Capacitive Coupling Noise in High-Speed

VLSI Circuits

Payam Heydari
Department of Electrical and Computer Engineering
University of California
Irvine, CA 92697

Massoud Pedram
Department of Electrical Engineering-Systems
University of Southern California
Los Angeles, CA 90089

1
1 Introduction

Shrinkage of the minimum feature size of the semiconductor devices to 130 nm and below
and increase in the the clock frequency to 3 GHz and above have caused crosstalk noise to
become a serious problem in integrated circuits. More precisely, crosstalk noise has
evolved as the key source of performance degradation and signal integrity problems in
high speed VLSI designs.

Various techniques have been proposed to evaluate the crosstalk noise in integrated cir-
cuits. The most accurate approach is to use a transistor-level circuit simulator. This
approach is, however, computationally inefficient, and hence, is not applicable to large
circuit structures. For example, our experiments show that simulating a small circuit struc-
ture consisting of a collection of 10 coupled lossy transmission lines with HSPICE takes
almost three minutes on a 1.5GHz Intel Pentium-4-based computer system. Since inter-
connects are modeled as linear time-invariant systems, model reduction techniques [1]-[6]
can be utilized to reduce the computational complexity. These model order reduction tech-
niques may be incorporated into the noise analysis and calculation programs to accurately
determine the noise behavior of the circuit under study. For example, reference [7] enu-
merates different types of environmental noise sources that have a major impact on digital
VLSI circuit performance. Next it proposes a fast methodology using noise graphs to ana-
lyze the noise. The main shortcoming of this work is that it does not accurately model the
on-chip interconnects. Reference [8] incorporates a model order reduction technique to
efficiently simulate the on-chip interconnects as distributed RC sections. However, it does
not present any analytical expression for the crosstalk noise. Moreover, in spite of employ-
ing model reduction techniques, this approach cannot completely solve the problem of
long computation times associated with this kind of noise analysis. In addition, neither [7]
nor [8] provide any insight to the circuit designers as to how to modify the circuit struc-
tures in order to reduce or control the crosstalk noise.

It is desirable to use closed-form expressions, instead of simulation tools, to predict the

noise effects in a circuit as long as their prediction accuracy is acceptable. This is espe-
cially true during the early stages of the design process when one does not afford simulat-

2
ing a large number of possible circuit structures and layout solutions. Consequently, a
number of researchers have addressed the problem of obtaining simple, closed-form
expressions for crosstalk noise in VLSI circuits. Vittal et al. in [9] provide bounds for the
crosstalk noise using a lumped RC model. This work, however, ignores the interconnect
resistance. Later, the same authors, in [10], make use of geometric considerations to obtain
expressions for the peak amplitude and the pulse width of the noise. Knowing the noise
pulse width is important because, in general, the noise margin of a gate depends on both
the noise peak amplitude and the noise pulse width. Their technique can handle arbitrary
input signals. In [11], Devgan proposes a clever technique for finding an upper bound on
the crosstalk noise. The author himself mentions that his model exhibits a large error when
the signals are fast and the rise and fall times are short. Unfortunately, this latter scenario
occurs frequently when practical values of the interconnect parasitics and signal frequen-
cies are used. We have observed that the percentage of the estimated error in such cases
can be as much as 60%. In addition, [11] does not predict the noise pulse width.
Kuhlmann et al. in [12] and [14] propose an exact crosstalk noise estimation method
for a distributed RC(L) model of the VLSI interconnect. They employ a moment-match-
ing technique and Devgan’s metric to effectively reduce the Laplace transform of the cou-
pled interconnect to a low-order rational function. More specifically, for a distributed RC
interconnect, a high-order transfer function is first reduced to a third order rational transfer
function by using a combination of moment-matching technique and Devgan’s metric.
The poles of the reduced-order transfer function are then derived to examine the stability
of the reduced system. If the reduced system is unstable, a second-order transfer function
will be used instead. Authors demonstrate the accuracy of their proposed method by com-
paring its performance with that of HSPICE. The metric proposed in [12] and [14] is less
intuitive in terms of the circuit interpretation, and requires two tree traversals (one for the
aggressor net and the other one for the victim net), a diagonal matrix-vector multiplication
for the DC component of the Taylor series expansion of the victim voltage in the Laplace-
domain, two additional tree traversals per moment, and two diagonal matrix-vector multi-
plication for high moments of the victim and the aggressor voltages. The tree traversal and
the multiplication are of the order of O(n) for n nodes. In addition, for every run of the cir-

3
cuit simulator on the system, roots of the reduced third-order model should be obtained for
the stability check. This operation is also of the order of O(n). The number of tree travers-
als and matrix multiplications to generate the third-order model are six times and twenty
times more than those in the Devgan’s metric, respectively [14]. [13] proposed a conven-
tional lumped 2-p RC circuit for the victim net, while postulating that the slew rate at the
coupling location is obtained by simple slew-rate calculation from the aggressor driver,
which is not accurately correct for long parallel wires. The reason is that the slew-rate cal-
culation for the aggressor net must account for the distributed nature of the coupling
between the adjacent lines. Although a lumped 2-π model can help us derive closed-form
analytical models for the noise attributes, the results lead to large and unaccetable errors
for long aggressor and victim nets. Using this model, the authors calculated the noise peak
and pulse width, and defined the amplitude pulse width product. Several experiments in
this paper have shown that the noise peak amplitude will have a more contribution to the
circuit failure than the noise pulse width. Therefore, a new definition is needed to empha-
size the bigger impact of the noise peak amplitude. Takahashi et al. in [15] proposes a 2-π
equivalent circuit to estimate crosstalk noise of partially coupled RC trees. This paper
assumes the aggressor waveform to be an exponential function, which then yields a more
accurate estimate than that case that either a step input or a saturated ramp input are
assumed. Unfortunately, the proposed analytical model for the interconnect is a 2-π RC
network, which cannot capture the distributed nature of a long RC interconnect. This is the
major shortcoming of the work by Takahashi et al. [16] has proposed a fast aggressor and
tree reductions to estimate the crosstalk. Similar to [13] and [15], [16] employs double-
pole approach for the crosstalk noise estimation.
In this paper, a new crosstalk noise metric is proposed, which is capable of predicting
the noise amplitude and the noise pulse width of an RC interconnect as well as an over-
damped RLC interconnect very efficiently. This paper is based on the work originally pro-
posed in [17]. Several experiments reveal that the proposed metric, on average, predicts
the peak crosstalk noise with the same or higher level of accuracy compared to the second-
order reduced model proposed in [14]. Unlike [14], the proposed noise metric gives a
rather simple analytical model of the crosstalk noise, which is efficient and sufficiently

4
accurate to be effectively incorporated in state-of-the-art noise calculators. The proposed
noise metric has a closed form expression that clearly highlights the dependency of the
noise on the aggressor and victim line circuit parameters as well as on the input signal rise/
fall times.
This paper is organized as follows. In Section 2, the interconnect coupling phenomenon
is reviewed and, through experimental results, it is shown that the inductive coupling on
chip is negligible for local wiring clocked at a target frequency of 1GHz. This result justi-
fies the focus of this paper, which is on the capacitive crosstalk effect. After a brief
description of Devgan’s metric in Section 2.1, we introduce our new noise metric in Sec-
tion 2.2. We next compare our metric with the analytical models proposed by Vittal [10],
Devgan [11], and Kuhlmann [12], [14] through a series of detailed simulation experi-
ments. Finally, Section 3 provides the conclusion of this paper.

2 Capacitive Coupling
The electromagnetic coupling of a signal from one conductor to another, which is called
crosstalk, can be induced through two coupling mechanisms: capacitive and inductive.

All signal conductors exhibit some interwire capacitances among themselves. When
the conductors are placed sufficiently close to each other, the capacitance becomes large
enough to couple significant energy from one conductor, called an aggressor or active
line, to another conductor, called a victim or passive line. Because, with each new process
technology, the thickness (height) of the wires is not scaled down as aggressively as the
width of the wires, and because the wires are packed increasingly closer to each other, the
ratio of the coupling capacitance to the total capacitance (includes area and fringe capaci-
tances) increases, and therefore, the capacitive coupling noise increases. Fig. 1 depicts a
highly simplified analysis (neglecting the resistive loss and the magnetic coupling of the
interconnect lines) of the essential attributes of the crosstalk noise. In this figure, coupled
noise on line 2 results from a transition on line 1. In this simplified circuit model, the
crosstalk voltage is obtained from a capacitive voltage division relationship as follows:
Cc
Vxtalk ( t ) =  ------------------------- VDD e–t ⁄ τ u ( t ) where τ = rDS, 12 ( Cc + C tot2 )
Cc + Ctot2

5
where u(t) represents the unit-step function, and rDS,12 is the on resistance of the NMOS
device of the victim line driver.
VDD VDD

MP11 MP21
Vin
MN11 Cd1 Cline1 Cg1 MN21

VDD Cc VDD
Ctot2=Cd2+Cline2+Cg2
MP12 MP22
VDD

MN12 Cd2 Cline2 Cg2 MN22

Fig. 1. A simplified circuit model of two capacitively coupled transmission lines.

High-speed digital circuits often employ dynamic logic families (e.g., Domino or True
Single-Phase Clocking) due to their higher switching speeds compared to the static logic
family. Dynamic circuits are, however, more susceptible to crosstalk noise compared to
the static logic because during some phase of the clock, the logic value is only stored on a
floating capacitor. An induced noise that changes the logic value on this floating capacitor
can cause the circuit to produce an incorrect result. Furthermore, as the circuit speed
increases and the signal transient times decrease, the effects of on-chip crosstalk noise
becomes more pronounced. Fig. 2 shows N neighboring wires. High frequency operation
of VLSI circuits causes on-chip wires to exhibit transmission line effects, and hence, elec-
trical and magnetic couplings start to take place between pairs of wires. These electric and
magnetic couplings reshape the signal waveforms and may induce delay in the signals
traveling through the lines, thus causing the circuit to possibly violate its timing con-
straints.

6
 ..
.

Fig. 2. Circuit schematic of N on-chip interconnects.

Our goal is to develop a circuit model for the capacitive coupling between on-chip cou-
pled interconnects and then use this model to derive a closed-form expression for the
crosstalk noise. We start our analysis by reviewing Devgan’s metric and its drawbacks in
estimating the crosstalk noise in RC circuits. For a more comprehensive explanation of
this metric, please refer to [11].

2.1 Devgan’s metric for crosstalk noise estimation

Consider two capacitively coupled RC networks as shown in Fig. 3.

Rs1+R11
V11 R12 V12 R1N V1N
......
C11 C12 C1N

Cc1 Cc2 CcN

Rs2+R21 R22 R2N

......
C21 C22 C2N
V21 V22 V2N
Fig. 3. Circuit schematic of capacitively coupled aggressor and victim nets.

The RC ladder network representing the aggressor net is driven by a flattened ramp
voltage whereas the RC ladder network representing the victim net is quiet. For this cir-

7
cuit, the node voltage vector at the victim net, V2 ∈ ℜ N × 1 , is related to the voltage vector

at the aggressor net, V 1 ∈ ℜ N × 1 , through the following equation:

–1 –1
[ ( sC 2 – A22 ) – sC c ( sC 1 – A 11 ) sC c ]V 2 = – sC c ( sC 1 – A 11 ) B 1 V s (1)

where C i = diag ( C ij + Cc j ) for i = 1, 2 and j = 1, 2, . . ., N and C c = diag ( – C cj )

for j = 1, 2, . . ., N. A11 and A22 represent the equivalent node resistance matrices of the
aggressor net and the victim net, respectively. The steady-state values of node voltages at
the victim net is calculated as:

–1 –1 VDD
V 2, ss = – A22 C c A 11 B 1 ----------
t - (2)
r

where tr is the rise-time of the input signal of the aggressor line driver. For simplicity, it is
assumed that the rise and fall times are equal. Note that this result is valid only if the driv-
ing voltages of the interconnects are infinite ramps. This is a critical assumption that seri-
ously and adversely affects the accuracy of capacitive crosstalk estimation. In practice, the
actual driving voltages of the interconnects are saturated ramp inputs rather than infinite
ramps. This means that the node voltages at the victim net reach their peak value approxi-
mately at t = t r . This peak value is obviously different from the steady-state value under
the infinite ramp input, and the error between these two values can be quite large if the
rise-time of the input is fast.
To better understand the shortcoming of this approach, consider two second-order RC
circuits with two floating capacitances connecting the corresponding nodes of these two
circuits as depicted in Fig. 4. Rs1 and Rs2 represent the resistances of the input source. In
reality, they represent the on-resistances of line drivers. The output impedance of line
drivers can be modeled as another RC section that is connected to the distributed RC inter-
connect. Consider typical values of these parasitics, i.e., assume that C1 = 60fF, C2 =
120fF, R2 = 50Ω, R1 = 20Ω, Cc = 180fF, Rs1 = 100Ω, Rs2 = 150Ω, and tr = 0.1ns.

8
 Rs1+R1 V11 R1 V12
C1 C1
Cc Cc
Rs2+R2 R2

C2 V21 C2 V22

F20 ig. 4. A pair of capacitively coupled second-order RC circuits.

From HSPICE simulation, the reported peak value of voltage V22 at the far-end of the
victim line is 0.416V. Devgan’s metric for the two coupled RC sections yields the follow-
ing equations:

VDD
V21, ss = 2 ( R 2 + Rs 2 )Cc ----------- (3)
tr

VDD
V22, ss = ( 3R 2 + 2R s 2 )Cc ----------- (4)
tr

Using Eq. (4), V22,ss is 1.053V. The estimated error is 153% verifying a well-known
observation established by earlier published works that this metric can be inaccurate for
deep sub-micron technologies [14].
Because the input signal rise-time is small, the crosstalk waveform rolls down quickly,
and consequently, the error becomes unacceptably large (cf. Fig. 5). Notice that for cases
where the input signal rise-time is large compared to the interconnect delays, Devgan’s
metric accurately predicts the peak value. Unfortunately, cases in which the estimations
are accurate (i.e., the slow slew rates for the pulses), are not the most important ones from
a circuit performance viewpoint. The reason is that the peak value of the crosstalk is
inversely proportional to the input rise-time. For slow slew waveforms, the crosstalk also
has a small peak value, and therefore, it has little impact on the circuit delay and the tim-
ing failure rate.

9
 Two Coupled RC Circuits
Time=4.43e-09
1.3

1.02
1
1.053V
Voltages (lin)

500m 0.416

-500m
6n 8n
Time (lin) (TIME)

Fig. 5. The output voltage and the crosstalk of two coupled second-order RC circuits.
C1 = 60fF, C2 = 120fF, R2 = 50Ω, R1 = 20Ω, Cc = 180fF, Rs1 = 100Ω, Rs2 = 150Ω, and tr
= 0.1ns.

In the next section, we derive a new, more accurate noise metric, and compare our
results with Devgan’s results and with HSPICE simulations.

2.2 A new metric for crosstalk noise estimation

Close examination of the HSPICE results shown in Fig. 5 reveals a major source of inac-
curacy in Devgan’s metric. The large error in this example comes from the fact that the
time constants of the exponentially rising portions of the victim node voltages, V2j for j =
1, 2, . . ., N, in the circuit of Fig. 5 are comparable to (or larger than) the input rise time.
The actual peak value of the crosstalk occurs approximately at t = tr (or tf , the fall-time,
whatever the case may be). In fact, other works (e.g., [13]) proved theoretically that the
peak noise is at tr, under one-pole or two-pole approximation. For a flattened ramp input,
it is easily seen that tr( f ) sets a lower bound on the time instance at which the peak value
of the crosstalk occurs. This is because, for an infinite ramp input, the voltage waveforms

10
at the victim line nodes monotonically increase toward their steady state values as pre-
dicted by [11]. As the current drive capability of the line drivers decreases or as the driver
sizes of the aggressor and victim lines become very different from one another, the peak
value of the crosstalk may occur further away from t = tr( f ). In contrast, as will be demon-
strated through a number of circuit simulations later in this section, our proposed metric
produces an accurate noise peak value and noise pulse width for all possible scenarios
with regard to the victim and aggressor line drivers.

To compute the noise peak value, we observe that the capacitive crosstalk noise at
every node of the victim net is a rising exponential function during the time interval that
the input signal of the aggressor line driver is rising. The actual peak value of the crosstalk
noise at each node of the victim net is in fact the value of the corresponding rising expo-
nential function at t = tr( f ). Recall that the steady-state value of this exponential function
is determined by Devgan’s metric.

   tr ( f )   
V2, m ax = V2, ss  I – exp  diag – --------
-   for j = 1, 2, ..., N (5)
   τ dj   

where diag (x) represents a diagonal matrix with all diagonal entries set to x. τ dj is the

time constant of the j-th node voltage in the victim net, and V2,ss is the vector of steady
state values of the crosstalk noise voltages at the victim nodes as calculated by Devgan’s
metric. Each node in the victim net sees two capacitances: a grounded area capacitance,
C2j, and a floating coupling capacitance, Ccj. The time constant at each victim node is thus
equal to the summation of individual time constants due to each of these two capacitances.
Similar to the open-circuit time-constant method that is employed for estimating the band-
width of high-frequency amplifiers [18], the time-constant due to each capacitance is
obtained by calculating the equivalent resistance seen across each capacitance with all the
other capacitances open-circuited. Therefore, to accurately estimate the time constants due
to capacitances C2j and Ccj, we first construct an equivalent circuit consisting of C2j , Ccj,
and the equivalent resistances seen across these two capacitances and replace all of the
other capacitances with open circuit connections. This circuit model is shown in Fig. 6.

11
 j
R1j,eq
V1,j-1 Ccj
R2j, eq = ∑ R 2i
i=1
j

R2j,eq C2j R1j, eq = ∑ R1i

i=1

Fig. 6. The equivalent circuit for computing the time constant of the j-th node of the victim net.

The characteristic polynomial of this second-order transfer function is:

τ vj

Λj ( s ) = R1j, eq R2j, eq C2j Ccj s 2 + [ (R 1j, eq + R 2j, eq )C cj + R 2j, eq C 2j ]s + 1 (6)

The time constant of this second-order circuit, which is roughly the inverse of the 3-dB
bandwidth of its system transfer function, is equal to the coefficient of the first-order term,
denoted by τ vj . In fact, τ dj must contain this coefficient as a part of its expression. Notice

that the input voltage source must be a unit-step function for τ vj to properly represent the

time constant at the j-th node of the victim net. This is obviously not the case for the dis-
tributed coupled RC circuits. The input voltage to the j-th node of the aggressor experi-
ences an RC delay due to the RC path from the input to the j-th node of the aggressor as is
shown in Fig. 7. The RC delay cannot be computed using the Elmore delay formula
because there is a signal path from the input node to each victim node V2j through the
floating capacitance Ccj that connects this victim node to the corresponding aggressor
node V1j. This RC delay also ought to be accounted for in the delay calculation of Fig. 7.
The overall RC delay is thus computed differently from the Elmore delay.

12
 Rs1+R11 V11 R12 V12 R1j V1j R1N V1N
....... .......
C11 C12 C1j C1N CL1
Cc1 Cc2 Ccj CcN
Rs2+R21 R22 R2j R2N
....... .......
C21 V21 C22 V22 C2j V2j C2N V2N CL2

Fig. 7. The RC signal paths through the aggressor line and floating capacitances

Furthermore, for RC circuits with orders greater than one, the initial slopes of the step
and ramp responses are zero. This zero initial slope leads to an increase in the circuit
delay. Fig. 8 indicates all these delay effects on the crosstalk noise waveform for a flat-
tened ramp input as well as a step input.
Step response vs. ramp response
2

1.8

1.6

1.4

1.2

Flattened ramp input Step input

Voltages (lin)

Crosstalk for a step input

800m

600m

Crosstalk for a flattened ramp input

400m

200m

4n 5n 6n
Time (lin) (TIME)

Fig. 8. Effects of zero initial slope and RC delay on the crosstalk.

13
 Consequently, the time constant of the j-th node in the victim net consists of two addi-
tive terms τ vj and τ aj , with τ vj (given by Eq. (6)) represents the time constant of the j-th

node in the victim net under a unit-step input excitation, and τ aj represents the propaga-

tion delay of the signal coming from other paths established by the floating capacitances
Cck (k = 1, ..., j − 1) as illustrated above. According to Fig. 7, τ aj includes the signal

delays of all additional signal paths through the coupling capacitances Cck (k = 1, ..., j − 1)
toward the j-th node in the victim net. The overall delay from the aggressor input source to
the j-th node in the victim net is:

τ d j = ζ ⋅ [ ( R 1j, eq + R 2j, eq )C cj + R 2j, eq C 2j + τ a j ] for j=1, 2, ..., N (7)

where τ aj is:

j–1
τ aj = R 1j, eq ( C cj + C 1j ) + ∑ [ R 1k, eq ( C ck + C 1k ) + R2k, eq ( C 2k + C ck ) ]
k=1

for j=1, 2, ..., N (8)

and ζ is a constant factor for the delay increase due to the nonzero, yet finite, input slope.
Its value is in the range [1.00, 1.02]. Throughout our analysis, we use ζ = 1.01 . Combin-
ing equations (7) and (8) yields the following expression for τ dj :

j
τ d j = ζ ⋅ R 1j, eq C cj + ∑ [ R 1k, eq ( C ck + C 1k ) + R 2k, eq ( C 2k + C ck ) ]
k=1

for j=1, 2, ..., N (9)

The peak amplitude of the crosstalk is easily obtained from equation (5) with the expres-
sion for τ dj given by equation (9).

As a special case, we first concentrate on the circuit of Fig. 4 in which two second-
order RC circuits are capacitively coupled. The peak voltage value of the node V22 is cal-
culated using three different approaches; HSPICE simulation, Devgan’s metric, and our
proposed analytical model. Applying equations (5) and (9) to the circuit of Fig. 4 yields

14
the following closed-form expressions for the peak values of the nodes, V21 and V22:

  tr  
V21, max = V21, ss  1 – exp – ------  (10)
  τ d1  

where τ d 1 = 1.01 [ ( R 1 + Rs1 ) ( 2Cc + C 1 ) + ( R 2 + R s2 ) ( Cc + C2 ) ]

  tr  
V22, max = V21, ss  1 – exp – ------  (11)
  τd2  

where τd = 1.01[ ( 3R 1 + 2R s1 ) ( Cc + C 1 ) + ( 3R 2 + 2R s2 ) ( C2 + C c ) + ( 2R 1 + R s1 )C c ]
2

To verify the accuracy of our approach on multistage RC networks in comparison to

other expressions proposed in [10], [13] and [14], we perform a number of experiments on
a two-line structure in a 130nm CMOS technology. In state-of-the-art CMOS technologies
the coupling capacitance accounts for approximately 70-95% of the total node capaci-
tances, which makes the coupling noise analysis even more important. In our implementa-
tion of the algorithm presented in [14], the second-order reduced transfer function is
utilized in order to avoid a potential stability problem. The coupled lengths of the adjacent
interconnects are varied from 200µm to 8mm. The supply voltage is VDD = 1.3V. Results
are reported for a range of rise-times varying between 30ps and 200ps, and for different
victim and aggressor driver resistances varying between 20Ω and 5kΩ . Table 1 contains
the result of these comparisons. Of particular interest is the situation where the driver
strengths of the aggressor and victim lines are vastly different. For instance, this circuit
configuration is encountered when a global signal line is in the close vicinity of a local
signal line. The last five experiments in Table 1 are devoted to this particular configura-
tion. The mean and maximum error values are reported in Table 2. These tables testify to
the higher accuracy of our approach compared to these other approaches. More precisely,
our proposed analytical model results in an average estimation error of only 5.82% which
is better than the 6.81% average estimation error resulting from the method of reference
[14] when a second-order reduced transfer function is employed. Interestingly, the pro-
posed noise metric exhibits a better accuracy compared to [14] when the driver sizes of

15
Table 1: The results of simulations on the two capacitively coupled transmission lines using star-HSPICE, comparing
methods of references [10], [13], [14] and our proposed metric using a 130nm technology and VDD =1.3V. The
aggressor line input is a flattened ramp signal
c1 r1 r2 c2 cc Cout1 Cout2 tr L HSPICE Pan Vittal Kuhlmann Ours
Rs1 Rs2
(pF/m) (kΩ/m) (kΩ/m) (pF/m) (pF/m) (pF) (pF) nsec (mm) volts volts volts volts volts

60 11.47 10.2 64 100 1k 2k 0.09 0.08 0.05 0.2 0.113 0.137 0.1396 0.119 0.1271
72 9.55 9.55 72 150 527 527 0.2 0.1 0.04 0.7 0.168 0.201 0.2111 0.192 0.181
83 8.2 7.0 90 160 1013 920 0.3 0.1 0.03 0.8 0.143 0.165 0.1937 0.152 0.16
92 9.3 10 80 170 270 400 0.06 0.2 0.06 2.5 0.331 0.383 0.401 0.352 0.361
101 12 12 101 150 140 150 0.1 0.2 0.08 1.2 0.189 0.221 0.2096 0.202 0.1901
120 10 10 120 132 2k 4k 0.06 0.07 0.1 1.1 0.255 0.286 0.3573 0.268 0.285
108 15 15 108 200 20 30 0.2 0.1 0.15 1.3 0.0901 0.0972 0.1234 0.0905 0.089
130 13 20 100 220 670 720 0.3 0.05 0.09 1.6 0.26 0.278 0.324 0.2782 0.267
140 17 17 100 200 350 350 0.3 0.2 0.12 2 0.222 0.261 0.271 0.239 0.221
140 17 17 100 200 20 30 0.07 0.08 0.06 2 0.25 0.263 0.26 0.273 0.25
90 11 11 90 120 160 160 0.3 0.2 0.08 4 0.219 0.254 0.266 0.238 0.232
85 12 8.5 75 140 190 100 0.8 0.2 0.l 5 0.144 0.156 0.1847 0.157 0.153
65 13 7 120 170 600 80 0.08 0.5 0.05 6 0.075 0.089 0.1 0.078 0.082
110 6.5 7 90 180 80 75 0.6 0.7 0.08 7 0.238 0.261 0.305 0.264 0.259
100 8 14 40 110 200 1k 0.8 0.05 0.06 8 0.465 0.512 0.6365 0.489 0.513
92 13 7 170 260 1.5k 50 0.08 0.3 0.08 0.8 0.0218 0.0231 0.0232 0.0224 0.0229
110 8 4 200 300 1.2k 25 0.1 0.65 0.1 3 0.0185 0.0207 0.0212 0.02 0.0192
70 10 1.82 20 95 20 3k 0.2 0.07 0.2 1.1 0.573 0.584 0.584 0.597 0.596
100 7 16 14 100 37 5k 0.8 0.1 0.1 2 0.647 0.689 0.7274 0.688 0.693
97 9 15 30 120 26 2k 0.3 0.06 0.07 1.2 0.675 0.712 0.7062 0.724 0.6862

Table 2: Percentage Error comparison for methods of references [10], [13], [14] and our proposed metric
%Error %Error %Error %Error
Pan’s Vittal’s Kuhlmann Ours
21.2 23.54 5.31 12.48
19.64 35.45 14.29 7.74
15.4 26.6 6.29 11.89
15.7 21.15 6.34 9
16.9 10.9 6.88 0.5
12.16 40.12 5.1 11.76
7.88 36.9 0.4 1.2
6.9 24.62 8.46 2.69
17.6 22.07 7.66 0.4
5.2 4.0 9.2 0
15.98 21.46 8.68 5.94
8.33 28.26 9.03 6.25
18.67 33.3 4 9.33
9.66 28.15 10.92 8.82
10.1 36.88 5.16 10.32
5.96 6.4 2.75 0.4
11.89 14.6 8.11 3.78
1.92 1.92 4.2 4.1
7.73 12.43 6.34 7.11
5.48 4.62 7.26 2.67
Average 11.72 21.67 6.81 5.82
Maximum percentage error 19.64 40.12 14.29 12.48
16
aggressor and victim lines are hugely different (i.e., the last five rows of Tables 1 and 2).
We expect that the method proposed in [14] gives rise to a higher accuracy once the third-
order reduced transfer function is used. However, this increased accuracy comes at the
expense of higher computational complexity due to the stability evaluation of the reduced
system. Notice that the metric proposed in [12] and [14] involves multiple tree-traversals,
diagonal matrix-vector multiplications, each of which is of order O(n) where n is the num-
ber of segment points, in order to compute each moment of each victim net node. From
these moments, the noise waveforms at all victim nodes are subsequently calculated by
solving a linear system of equations. In contrast, our metric presents closed-form expres-
sions for noise waveforms at all victim nodes (including the far-end termination) with a
computational complexity of O(n). The approach proposed in [13] gives rise to a slightly
better accuracy for short aggressor nets, as also demonstrated in Table 1.

The run-times of the proposed noise metric with that of [14] is compared in Table 3. In
Table 3, the peak value of the far-end crosstalk in a pair of geometrically identical aggres-
sor and victim lines in a 130nm technology is obtained using both the proposed metric and
[14]. The line length varies from 4-8mm, and the supply voltage is 1.3V. To accurately
model the interconnect, every 2µm of each line segment is modeled with an RC ladder
network.

Table 3: Run time comparison between the proposed noise metric and [14] using a
130nm technology and VDD =1.3V

Wire-length Run-Time (sec)

(mm) Kuhlmann [14] The proposed metric
4 3 1
5 3 1
6 4 2
7 4 2
8 5 3

17
 Figures 9 (a), (b), and (c) show the crosstalk voltage waveforms obtained by using HSPICE
simulation for the last three experiments in Table 1, where the aggressor and victim driver sizes
are very different. As mentioned earlier in this section, it is assumed that the peak value of the
crosstalk occurs approximately at t = tr( f ). As the current drive capability of the line drivers
decreases, the peak value of the crosstalk may occur further way from t = tr ( f ). As an example,
Fig. 10 shows the HSPICE simulation result for the far-end crosstalk at the victim net along with
the input waveform to the aggressor net. As seen in this figure, the time tmax approximated by
[15] is quite different from tr( f ). To examine the accuracy of the proposed metric for this case,
Table 1 includes experiments where the source resistances Rs1 and Rs2 are large, and the neigh-
boring lines are long. For instance, in one experiment given in Table 1, where the line length is
4mm, per-unit length parameters are c1 = c2 = 90pF/m, r1 = r2 = 11kΩ/m, cc = 120pF/m, and the
source resistances are Rs1 = 160Ω and Rs2 = 160Ω. Our metric predicts the peak crosstalk value to
be 0.232V, which constitutes less than 5.94% estimation error compared to the HSPICE result.
Notice that the error is kept below 5.94% although the peak value does not occur at tr. It is easily
proved that the proposed noise analytical model will result in accurate noise amplitude-pulse-
width product. Therefore, we can state that our metric calculates the two important attributes of
the capacitive crosstalk (i.e., the peak value and the noise pulse width) with a rather high accu-
racy.
Panel 1
1.3
Time=4.23147e-08 Time=4.95437e-08
1.4 Vxtalk,peak=0.647592V
1.2 Vxtalk,peak=0.573V 1.3

1.2
1
1
800m
800m
0.647592
600m 0.572999
600m
400m 400m

200m 200m
0 0

-200m -200m

-400m
-400m
-600m
-600m
44n 46n 48n 50n 52n 54n 56n
42n 50n
Time (lin) (TIME) Time (lin) (TIME)

(a) (b)

18
 Panel 1
Time=4.90402e-08
1.3

1.2 Vxtalk,peak=0.674467V
1

800m
0.674467

600m

400m

200m
0

-200m

-400m

-600m

50n 52n 54n 56n

Time (lin) (TIME)

(c)
Fig. 9. Crosstalk waveforms obtained using HSPICE simulations for the last three experiments in
Table 1.
The aggressor input and output; and the far-end crosstalk

1.2

1 Input to line 1
Output of line 1
Voltages (lin)

800m

600m

400m

200m

0
tr tmax
-200m
48n 49n 50n 51n 52n 53n
Time (lin) (TIME)

Fig. 10. The voltage waveforms at the input and output terminals of two adjacent interconnects of
4mm length and with per-unit length electrical parameters c1 = c2 = 90pF/m, r1 = r2 = 11kΩ/m, cc
= 120pF/m.

The proposed analytical model can be extended to any RC tree network. Shown in Fig. 11 is
the circuit consisting of a tree network a single interconnect line as a victim net. Such interconnect
topology is often encountered in an integrated circuit, where local clock distribution networks are
placed in the neighborhood of other signal lines. The circuit is realized in 130nm CMOS technol-
ogy.
19
 V1
l1 (mm)
l1 (mm) 0.04l1
V2
l2 (mm)
0.04l1 l1 (mm)
V3

Fig. 11. A simple tree-network capacitively coupled to a single victim net.

Our model is compared with [10], [11], and [14]. Table 4 reports the results of these
comparisons for a wide range of rise-times varying between 40ps and 400ps, different line
lengths, and different driver sizes. Without loss of generality, we assume that l2=0.6l1. To
accurately model the interconnect, we model every 20µm of each line segment with an RC
ladder network, which results in sufficiently accurate HSPICE simulation for l1 varying

Table 4: The results of simulations on the crosstalk in the victim net of Fig. 12 using star-HSPICE, paper [10], paper
[11], and our metric using a 130nm technology and VDD =1.3V

c1 r1 r2 c2 cc Cout1 Cout2 tr l1 HSPICE Devgan Vittal Ours

Rs1 Rs2
(pF/m) (kΩ/m) (kΩ/m) (pF/m) (pF/m) (pF) (pF) nsec (mm) volts volts volts volts

72 9.55 9.55 72 150 527 527 0.2 0.1 0.04 0.7 0.294 3.24 0.331 0.324
83 8.2 7.3 90 160 1013 920 0.3 0.1 0.03 0.8 0.301 5.3 0.321 0.343
92 9.3 10 80 170 270 400 0.06 0.2 0.06 2.5 0.469 4.5 0.511 0.502
101 12 12 101 150 140 150 0.1 0.2 0.08 1.2 0.282 0.662 0.328 0.297
120 10 10 120 132 2k 4k 0.06 0.07 0.1 1.1 0.345 6.86 0.402 0.394
108 15 15 108 200 20 30 0.2 0.1 0.15 1.4 0.118 0.118 0.127 0.117
130 13 20 100 220 670 720 0.3 0.05 0.09 1.6 0.368 4.36 0.425 0.401
140 17 17 100 200 350 350 0.3 0.2 0.12 2 0.326 2.02 0.39 0.362
140 17 17 100 200 30 30 0.07 0.08 0.06 2 0.46 0.85 0.531 0.488

Table 5: Error comparison for three noise metrics

%Error %Error %Error
Devgan’s Vittal’s Ours
1000 12.6 10.2
1660.8 6.64 13.95
859.5 8.96 7.04
134.8 16.31 5.32
1888.4 16.52 14.2
0 7.63 0.85
1084.8 15.49 8.97
425 9.6 5.1
84.78 15.43 6.1
Average 793.12 12.13 7.97
Maximum percentage error 1888.4 16.52 14.2
20
between 0.7mm and 2mm (cf. Table 4). Table 5 shows the estimated error of each model
as compared to HSPICE results.
The susceptibility of logic gates to noise depends not only on the peak amplitude of the
crosstalk noise but also on its duration. For example, digital circuits can often tolerate (and
indeed filter out) spike-like crosstalk noise with a large peak amplitude and very small
noise pulse width [10]. In static CMOS logic circuits, the peak amplitude of crosstalk does
not result in loss of signal values. Instead, it tends to cause an increase in propagation
delay along the victim line, which in turn may cause setup time violation in high-speed
circuits. These observations motivate the need for determining the noise pulse-width and
the time-domain noise waveform.

Given the equivalent time constants of the crosstalk noise, τ (drj ) and τ (dfj ) , the peak

amplitudes of the crosstalk noise, V2(,rSS) and V2(, fSS

)
, corresponding to the rising and falling
transitions of the input signal to the aggressor line, and the peak amplitudes of the
crosstalk waveform, the noise waveform is readily calculated. For the general case of
unequal rise and fall times, the noise waveform at each clock cycle time of Tc is computed
as follows:
Tc Tc
v2( t ) = v2r ( t )u ( t ) + v2f  t – ----- u  t – ----- (12)
 2  2

where v2r ( t )u ( t ) and v2f ( t )u ( t ) are defined as follows:

 (r)    t  
 V  I – exp  diag – ------- -   0 ≤ t ≤ tr
 2 , ss    τd( rj )  

v2r ( t )u ( t ) =  for j = 1, 2, . . ., N
 (r)   t – tr  
 V2, max exp diag – ----------
-  t ≥ tr
   τ (drj )  


21
  (f )    t  
 V  I – exp  diag – ------- 0 ≤ t ≤ tf
( f )  
, -
 2 ss
   τd j   

v2f ( t )u ( t ) =  for j = 1, 2, . . ., N
 (f)   t – tf  
 V2, max exp diag – ----------
(f)  
t ≥ tf
   τdj  


Notice that having the noise waveform gives us the maximum amount of information
regarding the noise behavior including the peak amplitude of crosstalk noise and the noise
pulse width. This information allows designers to find better solutions for noise avoid-
ance. Fig. 12 compares Eq. (12) with HSPICE simulation for a pair of capacitively cou-
pled nets. As one can see, our new metric accurately predicts not only the noise peak
amplitude but also the noise pulse width. Indeed, the effective pulse width is estimated
with a maximum error of only 5%. Our metric can easily be applied to the general case of
having several parallel runs of on-chip interconnects (on the same metal layer or on differ-
ent metal layers) by using the superposition principle [19].
Crosstalk noise waveform
0.35

Ours
HSPICE
0.3

0.25

0.2
Voltage

0.15

0.1

0.05

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time (sec) −10
x 10

Fig. 12. Crosstalk noise waveforms for two coupled transmission lines.

22
 Fig. 13 shows the change in crosstalk when the input rise time varies from 50ps to
300ps while all of the geometrical parameters are fixed.

Maximum crosstalk vs. the input rise−time

0.4

HSPICE
0.35 Ours
Devgan
Maximum corsstalk noise

Vital
0.3

0.25

0.2

0.15

0.1

0.05
2 4 6 8 10 12 14
−11
Input rise−time x 10

Fig. 13. Maximum crosstalk noise vs. input rise-time.

Comparing HSPICE results with our metric confirms that one achieves a high accu-
racy with our noise metric over a wide range of input rise-times. As expected, for long
rise-times, Devgan’s metric accurately predicts the peak amplitude of the noise. Vittal’s
metric exhibits higher fidelity compared to Devgan’s, i.e., its estimation error remains
roughly constant and does not present as large a dynamic range as Devgan’s. Recall that,
in [10], the authors use geometric considerations to obtain an expression for the crosstalk
noise. As a result, they do not account for effects of the non-ideal delays on the crosstalk
peak estimation. In contrast, our metric is based on the actual characteristics of capaci-
tively coupled RC circuits that are derived from simulations. In summary, our metric is
more accurate than those proposed in [10] and [11].

2.2 Crosstalk noise estimation for the saturating exponential input

23
 The single-pole model proposed in Section 2.2 can easily be extended to predict the
corsstalk noise when the input to the agressor line is a saturating exponential signal as
explained next. Consider

–t ⁄ τ s
V s ( t ) = V DD ( 1 – e )

where τs is the input time-constant. The crosstalk noise at the victim nodes becomes:

v2( t ) =
   1    t 
- – diag  ---
1 1   t 
VDD I – diag  --------------------- diag ----------
-  exp  diag – ----------
( ex )  τs- exp diag – τ---s- 
( ex )
 τd j – τs  ( ex )
τd    τd   
j j

for j = 1, 2, . . ., N

( ex )
where τd j represents the time constant of the j-th node in the victim net under an expo-
( ex ) ( ex )
nential input excitation. τdj is roughly equal to τd j = ( τs ln 2 ⁄ t r) τd j where tr is the rise-
time of the curve-fitting flattened ramp input, and τdj is the the time constant of the j-th
node in the victim net under the curve-fitting flattened ramp input. In fact, this approxima-
tion contributes to a larger noise estimation error.
Tables 6 and 7 demonstrates the accuracy of the proposed metric in calculating the peak
amplitrude of the coupling noise of two capacitively coupled interconnects in the presence
of a saturated expontential signal at the input of the aggressor net. The noise estimation of
the proposed metric is compared with the HSPICE simulation, and the techniques pre-
sented in [10], [12], and [14]. All comparisons are made using the device parameters for a
130nm standard CMOS process. The coupled lengths of the adjacent interconnects are
varied from 1.1-8mm.

24
Table 6: The results of simulations on the two capacitively coupled transmission lines using star-HSPICE, comparing
methods of references [10], [13], [14] and our proposed metric using a 130nm CMOS technology and VDD =1.3V.
The aggressor line input is a saturated exponential signal.
c1 r1 r2 c2 cc Cout1 Cout2 τs L HSPICE Pan Vittal Kuhlmann Ours
Rs1 Rs2
(pF/m) (kΩ/m) (kΩ/m) (pF/m) (pF/m) (pF) (pF) nsec (mm) volts volts volts volts volts

120 10 10 120 132 2k 4k 0.06 0.07 0.14 1.1 0.204 0.272 0.281 0.224 0.231
108 15 15 108 200 20 30 0.2 0.1 0.16 1.3 0.086 0.091 0.103 0.09 0.087
130 13 20 100 220 670 720 0.3 0.05 0.1 1.6 0.23 0.28 0.297 0.262 0.256
140 17 17 100 200 350 350 0.3 0.2 0.14 2 0.198 0.232 0.251 0.221 0.22
140 17 17 100 200 20 30 0.07 0.08 0.08 2 0.227 0.253 0.268 0.243 0.25
90 11 11 90 120 160 160 0.3 0.2 0.1 4 0.201 0.242 0.251 0.222 0.21
85 12 8.5 75 140 190 100 0.8 0.2 0.l2 5 0.127 0.143 0.162 0.141 0.138
65 13 7 120 170 600 80 0.08 0.5 0.06 6 0.072 0.083 0.089 0.074 0.079
110 6.5 7 90 180 80 75 0.6 0.7 0.1 7 0.218 0.241 0.256 0.232 0.239
100 8 14 40 110 200 1k 0.8 0.05 0.08 8 0.441 0.492 0.534 0.472 0.461

Table 7: Percentage Error comparison for methods of references [10], [13], [14] and our proposed metric
%Error %Error %Error %Error
Pan’s Vittal’s Kuhlmann Ours
33.1 37.75 9.8 13.24
5.81 19.77 4.65 1.16
21.7 29.13 13.91 11.3
17.2 26.77 11.6 11.1
11.45 18.06 7.05 10.13
20.4 24.88 10.45 4.48
12.6 27.56 11.02 8.66
15.28 23.61 2.78 9.72
10.55 17.43 6.42 9.63
11.56 21.09 7.03 4.54
Average 15.97 24.6 8.47 8.4
Maximum percentage error 33.1 37.75 13.91 13.24

5. Conclusions

In this paper, we presented an efficient analysis technique for the capacitive crosstalk
noise calculation in sub-quarter micron VLSI interconnects. We derived closed-form
expressions for the peak amplitude, the pulse width, and the time-domain waveform of
crosstalk noise. Experimental results show that the maximum error is less than 13% and
the average error is 5.82%.

25
References
[1] L. T. Pillage, R. A. Rohrer, “Asymptotic Waveform Evaluation for Timing Analysis”,
IEEE Trans. CAD, vol. 9, No. 4, pp. 352-366, April 1990.
[2] K. J. Kerns, A. T. Yang, “Stable and Efficient Reduction of Large, Multiport RC Net-
works by Pole Analysis via Congruence Transformations,” IEEE Trans. CAD, vol. 16, pp.
734-744, 1997.
[3] P. Feldmann, R. W. Freund, “Efficient Linear Circuit Analysis by Pade Approximation
via the Lanczos Process,” IEEE Trans. CAD, vol. 14, pp. 639-649, May 1995.
[4] A. Odabasioglu, M. Celik, L. T. Pileggi, “PRIMA: Passive Reduced-Order Intercon-
nect Macromodeling Algorithm,” IEEE Trans. CAD, vol. 17, No. 8, pp. 645-654, Aug.
1998.
[5] P. Heydari, M. Pedram. “Balanced Truncation with Spectral Shaping for RLC Inter-
connects,” IEEE Proc. ASP-DAC, pp. 203-208, Jan. 2001.
[6] P. Heydari, M. Pedram, “Model Reduction of Variable Geometry Interconnects Using
Variational Spectrally-Weighted Balanced Truncation,” Proc. IEEE/ACM Int. Conf. Com-
puter-Aided Design, San Jose, CA, Nov. 2001.
[7] K. Shepard, V. Narayanan, “Noise in Submicron Digital Design,” Proc. IEEE/ACM
Int. Conf. Computer-Aided Design, pp. 524-531, Nov. 1996.
[8] K. Shepard, V. Narayanan, P. C. Elmendorf, and G. Zheng, “GlobalHarmony: Coupled
Noise Analysis for Full-Chip RC Interconnect Networks,” Proc. IEEE/ACM Int. Conf.
Computer-Aided Design, pp. 139-146, Nov. 1997.
[9] A. Vittal, M. Marek-Sadowska, “Crosstalk Reduction for VLSI,” IEEE. Trans. Com-
puter-Aided Design, vol. 16, pp. 290-298, Mar. 1997.
[10] A. Vittal, L. H. Chen, M. Marek-Sadowska, K. P. Wang, S. Yang “Crosstalk in VLSI
Interconnections,” IEEE. Trans. Computer-Aided Design, vol. 18, no. 12, pp. 1817-1824,
1999.
[11] A. Devgan, “Efficient Coupled Noise Estimation for On-chip Interconnects,” IEEE
Proc. ICCAD, pp. 147-153. Nov. 1997.
[12] M. Kuhlmann, S. S. Sapatnekar, K. K. Parhi, “Efficient Crosstalk Estimation,” IEEE

26
IEEE Int. Conf. Computer Design, , pp. 266-272, Oct. 1999.
[13] J. Cong, D,. Z. Pan, P. V. Srinivas, “Improved Crosstalk Modeling for Noise Con-
strained Interconnect Optimization,” IEEE Proc. ASP-DAC, pp. 373-378, Feb. 2001.
[14] M. Kuhlmann, S. S. Sapatnekar, “Exact and Efficient Crosstalk Estimation,” IEEE
Trans. Computer-Aided Design, vol. 20, no.7, pp. 858-866, July 2001.
[15] M. Takahashi, M. Hashimoto, H. Onodera, “Crosstalk Noise Estimation for Generic
RC Trees,” IEEE Int. Conf. Computer Design, pp. 110-116, Sept. 2001.
[16] Li Ding, D. Blaauw, P. Mazumder, “Efficient Crosstalk Noise Modeling Using
Aggressor and Tree Reductions,” IEEE/ACM Int’l Conf. Computer-Aided Design, pp.
595-600, Nov. 2002.

[17] P. Heydari, M. Pedram, “Analysis and Reduction of Capacitive Coupling Noise in

High-speed VLSI Circuits,” IEEE Int’l Conf. on Computer Design (ICCD), pp. 104-109,
Sept. 2001.
[18] T. H. Lee, The Design of CMOS Radio-Frequency Integrated Circuits, Cambridge
University Press, pp. 146-152, 1998.
[19] Chua, L. O., Desoer, C. A., Kuh, E. S., Linear and Nonlinear Circuits, McGraw
Hill, New York, 1987.

27