltage. for every wire in the power grid. Similar NMC = ((log(1 − Ptarget ))/log(0. We remind the reader again that for the log.

and grain the circuit lifetime. and we choose fixed values for Ea and d. d. Vmax . Vdd . This can be attributed to the actual values chosen for mean and standard deviation of the process parameters which have been extracted from literature. For the second and third set. when we size. Besides showing the relative contributions of process parameter variations on the IR drop of power grids. our MC circuit simulation are performed using an and grain size.638 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS. simulation. The spread due to variation in both Ea and d varia- tion. which corre. the IR drop is not a fixed value but follows a distribution. 6. such as the RMS current and the ILD thickness. 36. These variations across benchmarks this distribution of IR drop values. TJ . This indicates that the variation in Ea has a relatively larger impact on the overall variation in IR drop. in the third simulation set. which induces probabilistic wire voltage. affect the distri- IR drop. bution. the spread is zero since no varia- TABLE VI I MPACT OF Ea AND d VARIATION ON THE N ORMALIZED 99. IR drop. and the spread as a function of the maximum resistance changes. in (15) to zero in our simulation. defined previously as the IR drop value of the incorporated. EM-Induced IR Drop Degradation 2) Impact of Temperature Variation Due to Joule Heating: Next.3%ile values. which leads due to Joule heating. compare this iterative approach with a noniterative method. spond to the mean values of their corresponding distributions. we define the maximum can be attributed to multiple factors that vary between the IR drop value. V. in the redundant paths of the power grid. for the power grid benchmarks. Fig. NO. Based on PG2) to 40% (for PG4). we observe that For the second set of simulations. is observed to be marginally more than the spread due to variation in only Ea .7%ile and 0. for every benchmark for each iterative approach can be attributed to the current redistribution of the three sets of simulations. For the first set. iterative approach where the wire resistance and the current for a specific lifetime tlife = 10 years. Ea . the normalized maximum IR drop value and the normalized PG1. in the sec- ond set. VOL. thereby impacting further 1) Impact of Ea and d on IR Drop: In order to quantify the EM-induced degradation. This Joule heating we repeat the third set in the previous circuit variation is a result of EM-induced increases in wire resis. where the currents are not updated after t = 0. relative to node which has the largest IR drop. and the spread. to statistical shifts in the IR drop. we perform three sets of circuit simulations. in the IR drop values. and in the third the iterative approach for all benchmarks. we do not density value are updated after incremental steps leading to consider any variation in both activation energy. but this time we artificially force temperature rise tances due to probabilistic variations in Ea and d. This results in lower . bution of current through the wires. as the 99. Effect of Joule heating on IR drop. 5. D. APRIL 2017 Fig. we tabulate the IR drop value of ues for each benchmark is lower when Joule heating is not power grid.7% ILE tion is assumed. The lower IR drop values in the spread. In order to capture the evolution of impact on IR drop due to variations in activation energy.7%ile point of this distri. These changes. since it is this node which its correct value after considering TJ . we also want to emphasize the fact that. we consider variations in the noniterative approach predicts a larger IR drop compared to the activation energy and a constant grain size. in turn. 7 illustrates the set. Table VI shows normalized maximum IR drop for a representative benchmark. and demonstrates that the maximum IR drop val- For each MC sample. as the which depends on the metal layers used. The max. 4. under statistical vari- ation of process parameters. as a function of time. For the power grid benchmarks. benchmarks. Fig. varies from 6% (for will decide the overall performance of the circuit. The percentage error in the IR drop. 3) Impact of Current Redistribution: The flow of current imum IR drop is normalized as a percentage of the supply in the power grid drives EM. we assume variation in both Ea and d. d. power grids. and is computed through an Fig. 6 shows the normalized maximum IR drop for the eight MC simulation on each power grid. Mortal wire comparison: our criterion versus traditional Blech criterion. we discuss the In order to quantify the impact of temperature rise due to statistics of performance degradation from part to part. clearly. as defined previously. the normalized VALUE AND N ORMALIZED S PREAD OF IR D ROP maximum IR drop increases and there is a clear spread in the IR drop. Ea . difference between the 99. In the first set.

8. increases to a value of 33% for tlife = 20 years as more no. thereby result- ing in larger EM-induced IR drop degradation. Phys. for is typically used by designers today.. K. Tan. the failure fraction is significantly larger than 1000ppm. For PG1 the error in the IR drop at circuit densities through wire width tuning. Note that for a well-designed power grid the fraction of samples crossing the IR drop threshold. vol. However. Our methodology can provide years lies completely to the left of this threshold. The data is generated for various interconnects and second. 10% as indicated by a variation in EM process parameters. [3] M. the IR drop is satisfactory. and automotive applications.. Appl. 9% of the samples cross World Sci. e. 4) Statistics of the Worst-Case IR Drop for the Power Grid: VI. respectively. the chosen threshold for larger values of tlife . that EM is a probabilistic phenomenon for Cu DD Vmax . This is due to the fact the power grid benchmarks are not completely optimized. This demonstrates that the providing resources that contributed to the results reported circuit lifetime. Li. pp. and 20 are substantial. that the power grid has inherent values of the lifetime. corresponding to the typical product lifetime values for an improved immortality criterion that accounts for steady- mobile. 2013. using our iterative MC approach. tlife = 10 years. can be longer than the lifetime of any of its component wires. A. 10. according to our definition of work: first. the IR drop threshold changes by only 2% in order to maintain a significantly large (98%) probability for correct circuit func- tioning. state behavior.MISHRA AND SAPATNEKAR: PROBABILISTIC WIRE RESISTANCE DEGRADATION DUE TO EM IN POWER GRIDS 639 voids nucleate and grow over this longer period. for example. C ONCLUSION In order to study the time evolution of variation of the IR We have developed a method for EM analysis of power grids drop.-Y..g. by adjusting the current than is really seen. the current stress for the EM-affected designer that the grid has EM problems. Tu. 8 shows the CDF of the IR drop at tlife = 5. 3790–3799. Depending on the product lifetime this value can be of the order 1000–1ppm Fig. should be a low value. Our results indicate that both effects Fig. USA. is observed to be 9% of the IR drop would be informed by the analysis of the type proposed here. 1–6. pp. ACM/EDAC/IEEE with lifetime since a larger fraction of the samples crosses Design Autom. tlife . P. the designer may encounter them during design ating further EM-induced resistance degradation. Electromigration in ULSI Interconnections. value at that instant.3× between tlife = 10 years to tlife = 20 years. and this value due to electromigration in confined metal lines. Singapore: corresponding to tlife = 10 years. 1993.” in Proc. Conf. University of Minnesota. IR drop evolution of PG1 (iterative versus noniterative). there is a 98% likelihood that the power grid will remain functional up to 20 years for the modified threshold. “Stress evolution the threshold. M. we perform an analysis that determines years. This reemphasizes the resiliency of the power grid to EM-induced wire resistance change. 8. In other words. for power grid PG1. and C. depending on the product lifetime [43]. For the CDF [2] C. Mishra and S. Sapatnekar. In contrast. such as critical stress can vertical line on the x-axis. resulting in functional failure. In contrast. the worst resistance increase corresponding to tlife = 10 years and tlife = 20 years to be 124% and 287%. which The authors would like to thank the Minnesota is above the typical 10%–20% resistance increase criterion that Supercomputing Institute. and this that for short lifetime applications (e. N. for long lifetime applications. which corresponds to the time during which within this paper. thereby attenu. Thus. tlife . and it is important to analyze them to inform the for the noniterative case.. We also observe Fig. Thus. S. TX. Further. and such optimization lifetime.” J. iterations. This implies a framework to address these additional variations. mobile).g. even though the EM-induced wire resistance change increases by 2. The designer can then wires remains constant predicting a relatively larger damage reduce the failure rate. respectively. grid will remain functional since the IR drop is below the threshold for all the samples. resilience to EM failures. The probability that the IR drop is larger than 10% increases [1] V. In addition to microstructural variations. Borgesen. current flow through the EM-affected wires. the For a given threshold value. thereby resulting in functional failure. computing. a significant fraction of samples will be EM-affected if the IR drop threshold is chosen as 10%. 7. On the other hand. “The impact of electromigration in copper interconnects on power grid integrity. the worst percentage ACKNOWLEDGMENT resistance degradation in PG1 at tlife = 5 years is 48%. Korhonen. 2010. we observe the CDF of the IR drop corresponding to taking into account two effects that were neglected in past the node with the largest value. . if the design specifications are changed so that the IR drop threshold criterion is relaxed from 10% to 12%. the CDF corresponding to tlife = 5 also influence the performance. Nevertheless. CDF plots for IR drop of PG1 for different circuit lifetimes. 73. the power grid is R EFERENCES robust to some EM failures in individual wires.. the power constitutes a part of our future work. Austin.

vol. pp. Fisher. Co-Optim. Yuan.Tech. Rel.. no. Chatterjee.-D. Xu.” Ph.” in Proc. “Statistical analysis of electromigration lifetimes and void evolution. R. [18] (Aug. 1573–1577..” J. Lee. and S. Phys. Monterey. 4. “Redundancy-aware electromi.-K. 2009. [37] K. India. 2005.. Kaldor.. CA. Japan. USA. IA. Phys. 5. Y. Rel.. Simulat. Liu et al. “Power grid analysis benchmarks.net/Links/2013ITRS/2013Chapters/ for quality. 1976. Devices Mater. “Electromigration [42] Z. 1803–1810. 4.” Appl.” J.640 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS. C. 1. Mehrotra. McKnight University Professorship and the Robert and Marjorie Henle Chair. 2002. X. USA. IEEE Int. Sapatnekar.. dissertation. Rel. Appl. L. Mishra and S. Phys.. 101. Symp. Sachin S. vol. Rel. Phys. F. Li and M. Symp. T. 2008. Integr. IEEE Veh. 1–2. 4. 109.” Microelectron. C. “Probabilistic immortality of Cu damascene intercon. Lloyd and J. CA. where he holds the Distinguished J.. [39] S. A. degree from the University pp. C. R..-Aided Design. pp. Champaign. Int. USA. V. C. characterization and modeling of the reliability of three-terminal dual- [11] D.” J.1–3D. 053708-1–053708-6. vol. Lett.. 1468–1472. S. V.” in Proc. 2011. Conf. 47. Huang. Department of Electrical and Computer Engineering.1–3E. [27] M.. vol. Process. 2000.-J. C. vol. 522–531.” per drift velocity and Black’s parameters extraction.-S. V. “Electromigration-resistance enhancement with Technology Bombay. vol. M.” Microelectron. 1994. J. Najm. Appl. [32] J. Res. nects. [31] V.D. 262. 2001. pp. Sep. Jain. 87. and C. 9–11.. P. Thompson. 102–104. pp. 175–179. properties of the confinement material on electromigration in metallic Comput. 16–32. IEEE. Hsinchu. 9. USA. 2014. CA.. 67–78. 013716-1–013716-9. and K. pp. Vivek Mishra (S’12) received the B. South Lake Tahoe..” in Proc. Phys. 083509-1–083509-4. USA. [22] W. Conf. M. Anaheim. no. C.. Conf. Austin. no. Conf. Tan. “Electromigration in thin aluminum films on titanium J. Najm. no. A. “Calculation of diffusion penetration curves for surface growth kinetics in electroplated Cu lines. “In situ scanning electron [12] S. Rel. “Study of upstream electromigration bimodality and its Taiwan. AIP Conf. Pey. 9. de Orio. 2015. Phys. wireless applications. 2005. 1601.. 2014. pp. Technol. Lett. “Optimizing standard cell design [Online]. Eng. Mater. 74–78. of Illinois at Urbana–Champaign. Phys. “The effect of Dr. 5. vol. 91. Alam. Beaulieu. Marek-Sadowska. 6. 104.-K. 2007. Phys. IEEE/ACM Int. [33] B. vol. H. 2117–2127. 2011. Sapatnekar. Mishra is a recipient of the 2015–16 Doctoral Dissertation Fellowship from the University of Minnesota. and D. 4. Cervenka. Pyun et al. Phys. Z. [36] C. NY. and pp. MN. “New electromigration validation: Via node vector method. analysis of power/ground networks. no. 906–910.” in Proc. Warner. Park. Kitchin. “The effects of the mechanical power grid electromigration checking. pp.-Aided Design. V. [24] M. CA. 2005. Troxel. 92. Doyen et al. for semiconductor devices. NO. 51. Thompson. SPIE Design Process Technol. 540–547. Semicond. vol.” in Proc. 3E. L. 2. and the SIA University Researcher USA. vol. He CoWP or CuMn for advanced Cu interconnects.” J. no. “Microstructure. N. IEEE Int. “Modeling the electromigration failure time distribu- [6] J. “Activation energy and the University of Minnesota. pp. USA. 2011. award. no. no. no.. 1587–1594. Ceric. Phys.” in Proc. “The electromigration failure distribution.pdf San Jose. Sep. Mumbai. USA. pp. vol. no. He was with Iowa State University. 1–4. Minneapolis. Hau-Riege and C. pp. USA. Abu-Dayya and N. Available: http://www.” J. M. and W. USA. South Korea. CA.. Christiansen et al. Vook. 2008. Ogawa et al. Appl. Pac. gration failure. Choi. R. [13] M. 36. the SRC Technical Excellence award. Mönig.itrs. 1951. 4. pp. Workshop Final Rep. in Proc.. Sukharev. 2011. resistance evolution under AC electromigration. Dept. 1999. W. Conf.” in Proc. and grain boundary diffusion. L.” Thin Solid Films. S. CA. IEEE/ACM Int.5. Phys. N. TX. Appl. Semiconductor Manufacturing Company.” in Proc. 2011. San Jose. Rosenberg. Appl. [16] E. 22. Hauschildt et al. 1797–1802. and K. grids: Algorithms and implementations. 2010. Phys. and F. Phys. [21] J. vol. Hu. 102. “Circuit delay variability due to wire tromigration in copper films. and P. and C. 488–493. Syracuse. “Effect of barrier thickness on electromigration reli- CA. “Comparison of methods of com- [9] X. Selberherr. 15. pp.” Proc. 712–724. pp. San Francisco.. R. vol. 129–134. 682–687. He has been with [28] Z. and the Ph. pp. D. 94. 1.D. E. Zeng.. Technol. 74. and F. Thompson. Sapatnekar is a recipient of seven conference Best Paper awards. 2014–2022.. void evolution in Cu interconnects. [34] A. [19] C. 20. C. 90530O-1–90530O-8. India. pp. L. Sapatnekar (S’86–M’93–F’03) received [25] K. Devices. Design Autom.. Mr. is currently pursuing the Ph. the M. 1222–1228. Comput. prefactor for surface electromigration and void drift in Cu interconnects. no. 2014. 376–381. ability of Cu/porous low k interconnects. Christiansen.Tech. USA. alloys for advanced chip interconnects. vol. Liniger. nos. Conf. 2015). K. “Electromigration failure modes in aluminum metallization tion in short copper interconnects. degrees from the Indian Institute of [20] C. 78. CA.. 061907-1–061907-3.3. pp. since 1997. Choi. 8. 553–560.-K. pp.. no. Black. 2013. S. no. Gan. IEEE Int.. Banerjee and A. Asia South gration checking for mesh power grids. Li.” in Proc. vol. 2002. pp. 4. 2007. pp. 5. Lloyd. L. Sci. 2015. 1995. degree with the Rel. from 2008 to 2011.. Fawaz. Nassif. IEEE/ACM Int. CA. Rel. Appl. “In situ study of void [41] J. Yang. Gignac.. “Global (interconnect) warming. IEEE Int. Jo and R. 3. 2003. VOL. 74–77. Kyoto. Conf.” in Proc. “Variation-aware electromigration damascene Cu interconnect trees. University of Minnesota.3. from 1992 to 1997. and S. thin-film lines. 47. Conf. 1203–1208. no.. Selberherr. V. Fawaz. CA. P. an ICCAD 10-year Retrospective Most Influential Paper award. Lee. International Technology Roadmap for Semiconductors. His current research improvement in Cu low-k interconnects. Rel. Design Autom. nitride. 1991. USA.” J.” in Proc. interests include circuit modeling and development Symp.. Jain. Sep. Seoul. in 2008. 57. “Amorphous Ru/polycrystalline Ru highly reliable of interconnect performance degradation due to electromigration and develop- stacked layer barrier technology. K. no. [26] A. CA. S.-C. L. 2013. San Jose. de Orio. pp.7. Chatterjee. Hau-Riege.” J. Minneapolis. 4. He was an ASIC Design Engineer with Taiwan 2007... vol. [29] R. vol. Hauschildt. Conf. Sweden. W. pp.” J. pp.” IEEE Trans. Cortadella. pp. “A vectorless framework for [38] S. pp. Interconnect ment of CAD techniques for timing and power analysis of integrated circuits. Hu et al. USA. M.” J. P. 3D. no. “A retargetable and accurate [15] C.. Appl. Burlingame.-Aided Design. 346–351. 69. T. Feng. “Statistical analysis of electromigration lifetimes and IL.” in Proc. Asia South Pac. vol. USA. pp. [43] C. MN. CA. Tu. N. San Jose.. “In-situ ultra-high vacuum studies of elec- [8] V. no. pp.. Comput. Appl. Phys. and S. pp. 7. Symp. 1–6. of efficient algorithms to analyze the circuit impact [23] S.” USA. microscope comparison studies on electromigration of Cu and Cu(Sn) “Circuit-level reliability requirements for Cu metallization. L. 1969. Hu. Texas Austin. nos. Appl. IEEE Microelectron. pp. degree from the Indian Institute of Circuits Devices Mag. 54. Comput. [14] S. USA. and S.3. 698–704. Japan. 1995. 2008. “The effect of stress migration on electromigration degree from Syracuse University. Yu. vol. Gan. “Physics-based puting correlated lognormal sum distributions and outages for digital electromigration assessment for power grid networks. [17] B. “A compact model for [30] L. pp. and J. APRIL 2017 [4] R.. D.” J.Tech. Appl.” IEEE the B. USA. Phys. CA.” in Proc. [40] P. H.” Appl. Univ. “Electromigration path in Cu methodology for logic-IP-internal electromigration assessment. Mumbai. Technology Bombay.-K.S.D. L. Stockholm. 571–576. . Dwyer. Krishnan. pp. R. pp. pp. [10] J. Blech.” in Proc. vol. Phys. interconnects.. pp. in dual damascene copper interconnects. San Jose. 17. Hu.. ACM/EDAC/IEEE Design Autom. “Black’s law revisited—Nucleation and growth in electromi. vol. Ames. Phys. vol. Mater. Badami.” in Proc. 2013Interconnect. impurity and metal cap effects on Cu electromigration. USA. 2945–2947. Li. [5] S. Wei.” Poster award. IEEE/ACM Int. a Best copper grain size statistics on the electromigration lifetime distribution. San Diego. C. Appl. IEEE Int. 2007. 4428–4437. “Fast static analysis of power challenges for advanced on-chip Cu interconnects. H. Manuf.” in Proc. Rel. [35] Y. and S. pp. Thompson. Anaheim..-Aided Design.. USA. 2010. no. “Experimental no. and J.-A. Conf. pp.3. Ceric. Chiba. 8. Monterey. 2014. “Electromigration multistress pattern technique for cop- early electromigration failures of copper dual-damascene interconnects. Heryanto et al. pp. Tipple.” [7] I.