DesignCon 2010

Rigorous Modeling of Transmit

Jitter for Accurate and Efficient
Statistical Eye Simulation

Fangyi Rao, Agilent Technologies

fangyi_rao@agilent.com, 408-553-4373
Vuk Borich, Agilent Technologies
vuk_borich@agilent.com, 408-553-2605

Henock Abebe, Agilent Technologies

henock_abebe@agilent.com, 972-699-2550

Ming Yan, Agilent Technologies

ming_yan@agilent.com, 818-879-6323
Abstract
Statistical analysis is well-established as an efficient alternative to Monte Carlo methods for low
BER simulation of serial link channels. Rigorous modeling of TX jitter presents both practical
and theoretical challenges in statistical analysis applications because of its time and pattern-
dependent nature. This paper reports on a new approach to statistical channel analysis that
incorporates rigorous and accurate models of TX jitter and its various components, including RJ,
PJ and DCD. Efficient simulation is made possible by a new, P-complete algorithm. Results
between the proposed method and long Monte Carlo simulations are compared and shown to be
in excellent agreement.

Author Biographies
Fangyi Rao received his Ph.D. degree in physics from Northwestern University in 1997. He
joined Agilent EEsof in 2006 as a Senior Development Engineer, where he works on Analog/RF
and SI simulation technologies in ADS and RFDE. From 2003 to 2006 he was with Cadence
Design Systems, where he made key contributions to the company's Harmonic Balance
technology and perturbation analysis of nonlinear circuits. Prior to 2003 he worked in the areas
of EM simulation, nonlinear device modeling, and optimization.

Vuk Borich is the Operating Manager for Analog/RF Simulation and Models at Agilent EEsof.
He earned a Ph.D. degree in Electrical Engineering from the University of Michigan, Ann Arbor,
in 2000. He joined Agilent EEsof in 2005, contributing to harmonic balance simulation of
frequency dividers. From 2003 to 2005 he was with Cadence Design Systems, where he worked
on the Ultrasim FastSpice simulator as a Senior Member of Technical Staff. From 2000 to 2003
he was with Applied Wave Research, first as Senior Development Engineer and then as Manager
of Simulation Technology, where he made several key contributions to the company's harmonic
balance technology, oscillator analysis, and phase noise simulation.

Henock Abebe received his Masters degree in Electrical Engineering from Southern Methodist
University. He joined Agilent EEsof in 2000 as a Development Engineer, where he works on
different parts of the ADS circuit simulator.

Ming Yan received his Ph.D. degree in electrical engineering from Illinois Institute of
Technology in 2005 for the research on EM simulation and semiconductor device modeling. He
joined Agilent EEsof in 2006 as a Software R&D Engineer, where he works on device modeling
and simulation technologies in ADS and RFDE. His most recent contributions are in the field of
SI simulation technologies in ADS.
Introduction
As data rates in serial link systems advance, the nature of channel behavior becomes increasingly
stochastic under the influence of various sources of jitter. As a result, statistical performance
measurements, including bathtub curves and BER contours, become critical in channel design
and verification. Since many applications specify eye openings at very low BER (e.g., 1e-12),
where brute-force Monte-Carlo simulation isn’t practical, methods such as StatEye have been
developed to directly compute eye probability density functions (PDFs) by statistical calculations
[1,2]. Statistical methods are efficient in calculating eye properties to extremely low BERs and
have been widely adopted by the signal integrity community.

Because of its importance to high-speed serial link performance, accurate modeling of TX jitter
and its various components -- including random jitter (RJ), periodic jitter (PJ) and duty-cycle-
distortion (DCD) – is required in statistical channel simulations. To model the TX jitter
contribution to the PDF at the channel output, StatEye-like methods use a formula of the form:

p(v, t )   p ISI (v, t   )  p jitter ( )d (1)

where pISI (v, t ) is the channel ISI PDF and p jitter (t ) is the TX jitter distribution. As the paper
will show, this formula relies on simplified assumptions about TX jitter that lead to optimistic
performance predictions [3]. In particular, they are valid only at the uniform jitter limit and fail
to model effects of the high frequency behavior of PJ, RJ and DCD, and phenomena such as
jitter amplification by the channel.

The recursive algorithm for fast ISI evaluation described in [1] is the key to efficient simulations
that have made StatEye-based methods popular and de-facto standard for statistical channel
analysis. The ideal transition time assumption that makes recursive ISI evaluation possible, as
the paper will demonstrate, is violated by the presence of TX jitter. As a result, rigorous jitter
modeling poses non-trivial challenges both in theory and in practical applications to highly
efficient channel simulations.

Drawing upon the principles laid out in [3,4], this paper offers a new approach to statistical
simulations that is both rigorous in its treatment of TX jitter – and therefore accurate – and
efficient in practical EDA simulations. We demonstrate TX jitter’s inherent time and pattern
dependent nature and discuss its impact on numerical complexity and simulation performance.
By including a rigorous, time- and pattern-dependent TX jitter model, our method is shown to
accurately capture important effects that include: asymmetric rise and fall edges due to DCD; PJ
frequency dependence; the effects of uncorrelated RJ sequence; jitter amplification by the
channel; and cross-talk and equalization effects. The numerical complexity introduced by TX
jitter is tackled by an advanced, P-complete simulation algorithm.

To verify the accuracy of our approach, statistical simulation results are extensively compared to
reference results obtained from a brute-force Monte Carlo method using long bit sequences,
under various TX jitter conditions described by our TX jitter model. Excellent agreement is
observed between the two methods. We also investigate simulation results obtained from a
StatEye-like approach using Equation (1) and explain their discrepancy with reference data.
A TX Jitter Model for Statistical Channel Simulation
A general TX jitter model is illustrated in Figure 1. It assumes uncorrelated RJ between different
rise and fall edges, represents PJ with a sinusoidal function of time, and includes DCD caused by
asymmetric rising and falling data edges. The rise time t r and the fall time t f of the ith TX pulse
are modulated by jitter functions  r and  f as:

t r (i )  nr (i )  T   r (i)
(2)
t f (i )  n f (i )  T   f (i )

where n r is the bit index of the first 1-bit of the pulse, and n f is the index of the first 0-bit that
follows it. T is the unit interval and n r T and n f T represent ideal rise and fall times. In terms of
RJ, PJ and DCD contributions,


 r (i )   r (i )  A cos[  nr (i )T   ] 
2
(3)

 f (i )   f (i )  A cos[  n f (i )T   ] 
2

where  r and  f represent random jitter at the rise and fall edges; A and  are the amplitude and
frequency of the periodic jitter; and  terms model DCD due to asymmetry between the rising
and falling edges, with opposite signs for  r and  f .

Equations (2) and (3) reveal several important characteristics of TX jitter. First, as shown in
Figure 1, jitter occurs only at physical rise and fall edges; in pulses with consecutive logic 1-bits
only the first and the last bits are affected by jitter. Second, random jitters  in Equation (3) are
uncorrelated between different edges. Third, the DCD term alternates in sign between edges. As
a result, all jitter components in the model are inherently time and bit pattern dependent.
Importantly, RJ and DCD vary with time at rates comparable to the data rate and thus contain
significant high frequency spectrum content.

Given a channel impulse response length M, measured in bits, a sample time t, and the expected
logic level at the channel output, the number of possible random bit patterns that contribute to
the output voltage v(t ) is 2 M  1 . For the mth pattern, the channel output v (m ) can be calculated
by the superposition of the rising step response R(t) and the falling step response F(t) of the
channel as:

 
v ( m) (t )   R[t  t r( m) (i)]  F[t  t (fm) (i)]  v0 (4)
i
where t r( m) (i) and t (f m ) (i ) are rise and fall times of the ith pulse in the mth pattern, and v 0 is the
output when all input bits are logic 0. Equation (4) shows that TX jitter and channel ISI are
interdependent and cannot be treated separately in general. Statistical calculations based on
rigorous TX jitter models must take into account jitter at all pulse edges in all patterns.
Enormous computational complexity is introduced to statistical simulations by jitter
dependencies on time and bit pattern.

i th pulse
0 1 1 0 0 1 0 1 0

t r (i ) nr (i)T n f (i)T t f (i) ideal edge

 r (i)  f (i)

Figure 1: TX jitter model. Solid vertical lines represent the actual rising and falling edges. Dash lines represent ideal
edges.

When TX jitter is zero, only ISI is responsible for the output jitter and Equation (4) becomes

( m)
v ISI  
(t )   R[t  nr( m) (i)  T ]  F[t  n (fm) (i)  T ]  v0 (5)
i

where nr( m) (i) is the bit index of the first 1 bit of the i th pulse in the m th pattern and n (fm ) (i ) is the
bit index of the first 0 bit follows that pulse. The ISI PDF p ISI of the expected logic level at time
t, as a function of output voltage, is given by the distribution of v (m ) for all 2 M  1 patterns.

1
p ISI (v, t )  
2 1 m
M
 [v  v ISI
( m)
(t )] (6)

where  is the Dirac delta function. When rise and fall edges are symmetrical, functions R and F
are the same and Equation (6) can be evaluated efficiently by recursive PDF convolution based
on the ideal single bit pulse response given by R(t)-R(t+T) [1,2].

StatEye-like TX jitter treatment, shown in Equation (1), follows from Equation (4) in the special
case where TX jitter is assumed to be constant. In that case, all  r and  f in Equation (2) have the
same value  , Equation (4) further simplifies to:

 
v ( m) (t , )   R[t  nr( m) (i)  T   ]  F[t  n (fm) (i)  T   ]  v0  v ISI
( m)
(t   ) (7)
i
The PDF of the expected logic level at the output is given by

1
p ( v, t )  
2 1 m
M   [v  v ( m ) (t , )]  p jitter ( )d (8)

where p jitter is the TX jitter distribution. Combining Equations (6) - (8), we arrive at the StatEye
jitter model in Equation (1):

p(v, t )   p ISI (v, t   )  p jitter ( )d (9)

This analysis demonstrates that TX jitter modeling by Equation (1) applies only if jitter is
independent of time/edge, or varies slowly within the length of the channel impulse response.
These conditions do not hold in practice where, as argued earlier, both DCD and uncorrelated RJ
carry high frequency components and PJ can vary significantly over impulse length.

Jitter Amplification by the Channel

Jitter amplification by the channel is a phenomenon arising from interactions between TX jitter
and channel ISI [5,6]. This effect is predicted by the rigorous TX jitter model in Equation (4).

Consider the output crossing time as given by Equation (4):

   
 R[t
i
(m)
c  t m nr( m ) (i )T 
2
  r( m ) (i )]  F [t c( m )  t m  n (fm ) (i )T    (f m ) (i )]  v0  vth
2 
(10)

where vth is the crossing threshold, t c(m ) is the zero-RJ crossing time for the m th pattern, and
t m the shift in crossing time induced by RJ  r(m) and  (m )
f . PJ is set to zero to simplify discussion.

Under the condition of small RJ, t (m) is given by linearization of Equation (10) around t c(m ) :

 R m (i )  r( m ) (i )  Fm (i )  (f m ) (i ) 
t m  i
(11)
 R
i
m (i )  Fm (i )

where

dR[t c( m )  nr( m ) (i )   / 2]
Rm (i ) 
dt
(12)
dF [t c  n (fm ) (i )   / 2]
(m)

Fm (i ) 
dt

Timing jitter at the output is characterized by the standard deviation  of the crossing time:
 2 
1
 
t c( m )  t m 2
 tc
2
(13)
2 1 m
M

where t c is the averaged crossing time defined as

1
tc  
2 1 m
M
t c( m ) (14)

The average  is taken over all  r(m) and  (m )

f . Equation (13) can be rewritten as

 2   DJ
2
  RJ
2

1
 DJ 
2
 t c( m )  t c
2 2
(15)
2 1 m M

1
 RJ
2
 M 
2 1 m
t m 2

As shown in Equation (15), the output jitter consists of deterministic jitter  DJ arising from ISI
and DCD, and random jitter  RJ induced by TX RJ. For uncorrelated TX RJ,

1 i Rm (i ) 2   r( m ) (i ) 2   Fm (i ) 2   (f m ) (i ) 2
 RJ
2
 M 
2 1 m
i
2
 
 Rm (i )  Fm (i )
 i 
(16)
1  Rm (i)   Fm (i)
2 2

  TXRJ
2

2 1 m 
M
i i
2

 Rm (i )  Fm (i )
 i 

where  TXRJ is the standard deviation of  r(m) and  (m )

f . Equation (16) shows that, due to ISI,

output RJ comprises contributions from multiple TX edges. When RJs in these edges are
uncorrelated, they contribute to the output independently and the effects are accumulated in the
timing variance. When the factor following  TXRJ in Equation (16) is great than one, TX RJ is
amplified by the channel. As shown in the next section, practical channels can exhibit jitter
amplification, especially at high data rates.

In the case of uniform TX RJ,  r(m) ,  (m

f
)
and t m are the same and  RJ   TXRJ . Equation (15)
becomes

 2   DJ
2
  TXRJ
2
(17)
Equation (17) also follows from Equation (1) at zero DCD, as a consequence of PDF convolution
between ISI and TX RJ. Therefore, in StatEye-like calculations the output RJ is always equal to
TX RJ. As a result, StatEye-like simulations fail to predict jitter amplification by the channel.

Rigorous TX Jitter Calculation in Statistical Simulation

To model TX jitter accurately and without approximation, the rigorous jitter model represented
by Equations (2) and (3) is now applied to statistical simulation. The output PDF of the expected
logic level is given by

1 2 1
p ( v, t ) 
2 
0
d 
2 1 m
M   [v  v ( m ) (t )]   g[ r( m ) (i)]  g[ (f m) (i)]  d r( m ) (i)  d (f m ) (i)
i
(18)

where  r( m) (i) and  (f m ) (i ) model random jitter at rising and falling edges of the i th pulse in the
m th pattern, and g ( ) is the RJ PDF, typically a Gaussian distribution. The product of g
functions in Equation (18) is the result of uncorrelated RJ at different edges. The channel output
v ( m) (t ) is given by Equation (4). The PDF is averaged over the PJ phase offset,  , to account for
its randomness relative to the data signal. As shown by Equation (18), our calculation includes
TX RJ, PJ and DCD due to asymmetric rising and falling edges.

As discussed earlier, TX jitter is both time and pattern dependent and must be treated
simultaneously with ISI. With the presence of TX jitter, pulse lengths are not ideal, so recursive
PDF convolution based on the single bit pulse response for pure ISI is no longer applicable. For
an impulse response M bits long, the numerical complexity of Equation (18) is O( 2 M ), making a
brute-force evaluation impractical.

To overcome the computational complexity introduced by TX jitter, an NP complete algorithm is

developed to evaluate Equation (18) efficiently. The channel output PDF is computed rigorously
and without approximation, such as step response linearization or low probability extrapolation,
making it possible to accurately evaluate link performance to arbitrarily low BERs. The method
has been extended to include the effects of crosstalk, equalization and receiver (RX) jitter.

Effects of crosstalk channels are taken into account as follows. When crosstalk is synchronous,
aggressors are applied at the same data rate and at a fixed phase relationship to the main channel.
When crosstalk is asynchronous, the aggressors’ phase relationship varies relative to the main
channel, either because their data rates are different or because the phase offset drifts. In the
proposed method, crosstalk signals are treated as additive noise at the output. When the primary
and aggressor transmitter bit sequences are mutually uncorrelated, the effects of crosstalk can be
calculated by convolving the output PDFs of the primary and crosstalk signals along the voltage
axis. For synchronous crosstalk,

p(v, t )   pmain (v  v1  v2   vn , t ) p xtlk

(1) ( 2)
(v1 , t ) p xtlk (n)
(v2 , t ) p xtlk (vn , t )dv1dv2 dvn (19)
 (i )
where pmain and p xtlk are the PDFs of the main channel and the i th crosstalk channel calculated
by Equation (18). For asynchronous crosstalk, in order to represent the random phase between
(i )
the main and crosstalk channels, p xtlk is averaged along the time axis across one crosstalk UI,
T (i ) . The averaged PDF is then convolved with pmain at each time point as follows:

1
( i )  xtlk
(i )
p xtlk (v )  p (i ) (v, t )dt
T (20)
p (v, t )   p main (v  v1  v2   vn , t ) p xtlk
(1) ( 2)
(v1 ) p xtlk (n)
(v2 )  p xtlk (vn )dv1dv 2  dv n

To model linear TX/RX equalization, continuous-time linear (CTLE) and feed-forward (FFE)
equalizers are combined with the channel impulse response by standard convolution techniques.
Under the approximation of zero decision error, decision feedback equalizers (DFE) are linear
and are included in Equation (4) as:

 
v ( m) (t )   R[t  t r( m) (i)]  F[t  t (fm) (i)]   ck  b ( m) (t  kT )  v0 (21)
i k

where c k is the k th DFE tap coefficient and b ( m) (t ) is the expected bit at the slicer output at time
t in the m th bit pattern.

RX sampling time jitter effectively shifts the eye along the time axis. The averaged eye is
calculated by convolving p(v, t ) with the RX jitter distribution.

Simulation Results and Discussion

A Molex channel, shown in Figure 2, is simulated under different TX jitter conditions. The
channel is represented by measured S-parameters as plotted in Figure 3. The same jitter model,
described by Equations (2) and (3), is applied in both the proposed statistical approach and in
Monte Carlo simulations.

Figure 2: Molex channel

 Insertion loss Return loss

Figure 3: S-parameters of the main channel

ISI Jitter
Figure 4 shows simulation results for 3 Gbps data rate with zero TX jitter, as calculated by the
proposed statistical approach. The output jitter is pure ISI so the bathtub curves at a low BER
indicate the worst-case eye closure. The ISI jitter RMS measured at the crossing level is 21.03ps.

Timing bathtub Voltage bathtub

Figure 4: Eye diagram and timing and voltage bathtubs at 3 Gbps with zero TX jitter.
Table 1 lists the simulated worst-case eye width and height and compares them to Monte Carlo
simulations with 103 and 106 bits. As expected, the longer the Monte Carlo run, the closer it
converges to statistical analysis predictions.

Method 1K bits Monte Carlo 1M bits Monte Carlo Proposed statistical approach
width 254ps 244ps 244ps
height 0.164V 0.157V 0.155V
Table 1: Width and height of the worst-case eye border at 3Gbps with zero TX jitter.

TX Random Jitter
In the next experiment, 5% UI (16.66ps) random jitter is applied to the TX. Eye diagrams and
bathtub curves are displayed in Figures 5 and 6. Results of the proposed approach, utilizing
Equations (2) and (3) to model jitter, are compared to a long Monte Carlo simulation and to a
StatEye-based approach making use of Equation (1) for TX RJ. In the Monte Carlo simulation,
where no sample data is available, the BER is set to zero (approximately below BER=1e-6).

The excellent agreement between the proposed statistical approach and the Monte Carlo
simulation within the applicable range (BER above 1e-6) demonstrates the accuracy of the
method presented in this paper. Note that, despite being induced by the same Gaussian TX RJ,
the output random jitters on different sides of the crossing time are asymmetric as a consequence
of pattern-dependence of TX jitter. Observe that the StatEye-based method overestimates the eye
opening: whereas the rigorous approach predicts a closed eye at 1e-12 BER, calculations using
Equation (1) predict an opening 41 ps wide and 33 mV high.
 (a) Proposed statistical approach (b) Monte Carlo with 1 million bits

(c) StatEye-based

Figure 5: Eye diagrams at 3 Gbps with 5% UI TX random jitter simulated by (a) the proposed statistical
approach, (b) one million bits Monte Carlo simulation, and (c) the StatEye-based method.

Timing bathtub Voltage bathtub

Proposed statistical approach Monte Carlo with 1e6 bits StatEye-based

Figure 6: Timing and voltage bathtubs at 3 Gbps with 5% UI TX random jitter.

Table 2 lists the total and random jitter components measured at the output. The random jitter is
extracted from the total jitter according to Equation (15) and using 21.03 ps ISI jitter RMS
reported in the earlier experiment. Results show that in the proposed method RJ is amplified by
the channel from 16.66 ps at the input to 20.94 ps at the output. As shown earlier in the paper,
the StatEye-based approach predicts no jitter amplification and underestimates random jitter at
the output.
 Method Total jitter RMS Random jitter RMS
Proposed statistical approach 29.68ps 20.94ps
StatEye-based 26.83ps 16.66ps
Table 2: Total and random RMS jitter at 3 Gbps with 5% UI TX random jitter. RJ RMS is extracted from total jitter
according to Equation (15), using 21.03 ps ISI RMS jitter with TX RJ equal to zero.

Figure 7 shows the RJ amplification factor of the channel as a function of data rate as calculated
by the proposed method. Amplification factor is defined as the RMS ratio between RJ at the
output and RJ at TX, which is fixed at 2%. As shown in the plot, the amplification effect
increases with the data rate, suggesting it is caused by ISI, as predicted by theory.

Figure 7: RJ amplification factor of the channel at different

data rates.

The efficiency of the proposed method is highlighted in Table 3 by the CPU time comparison
between statistical and Monte Carlo simulations (using 106 bits) on a typical workstation.

Method Proposed statistical approach Monte Carlo

CPU time 7 sec 198 sec
Table 3: CPU times of the proposed statistical analysis and the Monte Carlo simulation of one million bits at 3 Gbps
and 5% UI TX RJ.

TX Periodic Jitter
In this experiment, we compare statistical and Monte Carlo simulations with periodic jitter on the
TX. The PJ amplitude is 50 ps and two PJ frequencies, 3 kHz and 500 MHz, are considered.
Eye diagrams and bathtubs are shown in Figure 8 and 9, respectively and again show excellent
agreement at both PJ frequencies. Both methods predict a smaller eye opening at 500 MHz than
at 3 kHz, indicating frequency dependence of PJ effects.
 (a) Proposed method (500MHz PJ) (b) 1M bit Monte Carlo (500MHz PJ)

(c) Proposed method (3kHz PJ) (d) 1M bit Monte Carlo (3kHz PJ)

(e) StatEye-based

Figure 8: Eye diagrams at 3Gbps with TX periodic jitter simulated by (a) the proposed statistical method at 500MHz
PJ, (b) one million bits Monte Carlo simulation at 500MHz PJ, (c) the proposed statistical method at 3kHz PJ, (d)
one million bits Monte Carlo simulation at 3kHz PJ, and (d) the StatEye-based method. The PJ amplitude is 50ps.
 (a) Timing bathtub (3kHz PJ) (b) Voltage bathtub (3kHz PJ)

(c) Timing bathtub (500MHz PJ) (d) Voltage bathtub (500MHz PJ)

Proposed method 1M bits Monte Carlo StatEye-based

Figure 9: Timing and voltage bathtubs at 3Gbps with TX periodic jitter at 3kHz and 500MHz PJ
frequencies.

Figures 8 and 9 also include results obtained by a Stateye-based approach. In StatEye-based

calculations, the PJ contribution is modeled by Equation (1) and the PJ distribution is given by:

1
p PJ ( x)  (22)
 A2  x 2

where x is the jitter and A is the PJ amplitude. Equation (22) does not depend on frequency and
cannot account for the frequency-dependent effects. Note in Figure 8 and 9 that the StatEye
results match Monte Carlo at 3 kHz but differ at 500 MHz, confirming the expected failure of
Equation (1) to model high-frequency jitter.

The effect of PJ frequency dependence is further illustrated by plots of width and height of the
worst-case eye border as functions of PJ frequency in Figure 10. The PJ amplitude is 50 ps. Due
to limited sample size, Monte Carlo simulations always predict a larger eye border than the
proposed statistical method. Despite the difference, both results exhibit the same decreasing
trend at high frequency.
 Proposed method 1M bits Monte Carlo StatEye-based
.
Figure 10: Width and height of the worst-case eye border at 3 Gbps with TX periodic jitter at different
frequencies

TX DCD
Figure 11 shows eye diagrams at 3 Gbps with 15% TX DCD calculated by the proposed method
and a Monte Carlo simulation of 106 bits. Compared to Figure 4, the eye becomes asymmetric
and the crossing level is shifted upward by 33 mV, indicating longer logic 1 cycles. As discussed
in previous sections, Equation (2) models DCD caused by asymmetry between the rising and
falling edges of the data signal, where all rising edges shift in one direction and all falling edges
shift in the opposite direction. This type of DCD is manifested by a shift of crossing level toward
the logic level with longer duration. Such a shift cannot be reproduced by Equation (1) because
the convolution is performed along the time axis. Other types of DCD, caused by imbalance
between the 1 and 0 bits of the clock signal, can also be treated by the proposed method with
minor modifications.

(a) Proposed Method (b) 1M bits Monte Carlo

Figure 11: Eye diagram at 3Gbps with 15% TX DCD.

Crosstalk
Figure 12 shows the channel from Figure 2 subject to near end (NEXT) and far end (FEXT)
crosstalk. S-parameters between aggressor and receiver ports are plotted in Figure 13. Eye
diagrams and timing bathtubs with and without crosstalk are shown in Figures 14 and 15, as
calculated by the proposed statistical method. Both synchronous and asynchronous crosstalk are
simulated. For synchronous crosstalk, the relative signal phase between the aggressor and the
main channel is set to zero. As expected, the channel performance is impaired in all cases, with
more degradation caused by NEXT crosstalk because of stronger coupling to the main channel.

(a) NEXT (b) FEXT

Figure 12: Channels with NEXT and FEXT.

(a) NEXT (b) FEXT

Insertion loss Return loss

Figure 13: S-parameters between aggressor and receiver ports.

 (a) Synchronous NEXT (c) Synchronous FEXT

(b) Asynchronous NEXT (d) Asynchronous FEXT

Figure 14: Eye diagrams at 3 Gbps with crosstalk.

(a) NEXT (b) FEXT

w/o xtlk synchronous xtlk asynchronous xtlk

Figure 15: Timing bathtub curves at 3 Gbps with and without crosstalk.

Equalization
At 5 Gbps, the channel in Figure 2 requires equalization. We apply a 6-tap (2 pre-cursors and 4
post-cursors) FFE and a 6-tap DFE. Tap coefficients are optimized based on the channel impulse
response. A 2% UI Gaussian sampling jitter is applied on the RX. Eye diagrams, BER contours
and timing bathtub curves with and without equalization, as calculated by the proposed method,
are plotted in Figures 16-18. Table 4 lists eye width and height measured at 1e-12 BER with FFE
only and with FFE+DFE. As shown in Figure 16, the eye is completely closed by ISI without
equalization. It is opened by FFE alone and is further improved by DFE.

(a) w/o equalizer (b) 6-tap FFE only (c) 6-tap FFE + 6-tap DFE

Figure 16: Eye diagrams at 5 Gbps with and without equalization.

(a) 6-tap FFE only (b) 6-tap FFE + 6-tap DFE

Figure 17: Contours at BER from 1e-12 to 1e-4 at 5 Gbps with FFE and DFE.
 w/o equalizer 6-tap FFE only 6-tap FFE + 6-tap DFE

Figure 18: Timing bathtubs at 5 Gbps with and without equalization.

Equalizer 6-tap FFE 6-tap FFE + 6-tap DFE

Width 93ps 100ps
Height 0.201V 0.243V
Table 4: Eye width and height at 1e-12 BER at 5 Gbps with FFE and DFE.
.

Summary
Statistical methods are widely adopted for high speed serial link simulation. Well established
tools, such as StatEye, rely on approximate modeling of TX jitter and underestimate eye closure.
More rigorous modeling requires new methods, able to cope with the computational complexity
introduced by TX jitter. The paper reported on a highly efficient method that accommodates a
rigorous treatment of TX jitter influence on channel performance. It captures effects that include
RJ amplification by the channel, frequency dependent periodic jitter and asymmetric transitions
due to DCD. The method has been validated by extensive numerical experiments and detailed
comparisons with long Monte Carlo simulations.

References
1. B. Casper, M. Haycock, R. Mooney, “An accurate and efficient analysis method for multi-
Gb/s chip-to-chip signaling schemes,” VLSI Circuit Symposium, pp. 54-57, June 2002.
2. A. Sanders, M. Resso, and J. D’Ambrosia, “Channel compliance testing utilizing novel
statistical eye methodology,” DesignCon, 2004.
3. V. Stojanovic and M. Horowitz, “Modeling and analysis of high-speed links”, CICC 2003,
pp. 589-594, 2003.
4. B. Casper, G. Balamurugan, J. Jaussi, J. Kennedy, M. Mansuri, F. O’Mahony, and R.
Mooney, “Future microprocessor interfaces: analysis, design and optimization,” CICC 2007,
pp. 479-486.
5. S. Chaudhuri, W. Anderson, J. McCall, and S. Dabrai, “Jitter amplification characterization
of passive clock channels at 6.4 and 9.6 Gb/s,” Proc. IEEE 15th Topical Meeting on Electric
Performance of Electronic Packaging, Scottsdale, AZ, Oct. 2006, pp. 21-24.
6. S. Chang, D. Oh and C. Madden, “Jitter Modeling in Statistical Link Simulation”, EMC’08.