Interactive Tutorial on

Fundamentals of Signal Integrity

for High-Speed/High-Density Design

Andreas Cangellaris, Jose Schutt-Aine

and Umberto Ravaioli
ECE Department, University of Illinois at Urbana-Champaign
cangella@uiuc.edu, jose@decwa.ece.uiuc.edu, ravaioli@uiuc.edu
Alina Deutsch
IBM Corporation, T.J. Watson Research Center
deutsch@us.ibm.com

DAC 2001 © SEMCHIP 1

Outline
• Introduction (A. Deutsch)
– The interconnect bottleneck in high-speed systems
• Interconnect Modeling Fundamentals
(A.Cangellaris/U. Ravaioli)
– Time-domain & frequency-domain transmission line analysis
– Lossy lines and signal dispersion
– Crosstalk for short lengths of coupled interconnects
• On-Chip Interconnects (A. Deutsch)
– Modeling of on-chip interconnects
– Interconnect impact on system performance
– Future trends

DAC 2001 © SEMCHIP 2

Outline (cont.)
• Interconnects at the Package and Board Level
(J.Schutt-Aine/U. Ravaioli)
– Multiconductor transmission line theory
– Crosstalk modeling and measurement
– Lumped vs. distributed modeling of interconnects
• Concluding Remarks

DAC 2001 © SEMCHIP 3

Fundamentals of
Transmission Line Theory

DAC 2001 © SEMCHIP 4

Transmission-line theory quantifies signal propagation
on a system of two parallel conductors with cross-
sectional dimensions much smaller than their length

Coaxial Stipline Microstrip Familiar

Cross sections

I(z,t)
Generic symbol for
V(z,t) a two-conductor line

z (direction along interconnect axis)

DAC 2001 © SEMCHIP 5

For a uniform transmission line, the electric and magnetic fields

are transverse to the direction of wave propagation (and hence, to
the axis of the line). Thus, transmission line fields are called
Transverse Electromagnetic (TEM) Waves
y Electric field lines

x The electric field behaves like an electrostatic field.

Over the cross section, the potential difference between
any two points A and B on the two conductors is constant:



A→B
∫ E( x, y, z, t ) ⋅ dl = V ( z, t )

Magnetic field lines

Over the cross section, the magnetic field looks like a
magnetostatic field. Its line integral around one of the
conductors equals the total current in the conductor.



∫H ( x, y , z , t ) ⋅ d l = I ( z , t )
center
conductor

DAC 2001 © SEMCHIP 6

 In a transmission line configuration as much charge
moves down the “active” wire that much charge of
negative polarity moves down the “return path”

Active wire Active wire

(+++++)

Return path (- - - - -) Return path

Direction along interconnect Direction along interconnect

At every cross section of a transmission line the
Electric field
currents on the active and return wires
between wires
balance each other.
This balance leads to field confinement and
reduced interference.
DAC 2001 © SEMCHIP 7

Signal propagation is quantified in terms of the

solution of the so-called Telegrapher’s equations

Time-Domain Form Frequency-Domain Form

∂v ( z , t ) ∂i ( z , t ) dV ( z , ω )
= − Ri ( z , t ) − L = − RI ( z , ω ) − jω LI ( z , ω )
∂z ∂t dz
∂i ( z , t ) ∂v ( z , t ) dI ( z , ω )
= − Gv ( z , t ) − C = − GV ( z , ω ) − jω CV ( z , ω )
∂z ∂t dz

I(z,t) I(z,t) R(∆z) L(∆z)

I(z+∆z,t)

V(z,t) C(∆z) G(∆z)

V(z,t) V(z+∆z,t)

z ∆z
DAC 2001 8
 Transmission-Line Parameters
• Per-unit-length capacitance C
• Per-unit-length conductance G
• Per-unit-length inductance L
– Loop inductance
– Frequency dependence due to Current Loop
skin effect
• Per-unit-length resistance R
∆z
– Strong frequency dependence
due to skin effect

DAC 2001 © SEMCHIP 9

Time-Domain Solution of Telegrapher’s Equations

Neglecting losses for simplicity:
∂v ( z , t ) ∂i ( z , t ) 
= −L
∂z ∂t  ∂ 2v 1 ∂ 2v
 ⇒ − =0
∂i ( z , t ) ∂v ( z , t )  ∂z 2 v 2p ∂t 2
= −C
∂z ∂t 
where v p = 1 is the wave velocity on the line.
LC
General solution: v ( z , t ) = f + ( z − v p t ) + f − ( z + v p t )
 
forward wave backward wave

1 + 1 −
Current wave: i ( z , t ) = f ( z − v pt ) − f ( z + v pt )
Z0 Z0
   
forward wave backward wave

L
where Z 0 = is the characteristic impedance of the line.
C

DAC 2001 © SEMCHIP 10

A voltage signal f (t) launched on a lossless line
propagates unaltered with speed vp dependent on the
transmission-line properties.
I+(z0,t)
1
Vs V+(z0,t) vp =
LC
z V + ( z, t ) L
 Z =
I + ( z, t )
0
V+(z0,t) C
Vs(t) = f (t)
Characteristic
t Impedance
I+(z0,t)
t
Time of
flight t0 = z0/vp t

DAC 2001 11

The characteristic impedance dictates the

amplitude of the voltage waveform launched
on the line
+
RS I (0,t)

Vs V+(0,t) To infinity (or matched)

VS (t ) = V + (0, t ) + I + (0, t ) RS  + Z0
 ⇒ V (0, t ) = V ( t )
Z 0 + RS
S
V + (0, t ) = Z 0 I + (0, t ) 

DAC 2001 © SEMCHIP 12

Discontinuities in the characteristic impedance
of a transmission line give rise to reflections
At the junction it is:
Z01 Z02 V1 = V + + V − = V2 = V ++
V ++
(V +, I +)
(V −, I −)
(V ++, I ++) I1 =
1
Z 01
(
+ −
V −V = I2 = )
Z 02

Z02 − Z01 
V − = ΓV +  V ++ = TV +  Reflection Coefficient: Γ=
  Z02 + Z01 
− V −  and ++ V ++  
I =−  I = 2Z02 
Z01  Z02  Transmission Coefficient: T =
Z02 + Z01 
Maintaining a fairly constant value of the characteristic impedance
along an interconnect path is essential for reflection suppression.

DAC 2001 © SEMCHIP 13

Source and load impedances impact transmission line

performance of the interconnect

Vs ZS(f) Z0 ZL(f)

Source reflection coefficient Load reflection coefficient:

Z S ( f ) − Z0 ZL ( f ) − Z0
ΓS ( f ) = ΓL ( f ) =
Z S ( f ) + Z0 Z L ( f ) + Z0
Load transmission coefficient:
2Z L ( f )
TL ( f ) = 1 + Γ L ( f ) =
Z L ( f ) + Z0

DAC 2001 © SEMCHIP 14

 Example: Unterminated interconnect (ZL=∞) driven by high
source impedance driver with ZS>>Z0 (e.g. unbuffered CMOS)
Source (ΓS ≈ 1) Load (ΓL = 1)
Excitation: Step Pulse of amplitude V0
V+ Z0 Z − Z0
V + = V0 << V0 , Γ S = S ≈1
T=d / v Z S + Z0 ZS + Z0
ΓLV+
(One-way
Z L − Z0
Delay) ΓL = = 1, TL = 2
Z L + Z0
2T
ΓSΓLV+ VL(t ) Steady-state value V0

3T Slow response
ΓL ΓSΓL V+
~6V+
4T ~4V+
t Bounce diagram t 2V+

T 3T 5T t
DAC 2001 15

Example: Unterminated interconnect (ZL=∞) driven by low

source impedance driver with ZS<Z0 (e.g. ECL or strong TTL)
Source (ΓS ≈ –1) Load (ΓL = 1)
Excitation: Step Pulse of amplitude V0 ; Z0 = 7ZS
Z0 7 Z −Z
V+ V + = V0 ≈ V0 , ΓS = S 0 = −0.75
ZS + Z0 8 ZS + Z0
T=d / v
ZL − Z0
ΓLV+
(One-way
ΓL = = 1, TL = 2
Delay) ZL + Z0

2T VL(t ) Overshoot & Ringing

ΓSΓLV+ = - 0.75 V+ 2V+
1.625V+
3T
ΓL ΓSΓLV+ LV+ = -0.75 V+ Steady-
0.5V+ State: V0
4T
t Bounce diagram t T 3T 5T t

DAC 2001 © SEMCHIP 16

 A capacitor CL represents the load at the gate
input of the receiver. Its presence adds delay.
(V+, I+) VL = V + + V −

(V-, I-) C
IL =
1
Z0
(V + −V − )
L
dVL
Interconnect delay = T I L = CL , VL (t = T ) = 0.
dt
VL (t ) = V + (1 − exp( − (t − T ) / τ ) ) , t > T
where τ =Z 0 C .
Let Td be the time at which VL (t = Td ) = 0.9V + .
Td = T + 2.3τ ⇒
Extra delay due to the capacitor is 2.3Z 0 C

DAC 2001 © SEMCHIP 17

Delay is introduced by all capacitive and inductive

discontinuities present in a signal path

Wire bonds
(primarily inductive) Interconnect bends
Via
(primarily capacitive)
(primarily inductive)

Z0 Z0 Z0 Z0

Equivalent circuit for SPICE-based transient simulation

DAC 2001 © SEMCHIP 18

 The delay due to a capacitive o an inductive
discontinuity depends on the values C or L and Z0
Capacitive Discontinuity Inductive Discontinuity
(V++, I++) (V+, I+) (V++, I++)
(V+, I+)

C Z0 Z0 L Z0
Z0

(V-, I-) (V-, I-)

V ++ (t ) = V + (1 − exp( −t / τ C ) ) V ++ (t ) = V + (1 − exp( −t / τ L ) )
CZ 0 L
where τ C = where τ L =
2 2Z0
V (t ) reaches 0.9V + at t = 2.3τ C
++
V ++ (t ) reaches 0.9V + at t = 2.3τ L
Hence, the capacitor adds a delay of Hence, the inductor adds a delay of
Td = 1.15CZ 0 . Td = 1.15 L / Z 0

(C=1 pF, Z 0 = 50 Ohm; Td = 57.5 ps) ( L = 2.5 nH, Z 0 = 50 Ohm; Td = 57.5 ps)

DAC 2001 19

Slots in ground planes increase interconnect delay and

enhance noise generation and interference

The slot in the ground plane

acts as a slot antenna.

interconnect
The return current in the ground plane
flows around the slot. Hence,
• Extra L ⇒ extra delay
• Unbalanced currents lead to
enhanced emissions Ground plane
• Interference (crosstalk) with
other wires beyond immediate Interference occurs on
neighbors either side of the plane

DAC 2001 © SEMCHIP 20

 Transmission line models of interconnects
predict only “differential-mode” currents
Displacement
I1 current (radiation)

- Conduction (return)
current
I2 I1 − I 2
Transmission-Line
Model ID
Differential-mode current: I D =
2
I +I
+
- ID
Common-mode current: I C = 1 2
-
2
I1 = I C + I D
Radiated emissions calculation
can be grossly incorrect if the I 2 = IC − I D
common-mode current is not
taken into account
DAC 2001 © SEMCHIP 21

The input impedance of a match-terminated interconnect with a

continuous return path remains essentially contact over a broad
frequency range




'(





 
 


)(



  
 !"#$%&
Plots generated using UIUC’s fast time-domain solvers (Prof. E. Michielssen)

DAC 2001 © SEMCHIP 22

The disruption of the return path caused by the slot manifests itself
as an added inductance at lower frequencies and radiated emissions
(radiation resistance) at higher frequencies



*


 '(

*
)(


*

 +,)


 +,",)


  
 !"#$%&

Plots generated using UIUC’s fast time-domain EM solvers (Prof. E. Michielssen)

DAC 2001 © SEMCHIP 23

Mesh (Grid) Planes in PCBs increase the

characteristic impedance of the lines

Return current
Per-unit-length inductance, L, increases.
Per-unit-length capacitance, C , decreases.
L
Z0 = increases
C

DAC 2001 © SEMCHIP 24

In the case of mesh (grid) planes, high-speed lines
should be routed right above the plane metallization
interconnects

Grid Plane
(cross section)

Advantages
-Better impedance control
-Reduced cross-plane interference

DAC 2001 © SEMCHIP 25

Lossy Transmission Lines

• Ohmic loss in the metallization

– Frequency-dependent R and L (skin effect)
• Insulating substrate loss
– Frequency-dependent G
• Semiconductor substrates
– Frequency-dependent R and L
– Frequency-dependent G and C

DAC 2001 © SEMCHIP 26

Frequency-Domain Solution of Telegrapher’s Equations
In the frequency domain, interconnect loss can be accounted for easily.
dV ( z , ω ) 
− = [ R (ω ) + jω L (ω )] I ( z , ω ) 
dz  d 2V ( z , ω )
 ⇒ − Z (ω )Y (ω )V ( z , ω ) = 0
dI ( z , ω ) dz 2
− = [G (ω ) + jω C (ω )]V ( z , ω ) 
dz 
where Z (ω ) = R (ω ) + jω L (ω ) is the per-unit-length impedance of the line
and Y (ω ) = G (ω ) + jω C (ω ) is the per-unit-length admittance of the line.

V ( z , ω ) = V + (ω ) exp( −γ z ) + V − (ω ) exp(γ z )

General solution :  1
 I ( z , ω ) = Z (ω ) V (ω ) exp( −γ z ) − V (ω ) exp(γ z ) 
+ −

 0

γ (ω ) = [ R (ω ) + jω L (ω )][G (ω ) + jω C (ω )] is the complex propagation constant ,

R (ω ) + jω L (ω )
and Z 0 (ω ) = is the characteristic impedance.
G (ω ) + jω C (ω )

DAC 2001 © SEMCHIP 27

The characteristic impedance of a lossy line

is a complex number!
When G ≈ 0 it is the per-unit-length ohmic loss in the wires that
dominates the loss; hence,
R (ω ) + jL (ω ) L (ω ) R (ω )
Z 0 (ω ) = = 1− j
jω C C ω L (ω )
For the interconnect structures of interest , L is in the order of nH/cm;
hence, for f < a few tens of MHz, ω L << R (especially for thin-film wire).
1 − j R (ω )
Thus, for low frequenc ies: Z 0 (ω ) ≈
2 ωC
Notice that the real and imaginary parts are of the same magnitude.
On the other hand , for high frequ e ncie s such that R << ω L ,
L  1 R (ω ) 
Z 0 (ω ) = 1 − j 
C  2 ω L (ω ) 
Notice that, since at high frequencies R (ω ) ∝ ω , the characteristic
impedance is predominantly real.
DAC 2001 © SEMCHIP 28
 The presence of loss is responsible for signal attenuation
and distortion
The propagation constant becomes frequency dependent:
γ (ω )= [ R (ω ) + jω L (ω )][G (ω ) + jω C ] = α (ω ) + j β (ω ).
α (ω ) is the attenuation constant; β (ω ) is the phase constant.
V + (ω , z ) = V0+ exp( −α (ω ) z ) exp( − j β (ω ) z )
    
attenuation phase shift

The characteristic impedance and the phase velocity are frequency dependent:
R (ω ) + jω L(ω ) ω
Z 0 (ω ) = , v p (ω ) =
G (ω ) + jω C β (ω )

Different frequencies in the spectrum of a pulse propagate at

different speeds and suffer different attenuation.
This results in pulse distortion often referred to as dispersion

Input Output
Lossy line
Pulse Pulse

DAC 2001 29

Input Impedance of a Transmission Line

(or, how long wires “transform” load impedances)
V (d ) Z + Z 0 tanh γ d
Z in ( d ) = = Z0 L
I (d ) Z 0 + Z L tanh γ d I(d)
Neglecting losses, γ = j β , Z 0 is real, and it is:
Z L + jZ 0 tan β d Z0,γ ZL
Z in ( d ) = Z 0 V(d)
Z 0 + jZ L tan β d
− Periodic function with period λ / 2 d
Z L − Z0 Z L − Z0
1+ 1−
Z L + Z0 Z L + Z0
− max( Z in ) = Z 0 ; min( Z in ) = Z 0
Z − Z0 Z L − Z0
1− L 1+
Z L + Z0 Z L + Z0
• Matched Load: Z L = Z 0 ⇒ Z in ( d ) = Z 0
• Shorted Line: Z L = 0 ⇒ Z in ( d ) = jZ 0 tan β d
- A shorted line of length equal to an odd multiple of λ / 4 has infinite input
impedance and th us appears as an open circuit.
DAC 2001 © SEMCHIP 30
 Grounding wires running some distance along a
ground plane or chassis exhibit transmission-line
behavior at RF frequencies.
d

Zg Z0

• Safety earth is not an RF ground.

• At high frequencies, the claim that
“everything is connected to earth”
through safety earth is
meaningless

• At high frequencies, “single point

ground” is meaningless

DAC 2001 © SEMCHIP 31

Skin-Effect Resistance
Skin effect

At frequencies such that the skin depth is At high frequencies, where the skin
larger or comparable with the conductor depth is smaller than the conductor
thickness, the current distributes uniformly thickness, current crowding around
over the conductor cross section. the perimeter occurs.

1
Skin depth: δ =
π f µσ
• At f = 1 GHz, for aluminum with conductivity σ = 4×107 S/m and
permeability µ = 4π × 10-7 H/m, the skin depth is 2.5 µm.
• For high enough frequencies, the p.u.l. resistance increases as √f

DAC 2001 © SEMCHIP 32

 The contribution of the return path to interconnect
resistance may need to be taken into account

R per-unit-length (Ohms/cm)
Frequency dependence
of the p.u.l. resistance 3
PEC reference planes
(top) and inductance Copper reference plane
2
(bottom) of the single
stripline configuration 1
with w=50 µm, t=10 µm,
0
g=10 µm, and h=100 µm. 10
-2 -1
10
0
10
1
10
Frequency (GHz)

L per-unit-length (Ohms/cm)
 3.4


  3.2


3

2.8
-2 -1 0 1
10 10 10 10
Frequency (GHz)

DAC 2001 © SEMCHIP 33

Extraction of the frequency-dependent p.u.l. interconnect

resistance must take into account the presence of
adjacent conductors (proximity effect)

DAC 2001 © SEMCHIP 34

 The per-unit-length resistance matrix has non-zero off-diagonal
elements. Taking these off-diagonal elements into account is
important, especially for the tightly coupled wires
5

4.5

Per-unit-length Resistance (Ohms/cm)

4
g
3.5
w
3

1 t 2.5

2
s h
1.5
R11 R12
2 1

0.5

g 0
-4 -3 -2 -1 0 1
10 10 10 10 10 10
Frequency (GHz)
3.5

w= 50 µm, t=10 µm, s= 30 µm 3

L11
h= 110 µm, g= 10 µm, εr=4

Per-unit-length Inductance (nH/cm)

2.5

Aluminum: σ=3.3E7 S/m 2

Copper: σ =5.8E7 S/m 1.5

L12
1

0.5
DAC 2001 10
-4
10
-3
10
-2
10
-1
10
0 3510 1

Frequency (GHz)

Impact of different metallization on delay

Vs length=20 cm g
– + w
Zo 1 t
1 pF
s h
Zo 2
1 pF g
Risetime=50 ps
Far-end Voltage of Driven line Far-end Voltage of Victim line
0.2

1
90% 0.15

0.8

0.1
Voltage (V)
Voltage (V)

0.6

0.05
0.4

~ 1 ns 0
0.2

0 -0.05
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (sec) -8 Time (sec) -8
x 10 x 10

DAC 2001 © SEMCHIP 36

 Impact of frequency-dependent loss on
interconnect transient response




 
  


Cross-section of a stripline
geometry.
s=50 µm, w=50 µm,
t=10 µm, g=10 µm,
h=200 µm and εr=4.
Copper metallization.

Constant Model (Rdc, L
, C)

Frequency dependent model: PEC reference planes (R(f), L(f), C)
Frequency dependent model: Copper reference plane (R(f), L(f), C)
DAC 2001 37

The effect of frequency-dependent loss is

particularly apparent in the cross-talk levels




 
  


Constant Model (Rdc, L
, C)

Frequency dependent model: PEC reference planes (R(f), L(f), C)
Frequency dependent model: Copper reference plane (R(f), L(f), C)
DAC 2001 38
 Insulating Substrate Loss
• Characterized in terms of the substrate material
conductivity or loss tangent
 ε ′′(ω ) 
ε (ω ) = ε ′(ω ) − jε ′′(ω ) = ε ′(ω ) 1 − j  = ε ′(ω ) (1 − j tan δ (ω ) )
 ε ′(ω ) 
  σ (ω )  
σ (ω ) + jωε ′(ω ) = jω ε ′(ω ) 1 − j 
  ωε ′(ω )  
ε ′′(ω ) σ (ω )
tan δ (ω ) = =
ε ′(ω ) ωε ′(ω )
• Transverse electric field between conductors results in a
transverse leakage current and, thus, ohmic loss


ε,σ
E ε J G C

DAC 2001 © SEMCHIP 39

The assumption of constant loss tangent leads to

physically inconsistent models for G

Coaxial 2πε 2πσ

C= , G= ⇒
ln(b / a ) ln(b / a)
2a
G σ G
= ⇒ = ω tan δ
C ε C
2b
• Assuming tanδ is constant yields G(ω) ∝ ω
– Such a behavior violates causality!
– For a causal circuit
Re{Y (ω )} is an even function of frequency Y(ω) G C

Im{Y (ω )} is an odd function of frequency

DAC 2001 © SEMCHIP 40

Simply assuming the loss tangent to remain constant
over a broad (multi-GHz) frequency range leads to
a non-physical behavior of G(ω)
• A physically correct model needs to start with a physically-
correct description of the frequency dependence of the complex
permittivity.
– Use measured data for the complex permittivity to synthesize a Debye
model for it
– Use the synthesized Debye model for the extraction of C(ω) and G(ω)
K
εk
ε (ω ) = ε ′(ω ) − jε ′′(ω ) = ε ∞ + ∑ ⇒
k =1 1 + jω τ k
K
ω ε kτ k
ε ′′(ω )
∑ 1+ ω τ
k =1
2 2
tan δ (ω ) = = k
ε ′(ω ) K
εk
ε∞ + ∑
k =1 1 + ω τ k
2 2

G (ω ) ∝ ω tan δ ⇒ even function of frequency

DAC 2001 41

Capacitive and Inductive Crosstalk

in Short Interconnects
V1 V2 V3 VN Reference
y conductor
x
V1   R11 R12  R1N  I1   L11 L12  L1N  I1 
    R2N    L22  L2N  
d V2   R21 R22   I2  + jω  L21   I2 
−  =
dz   
  
 

VN  RN1 RN 2     
 RNN 
 IN  LN1 LN 2  LNN 
 IN 
 I1   G11 G12  G1N V1   C11 C12  C1N V1 
    G2N    C22  C2N  
d  I2  G21 G22  V2  + jω C21  V2 
−  =
dz  
  
 

IN  GN1 GN2   
 GNN VN  CN1 CN2  CNN VN 

DAC 2001 42
 Crosstalk in Coupled Lines
• For interconnects with more than two (active) conductors,
crosstalk analysis is most effectively performed in terms of a
circuit simulator that can support MTL models (*).
– Most common (and computationally efficient) SPICE
equivalent circuits for MTL assume lossless transmission
lines.
– Models for MTLs with losses (including frequency-dependent
losses associated with skin effect) are available also. They are
essential for accurate analysis of interconnect-induced delay,
dispersion, and crosstalk at the board level for signals of GHz
bandwidths.
– It is assumed that the interconnect structure is uniform
enough for its description in terms of per-unit-length L,C,R,
and G matrices to make sense.
(*) V.K. Tripathi and J.B. Rettig, “A SPICE Model for Multiple Coupled Microstrips and Other
Transmission Lines,” IEEE Trans. Microwave Theory Tech., vol. 33(12), pp. 1513-1518,
Dec. 1985.

DAC 2001 © SEMCHIP 43

For the case of a three-conductor, lossless line in

homogeneous dielectric, with resistive terminations,
an exact solution is possible.
• Exact solutions are useful because:
– they help provide insight into the crosstalk mechanism;
– they can be used to validate computer-based simulations.
• The following results were first published by C.R. Paul (C.R.
Paul, “Solution of transmission line equations for three-
conductor lines in homogeneous media,” IEEE Trans. On
Electromagnetic Compatibility, vol. 20, pp. 216-222, 1978.
Rs IG (0) l IG (d) RL

RNE VG (0) VG (d ) RFE

Vs ~
IR (d)
IR (0) VNE VFE

DAC 2001 © SEMCHIP 44

Exact solution for crosstalk in a lossless, three-
conductor line with resistive terminations
L M   CG −CM 
Per unit length parameters: L =  G  = −C CR 
, C
 M LR   M
Near - end and far - end crosstalk voltages :
S  RNE  j 2π l / λ  
VNE =  jω Ml  C + α LG S  IGDC  +
D  RNE + RFE  1− k 2  
S  RNE RFE  j 2π l / λ 1  
 jω CM l  C + S VGDC 
D  RNE + RFE  1 − k 2 α LG   Inductive &
Capacitive
S RFE R R 
VFE = − jω MlI GDC + NE FE jω CM lVGDC  Coupling
D  RNE + RFE RNE + RFE  Coefficients

sin β l M CM
where, C = cos β l , S = , k= = ≤ 1, and
βl LG LR CG CR
DAC 2001 45

Exact solution for crosstalk in a lossless, three-

conductor line with resistive terminations
 (1 − αSGαLR )(1 − αLGαSR )  + ω τ + τ ;
D = C 2 − S 2ω 2τ Gτ R 1 − k 2  j CS ( G R )
 (1 + αSRαLR )(1 + αSGαLG ) 
R R R R
αSG = S , α LG = L , αSR = NE , α LR = FE ;
ZCG ZCG ZCR ZCR
ZCG = LG / CG , ZCR = LR / CR are the characteristic impedances of each
line in the presence of the other one;
LGl RR LRl R R
τG = + CGl S L , τ R = + CRl NE FE ,
RS + RL RS + RL RNE + RFE RNE + RFE
are the time constants of the coupled lines;
RL VS
VGDC = VS , IGDC = , are the voltage and current of the
RS + RL RS + RL
generator circuit under dc excitation (no coupling to the receptor circuit).
DAC 2001 46
 Under the assumptions of electrically short lines, and weak
coupling, the crosstalk equations simplify considerably
• A line is said to be electrically short if its length is a small fraction of
the wavelength at the highest frequency of interest.
Package interconnects fall in this category
• Two lines are said to be weakly coupled if the coupling coefficient,
k , is sufficiently smaller than 1.

Under these assumptions the equations for the near-end and

far-end crosstalk voltages become:
1 RNE RNE RFE RL 
VNE =  jω Ml + jω CM l 
D  ( RNE + RFE )( RS + RL ) ( RNE + RFE )( RS + RL ) 
1 RFE RNE RFE RL 
VFE = − jω Ml + jω CM l 
D  ( RNE + RFE )( RS + RL ) ( RNE + RFE )( RS + RL ) 

DAC 2001 © SEMCHIP 47

For weakly coupled, electrically short wires, the crosstalk is a

linear combination of contributions due to the mutual
inductance between the lines (inductive coupling) and the
mutual capacitance between the lines (capacitive coupling).

jω RNE
VNE = VNE
IND
+ VNE
CAP
= ( Ml + (RFE RL )CM l )
D ( RNE + RFE )( RS + RL )
jω RNE
VFE = VFINE D + VFE
CAP
= ( −Ml + (RNE RL )CM l )
D ( RNE + RFE )( RS + RL )
Notice that:
• The higher the frequency the larger the crosstalk
• Inductive coupling dominates for low-impedance loads
• Capacitive coupling dominates for high-impedance loads

DAC 2001 © SEMCHIP 48

 Lumped versus Distributed Modeling

• When the interconnect length is much smaller than

the wavelength of interest, lumped models provide
sufficient accuracy and can be used
– Typical case for package interconnects at RF
frequencies
– Inaccurate for interconnects at the MCM and PCB level
• What does “interconnect length is much smaller
than the wavelength of interest” really mean?
– Typical rule of thumb: d < λ/10
– …but one can take a closer look at this rule of thumb as
shown next

DAC 2001 © SEMCHIP 49

There are four possible implementations of lumped

models for a two-conductor interconnect of length d

L/2 L/2 L L = Lp.u .l .d

C = C p.u .l .d
C C

T- model Right L - model

L L

C C/2 C/2

Left L - model π - model

DAC 2001 © SEMCHIP 50

The accuracy of a lumped model can be examined by considering
the input impedance obtained when the model is terminated at the
characteristic impedance of the line
(see B. Young, Digital Signal Integrity, Prentice Hall, 2000) L/2 L/2

For the model to exhibit "transmission line" behavior, Zin C Z0

its input impedance should equal the load impedance Zˆ0 :
2
1 L ω  2
Zin = jω L / 2 + = Zˆ0 ⇒ Zˆ0 = 1 −   , where ωT =
jωC +
1 C  ωT  LC
Z0 + jω L / 2
ˆ
• For ω <<ωT , Zˆ0 ≈ Z0
• A bandwidth of validity of the T-model can be obtained by finding ωmax such that
Z − Zˆ ≤ aZ for ω ≤ ω .
0 0 0 max

0.62 0.62
For a = 0.05, ωmax = = .
LC d Lp.u.l C p.u.l .
Application: d = 2 cm, Lp.u.l . = 4 nH/cm, C p.u.l = 1 pF/cm; ωmax = 4.9 GHz

DAC 2001 © SEMCHIP 51

5% accuracy bandwidth for the four lumped models

T- model π - model
L/2 L/2 −1/2 L
L   ω LC  
2 2
L  ω LC 
Zˆ0 = 1−   Z0 =
ˆ 1− 
C C  2  C   2   C/2 C/2
 
0.62 0.62
ωmax = ωmax =
LC LC

Right L - model Left L - model

2 2
L L  ω LC  L  ω LC  L
Zˆ0 = jω L / 2 + 1 −   Zˆ0 =− jωL /2 + 1− 
C  2  C  2  C
C 0.1 0.1
ωmax = ωmax =
LC LC

DAC 2001 © SEMCHIP 52

The number of segments is dictated by the
maximum frequency of interest that must be
represented accurately in the simulation
• The effective bandwidth criteria described earlier can be used to
select the segment size.
Example : Interconnect lead with total capacitance
C = 20 pF, and (loop) inductance L = 50 nH.
Let f max = 10 GHz be the maximum frequency of interest.
Find the minimum number n of lumped segments required
for accurate modeling.
Lseg = L / n and C seg = C / n.
0.62 0.62 n 2π f max LC d
2π f max ≤ = ⇒n≥ or n ≥ 10
Lseg C seg LC 0.62 λ min
For the given numbers, nmin = 101

DAC 2001 © SEMCHIP 53

Use of an insufficient number of segments leads to

artificial filtering and phase distortion of the
transmission-line response

The load response for

a source and load
match-terminated, lossless
transmission line is:
 2π  
VL = 0.5exp  − j 
 λ 

DAC 2001 © SEMCHIP 54

If done properly, distributed RLCG circuit
modeling of MTLs works

Such an approach is preferable when:

• the wire resistance must be taken into account; (as
already mentioned, some SPICE vendors provide lossy
line modeling through extensions of the exact model
mentioned earlier);
• the line is electrically short, and a few lumped-circuit
segments are sufficient for its modeling;
• the MTL exhibits non-uniformity (variable cross section)
along its axis;
• modeling of radiation coupling to the MTL is desired.

DAC 2001 © SEMCHIP 55