DesignCon 2019

Baseline Wander: Systematic

Approach to Rapid Simulation and
Measurement

Pavel Zivny, Tektronix, 503-627-4755

pavel.zivny@tektronix.com

Vladimir Dmitriev-Zdorov, Mentor A Siemens Business

vladimir_dmitriev-zdorov@mentor.com, +1 720-494-1196

Maria Agoston, Tektronix,

maria.agoston@tektronix.com
I. Abstract
High speed serial data transmissions can suffer loss of margin as well as data errors due to
baseline wander caused by mechanisms such as AC coupling. Thorough simulation and
measurement are difficult since they involve interaction between physical layer fidelity
features from sub-MHz to tens of GHz of signal spectra, and must include effects
accumulated over millions of UIs.
We show a method of rapid analysis of baseline wander that quickly spans the large
frequency range required, and consider its implementation in both bit-by-bit and statistical
analysis. We demonstrate this method on extreme cases of signals in PAM4 signaling, and
show measurement and simulation correlation.

II. Authors Biographies

Pavel Zivny
Pavel Zivny is a Domain Expert engineer with the sampling oscilloscopes group of
Tektronix. He holds an MSEE degree and has been with Tektronix for over 20 years, working
in test, design, and marketing of both real time and sampling oscilloscopes. Pavel was
granted oscilloscope related patents, authored industry articles and papers, and represents
Tektronix to high-speed Serial Data standards committees.

Vladimir Dmitriev-Zdorov
Dr. Vladimir Dmitriev-Zdorov is a principal engineer at Mentor, A Siemens Business. His
work includes development of efficient methods of circuit/system simulation, transformation
and analysis of multi-port systems, analysis of SERDES links. The results have been
published in numerous papers and conference proceedings.

Maria Agoston
Maria Agoston is Principal Engineer with Tektronix. Her contributions to Tektronix
instruments have been in the area of digital signal processing and jitter and noise analysis
techniques. She holds a MS in Computer Engineering from Oregon State University and a
MSEE from Polytechnic University of Bucharest. Maria has been awarded several patents in
area of sampling oscilloscope waveform processing, and has authored several papers on high
speed serial link analysis.
III. Introduction. What is Baseline Wander?

“Baseline wander is actually the effect where the base axis (X-axis) of any signal viewed on a
screen appears to 'wander' or move up and down rather than be straight. … I believe it is due
to improper electrodes (like rusted, or broken)”
www.crazyengineers.com

“...A long string 0s or 1s can cause a drift in the baseline (baseline wandering) and make it
difficult for the receiver to decode correctly”
https://www.coursehero.com

What is baseline wander (BLW)? Is this something caused by the lack of low frequency gain?
Is it caused by having imbalance in symbols in the transmitted pattern? In fact, both
conditions are necessary to cause BLW. It happens when channel has low frequency
bandwidth limitations that block or significantly reduce part of the signal spectrum. The
missing portion of the signal creates an error that manifests itself as a slow drift of the
signal’s median level. For example, PCIe channels include series decoupling capacitors. An
unconstrained data stream travelling through such a channel would cause BLW, as an
unconstrained signal can have a considerable amount of slowly changing data symbol
imparity. To avoid BLW, the first generations of PCIe used data encoding protocols (8b10b)
ensuring the signal pattern is DC-balanced. However, later generations employ less restrictive
encoding to reduce encoding overhead and improve data transmission efficiency. They use
scrambling that helps to distribute the energy of the signal evenly along frequency spectrum,
thus reducing occasional low frequency spikes that occur in a raw un-encoded data stream.
However, scrambling doesn’t eliminate the low frequency portion of the signal spectrum
completely, and hence – BLW remains a problem.
There are few works devoted to BLW in particular. However, the underlying phenomena
were addressed over time in many sources on digital transmission, such as [1-3]. We can find
there two major ways of mitigating BLW. One is using DC-balanced signal encoding
protocols that prevent accumulation of signal disparity. The other is BLW correction through
digital quantized feedback [4, 6]. Here, the missing BLW portion is “synthesized” on the
receiver end by applying the restored digital signal to a low pass filter, and adding filtered
waveform to the input of the receiver.
Both techniques have their limitations. DC-balanced protocols bear less net information per
symbol than unconstrained patterns, and ever growing demand for speed forces the designers
to trade BLW for efficiency. BLW correction is not widely used in modern high-speed
communication. It is costly, and it’s difficult to construct low pass filter that accurately
approximates missing portion of the channel’s transfer function.
Therefore, BLW noise remains an important factor that should be considered on the SERDES
design and verification stage. Growing popularity of multilevel signaling (especially PAM-4)
makes BLW impairment even more critical because of reduced separation between signal
levels. Hence, we need accurate and efficient methods of BLW simulation and measurement.

BLW and current simulation tools

Unfortunately, BLW is not adequately addressed in existing analysis tools. Consider
traditional bit-by-bit simulation of SERDES links, StatEye-type simulators, or IBIS AMI
analysis, both time domain and statistical. Which of them consider impairments caused by
 BLW? Most of the fast eye-diagram/BER evaluation methods are based on convolving a long
digital stimulus with a relatively short channel’s elementary response (impulse or step). For
efficiency, they use overlap-save version of IFFT, but cannot afford a very long channel
response (up to 1M symbols) that includes accurate near-DC bandwidth limitations. Even
generating such response for statistical analysis is problematic. Similarly, new channel
compliance evaluation methods, such as COM, ERL and others don’t have provisions to
consider effect of BLW. These approaches require that channel S-parameters be measured
starting from 40-50MHz with a step of about 10MHz. With such sampling, the low frequency
portion of the transfer function is completely ignored. The IFFT methods these tools employ
require uniform sampling, and the computational cost of using ~10KHz steps going up to
~50GHz is prohibitive.
We can name only a few works evaluating BLW. In [5] the authors describe an approximate
evaluation of BLW distribution, an important step in BLW analysis; [6] studies the effects of
BLW on signal integrity for PCIe Gen 3 protocols, [7-8] investigate the effect of BLW on bit
error ratio, and finds a short periodic test pattern that approximates its distribution.

Overall, there is no practically feasible systematic approach that can be used for BLW
simulation and measurement. In this work, we propose a fast method of bit-by-bit simulation
and a statistical BLW analysis. The latter assumes that the input pattern is either random and
uncorrelated, correlated with known correlation function or spectral density, or is a
deterministic periodic pattern. The method is majorly based on the idea of rational function
approximation of the BLW transfer function, making it possible to avoid IFFT and use
recursive convolution as a solution vehicle. Although is somewhat less convenient, it can also
be used in a more traditional way, based on step/pulse responses. In either case, we need two
convolution solvers, one for low-frequency BLW and the other for high-frequency ISI
effects.

IV. Looking Deeper into BLW. Is it a Part of ISI or

Something Else?

Despite the fact that phenomenon of baseline wander (BLW) has been known for years [2-7],
it somehow remains a mystery, not sufficiently covered in literature or addressed in analysis
tools. And yet, BLW is a serious factor: it produces slowly progressing low frequency noise
that can reach several mV by magnitude, sometimes comparable or exceeding crosstalk. It
affects bit error ratio, and is especially harmful for multilevel signaling protocols with smaller
level separation.
We know that BLW appears in SERDES links which do not have a conductive connection
between the signal transmitter and receiver, for example those using series DC-blocking
capacitors or transformers. We also know that BLW is a relatively slow process that depends
on both the channel’s transfer function at low frequency, and the spectrum of the digital
stimulus applied to the channel over considerable number of symbol intervals. In that respect,
BLW is closely related to inter-symbol interference (ISI), as both are manifestations of the
channel’s bandwidth limitations. If so, why can’t we consider BLW together with inter-
symbol interference (ISI) and use similar techniques for their simulation and measurement?
What is the difference between BLW and ISI?
To answer these questions, let’s consider channels with bandwidth limitations at low
frequency, high frequency, and both bandwidth boundaries. We understand that BLW is
caused by low-frequency bandwidth deficiency (LFD), while ISI results typically from high-
frequency bandwidth deficiency (HFD) or reflections. Figure 1 illustrates 3 cases: LFD,
HFD, and both of them.

Figure 1. Channel transfer function (top row), missing bandwidth (middle), and error transfer function (bottom
row). Left column corresponds to low-frequency deficiency, middle column – high-frequency bandwidth
deficiency causing ISI, and right column has them both

Figure 1 illustrates channel transfer function with low- and high frequency limitations (top
row). Middle row shows a missing portion of it, and bottom – an error transfer function. It is
important to recognize the scale on frequency axis (which is far from being linear), and
important frequencies and associated time constants. Let Fb be signal rate, and T  1 / Fb -
symbol interval. In Figure 1(a, H) there is a band between frequencies FH 1 and FH 2 - marked
blue - where most of the ISI effects reside. Typically, the length of the channel response
needed for ISI estimation varies from a few hundred to a thousand of unit intervals, hence
FB / FH 1 100 1000 . Since we should allow several samples per UI, the waveform
sampling rate, determined by FH 2 , should be higher by an order or two: FH 2 / FB 10 100 .
Low frequency bandwidth limitations span the range from FL1 to FL 2 , Figure 1(a, L). This is a
potential bandwidth of BLW error. Since FL 2 defines the upper LFD limit, it is related to the
time step that is small enough to capture all details of BLW noise. Typically, in channels with
DC blocking capacitors FL 2 doesn’t exceed a few MHz, which is well below the signal rate
FL 2 Fb . The lowest frequency of interest is FL1 ; this is where the channel’s transfer
function barely takes off. This frequency is in inverse proportion to LFD error response
duration, or the time needed for BLW effect from a particular symbol to disappear. If we
assume exponential shape of LFD response, represented with ~1% accuracy, it would take
about 400 uniform samples. However, it could be much larger with a more complicated
frequency response, e.g. having low frequency resonances. Roughly, we can assume that
FL 2 / FL1 100 1000 .
As we see from above, low-frequency LFD-related noise has its own time scale, non-
commensurable with simulation step and duration that we use for ISI. Sample interval for
BLW noise doesn’t have to be as small as for ISI. Depending on the ratio Fb / FL 2 1 , even a
symbol interval could be an unnecessarily fine granulation. If we have to simulate ISI
together with BLW, it would take around a billion solution steps only to pass clear of a single
BLW response duration. Therefore, it makes sense to separate computations of ISI and BLW
and use individual time steps for both.
Figure 2 illustrates pulse responses of the channels with LFD and HFD separately, and
combined. Its structure is similar to Figure 1. For convenience of drawing, the time constants
related to ISI and BLW are made much closer to each other. A channel that doesn’t modify
the DC component of the input signal preserves the integral of the input pulse, which is H 0T ,
see Figure 2, (a, H). The error, found as a difference between this pulse response and the
ideal pulse is shown in Figure 2, (c, H). Integral over the error response is zero. Although ISI
could be considerable, its accumulation doesn’t create a slowly changing bias component.
On the contrary, for the channels with a non-transparent DC component (cases L and LH) the
integral of the pulse response is zero, but not the integral on the error. The responses in the
right column may look similar to those in the middle, but notice a small negative tail of the
pulse response that stays long enough to zero-out the average. Respectively, the error
component has a negative tail making the integral equal  H 0T . Remarkably, this value
doesn’t depend on the low-frequency time constant, or exact shape of the channel response.

Figure 2. Pulse responses, characterizing the channels with LFD, HDF, and LDF & HFD. The structure is
similar to Figure 1
 What happens if we increase the value of the DC-blocking capacitor? The time constant of
the error component, as in Figure 3 (c, L) will increase. Its peak value gets smaller, however
its tail gets longer, as needed to preserve the value of the integral. As a result, BLW will be
accumulated from increasingly many symbols, its behavior becomes less predictable and less
correlated with the input pattern, at least within a time window we typically consider for ISI
analysis. Even though the energy of the error coming from a single pulse remains constant,
with more symbols involved, it’s less probable that their magnitudes will be “biased” in
concert, therefore the standard deviation of BLW decreases approximately as
T
 BLW  H 0 x . Here, we assumed that input pattern is random uncorrelated, symbol
2 LF
amplitude has zero average and standard deviation  x ; T is symbol interval, H 0 - error transfer
factor at DC, see Figure 1(b, L), and  LF is a large time constant associated with error transfer
function with response close to exponential.
A brief conclusion for this part:
1) Unlike ISI, BLW response is characterized by an error with a non-zero average. This
explains its accumulation over time if the input pattern is not DC-balanced.
2) The duration of a BLW response is orders of magnitude longer than that for ISI
effects, which makes it reasonable to compute them independently.
3) The, sampling interval needed to represent slowly changing BLW exceeds the symbol
interval if FL 2 Fb . In this case it seems reasonable to aggregate several consecutive
symbols together and use the “average” magnitude as an input to BLW filter.

V. Separating BLW transfer function from the full

channel characteristic

In order to efficiently simulate BLW, we should separate its transfer function from the rest of
the channel response. A proper way is to include the effect of linear equalization (FFE,
CTLE) as well, but we won’t consider it now because it doesn’t modify the procedure
described in this section.
One simple approach is first to find the characteristic by taking the samples on a coarse linear
grid, as needed for accurate ISI representation. For example, if the channel characteristic
from Figure 1 (a, LH) is sampled linearly with 10MHz spacing, starting from 50MHz, the
low-frequency (BLW) portion will be completely ignored. Then, we can resample low
frequency portion of the extracted response using a variety of interpolation methods (by
Fourier transform, spline interpolation, rational fitting, etc.). This will give us characteristic
behavior similar to Figure 1 (a, H). Then, we can find the characteristic of the same channel
with much finer granulation going from DC up to 50MHz, or whatever frequency is needed
to cover the effect from the blocking capacitor. The difference between the first characteristic
(its LF portion) and the second gives us “BLW error transfer function”, similar to the one
shown in Figure 1 (b, L).
VI. Representing BLW transfer function by IIR filter
Since the BLW response duration may reach millions of UIs, time domain simulation
methods using direct convolution (or its modifications, like overlap-save method) appear
impractical. This group of methods belongs to finite-impulse-response filters (FIR). Instead,
we’ll use an infinite impulse response filter (IIR) by finding a rational approximation of the
BLW transfer function [10, 11]. This approach is sufficiently accurate (max error below
0.1%), ensures causality, and allows fast recursive convolution, making it superior to direct
convolution.
An arbitrary BLW transfer function can be represented as sum of elementary frequency
components:
M
1  Am Am* 
H ( s)  H      * 
. (1)
m 1 2  1  s /  m 1  s / m 
Here, M is the number of poles; Am and m are real or complex factors and poles, with
Re{m }  0 . Then, for a given series of piece-wise constant (PWC) input - a sequence of
digital symbols of magnitudes xn - the output BLW becomes:
M
y (t n ) = H  xn +Re  zm,n , (2)
m 1

zm ,n = emhn zm,n1 + Am [1  e m hn ] xn 1 . (3)

Formulas (2, 3) give accurate solution in case of PWC input when the time points are taken at
its nodes. These are modifications of PWL formulas such as those from [12]. Two
simplifications are possible for BLW noise. First, its transfer function disappears at high
frequency thus making H   0 . Second, PWC node points are evenly spaced, therefore we
can set it to symbol interval hn  T . Then (2, 3) become:
M
y n =  Re{Km zm,n } (4)
m 1

zm ,n = Em zm ,n 1 + xn 1 , (5)
mT  mT
with constant complex factors K m  Am (1  e ) and Em  e . Complex zm are internal
state variables; they should be updated once per solution point.
Expressions (4, 5) evaluate the output BLW error once per UI at the pinnacle of the pulse
response and can be used as an efficient vehicle for time-domain BLW computations. We’ll
also use them to derive analytical form of the BLW noise pulse response - as in Figure 2 (c,
L) - for statistical analysis.

VII. Representing BLW transfer function by FIR filter

(direct convolution)

Once the channel response and BLW noise are separated in frequency, direct convolution
methods employing FIR filtering become possible, too. For example, we can use step or pulse
response representing BLW filter with much coarser time granulation compared to the signal
response. With two separate responses, and two convolution engines (for BLW and ISI
components) computations can be organized in a more efficient way.
 In fact, there is a simple relation between IIR equations (4, 5) and BLW time response. The
“cursors” of the BLW response on a single pulse, taken by symbol intervals, make a series
M
Pn   Re{Km Emn1} , n  1... N . Of course, with direct convolution, the length of this series
m 1
is always limited.
In this paper, we prefer using IIR filter described in the previous section. If necessary, it
allows us to combine BLW and ISI effects together in a single filter. It also provides a
convenient formal representation of responses, which we’ll use in the following sections.

VIII. Time domain analysis of BLW

Here we apply the techniques outlined in Section V and VI to the analysis of a FPGA design
with 4uF series capacitors, operating at 5 Gbps. The wideband transfer function (TF) has
been extracted from the design using a combination of logarithmic and linear spacing,
spanning the range from 10Hz to 50 GHz. The BLW noise TF (Figure 3, blue curve) was
found as a difference between the one with LF portion extrapolated down to DC (red curve),
and the full broadband TF (green).

(a) (b)
Figure 3. (a) Broadband transfer function, its version using flat continuation for the missing LF portion, and
BLW noise transfer function, the difference between the two; (b) sampled transfer function of BLW and its fit

We used a non-periodic random unconstrained NRZ pattern with 5 Gbps. The BLW transfer
function drops by 60dB at frequency that is about 3 orders of magnitude below the bit rate.
The time constant of the BLW components is two more orders of magnitude below that,
where the green and blue curves cross. Our expectation therefore is that BLW accumulates
the effect from approximately 100K symbols, but is only sensitive to symbol imparity
averaged over about 1,000 symbols and not to its faster variations.
The fitted version of the BLW function is represented by 7 real and complex poles. Poles and
residues were used to construct an IIR filter, as defined by equations (4, 5). We performed a
series of simulations updating state variables per equal time intervals. First, once per symbol,
with xn in (5) being actual symbol values (1, 1) . Then, the state variables were updated every
p symbols, with p = 10, 100…100K. For each group of p consecutive symbols we
determined the average and used it as an “aggregated symbol” when updating state variables
in (5).
Figure 4. (a) BLW computed over 10M bits with different granularity; (b) same waveforms, zoomed; (c) BLW
computed over 10B symbols with updates every 1K symbols; (d) BLW noise histogram build from 10B UI
waveform

As we see from Figure 4 (a, b), only extremely rare updates, once per 10K (marked green
star) or 100K symbols (marked with triangles), differ from those found with finer
granulation. This result agrees with our observations of frequency responses from Figure 3.
We found that updating state variables once per 1K symbols provides high simulation
accuracy of BLW, and is much faster than updating every symbol. Therefore, the next
experiment comprises of computing BLW over 10B symbols with 1K symbol updates, going
over 2 seconds of the “model time”. The resulted noise waveform and the histogram built
from it are shown in Figure 4, (c) and (d).
This approach is indeed very fast. On a Dell M4800 it takes about 1 sec of CPU time per
100M symbols. In this example, that’s only about 70 times slower than real-life hardware
operation.
It’s tempting to say that this fast time domain simulation of BLW provides a way to build
distribution down to low probability levels. Let’s imagine that we have a fast solver that can
simulate up to 1012 symbols in this manner, together with BLW noise. After we build the eye,
can we say that BLW noise distribution was accounted down to 1012 probability level?
Unfortunately, the answer is “no”. As we mentioned before, BLW and ISI effects have quite
different bandwidth limits. When we estimate histogram of a random process, only
uncorrelated data samples are statistically meaningful. Since correlation interval of BLW
process is about ~  LF / T of UIs, we need to increase the time of BLW estimation in the same
proportion. In our test example this ratio is 6 order of magnitudes! Whatever the time domain
simulation speed is, we cannot hope to cover 1018 symbols.
That’s why statistical analysis of BLW is extremely important.
IX. Statistical analysis of BLW
Let’s unwrap recursive convolution formulas (4, 5) and explicitly express the output BLW
samples through symbol values. By applying (5) repeatedly for N consecutive symbols, we get
the value of the m-th state variable:
zm,N  EmN 1 x1  EmN 2 x2  ...  Em xN 1  xN . (6)
M N
Therefore, the output becomes yN  Re  Km  EmN n xn . By changing the order of
m1 n 1
N M
summation, we can write it as yN   Pn xn , where Pn   Re{Km EmN n } . For the purpose of
n 1 m1
statistical analysis, we can assume that the property of the input pattern doesn’t change with
time inversion: it can be played backward and reveal the same statistical properties. If so,
symbol values xk can be replaced by x N k 1 and the output values represented as:
N M
yN   Pn xn , where Pn   Re{Km Emn1} . (7)
n 1 m 1

Random uncorrelated input pattern, even centered distribution

Let’s assume that input symbols have L-level modulation (PAM-L), with all states equally
probable, meaning they average to zero and are uncorrelated. Due to the considerable
difference between the signal rate and BLW processes, the low-pass filter will efficiently
average the effect from individual symbol over hundreds or thousands of bits, thus making
the distribution of BLW noise close to Gaussian.
We’ll estimate standard deviation of the output samples described by (7). A square of
standard deviation of y N is  y2  ( yN  y )2 . We assume that the mean value of a symbol
magnitude is zero, thus making the mean value of the output zero, too. Therefore, we need to
find an average of the product of the two sums:
 N  N 
 y2    Pn xn    Pq xq  . (8)
 n 1   q1 
Note that since xn are statistically independent, the average of the product xn xq is zero for all
n  q . Therefore, (8) reduces to a sum of components with squared factors only:
N N
 y2   Pn2 xn2   x2  Pn2 . (9)
n 1 n 1

Dispersion of the input symbols is defined by their largest magnitude As and the number of
signal levels as
L 1
 x2  As2 . (10)
3( L  1)
Hence the only remaining task is to find the sum in (9).
The summands in (9) don’t form a geometrical progression, as we’d have for an exponential
component; they may have components oscillating with different frequencies and attenuating
with different rate. Therefore, the sum (9) doesn’t converge monotonically with N. However,
since we know the slowest pole in approximation (1), we can easily estimate the number of
summands that guarantees convergence with predefined accuracy. Let  min  min Re{ m } .
For a given error   1 , we can estimate the number of summand from e  N minT   . In our
case min  2 *3.128e3 therefore for   1010 we should consider about 5.8M summands.
Figure 5 below, illustrates convergence of the sum while considering 10M summands. While
the number seems large, the computations are simple and take about a second. As we see, the
summands in (9) (red curve) start to decrease monotonically only for N>100K, after the faster
poles disappear. Their increments (green) are much smaller but follow the trend. The estimate
of sigma – a square root of the current sum (shown black) doesn’t visibly change after
N  5e5 . However 15 digits of sigma estimate settle after about N  3.2M symbols.

Figure 5 (a) convergence of the sum and BLW standard deviation; (b) comparison of BLW histogram found by
10B time-domain analysis and PDF predicted by statistical method

The resulting sigma is 1.228 mV, considering that for unit-magnitude NRZ  x2  1 . Found
statistical distribution perfectly matches the histogram from 10B-symbol BLW simulation.

Correlated input pattern

What if the input symbols are not statistically independent? If so, they are correlated, and the
average xn xnk is not zero even for k  0 . In this case, finding the average of the expression
(8) becomes more complicated. We assume that input pattern is a stationary process and its
statistical properties don’t change over time. Therefore, the average Cx ,k  xn xn k , Cx ,0   x2
could be non-zero but doesn’t change with n. Then dispersion of BLW becomes:
 N  N  N N N k
 y2    Pn xn    Pq xq    Pn2 xn2  2 Pn Pnk xn xnk 
 n 1   q1  n 1 k 1 n 1
N K
 N 
C x ,0  Pn2  2 C x ,k   Pn Pn k  . (11)
n 1 k 1  n 1 
Mathematically, (11) is a sum of point-to-point products of the discrete correlation functions
from both input pattern and the BLW pulse response. Factor 2 at the second sum means that
summation should be done for positive and negative indices, which however were combined
together due to symmetry. Upper index K in the second sum is a maximal correlation distance
between the symbols of the input pattern so that C x ,k  0 for k  K . In many practical cases K
is less than N , the length of the BLW pulse response in UI.
N
As we see from (11), it’s enough to find K sums  Pn Pn k and multiply them on the pattern
n 1
correlation coefficients, then add the result to dispersion of an uncorrelated pattern.
Having additional terms in (11) doesn’t mean that the resulted BLW noise increases.
Correlation between samples of the BLW pulse response is strong, and their products remain
positive in a wide range. However, correlation coefficients of the input pattern could be
negative. Let’s consider NRZ 8b10b input pattern applied to the same channel.

Figure 6. (a) Correlation function of 8b10b pattern, found from averaging over 10B symbols; (b) correlation
function of BLW pulse response; (c) Their product, illustrating summands of (11)

In Figure 6, the right side is a zoomed version of the left. Correlation function of 8b10b, (a),
was found by averaging the products xn xnk over 10B symbols. Correlation length, however,
is only a few symbols. Since 8b10b is DC balanced, the sum of the correlation coefficients is
zero. Correlation function of BLW pulse response (b) changes very slowly and remains
positive. By summing their sampled products we get  y2  1.7562e-10 [V2], or BLW noise
sigma equal 13.254 [uV]. This is about 2 orders of magnitude below the one for uncorrelated
pattern.

BLW caused by periodic test pattern (PRBS-N and others)

BLW noise is often modeled by applying periodic test patterns [9], which is simpler than
performing full time domain or statistical analysis. Here we show how to find the response to
such input.
As shown above, we can find m-th state variable after N consecutive symbols using (6).
However, our assumption was that all state variables were initially zero. Let’s consider two
cases.
If the period N of the test pattern is shorter than the length of the BLW response, we should
first initialize state variables for any n-th symbol within a period as
Sm,n  EmN 1 xn N 1  EmN 2 xn N  ...  Em xn1  xn . If the input pattern is N-periodic, state
variables become cyclostationary. Then, the stationary values for the same symbol within a
period become zm,n  Sm,n (1  EmN  Em2 N  Em3N ...) . Here, we used the fact that the contribution
  
Kp
from more distant periods is attenuated by an exponential multiplier EmN , where integer
K p shows how many periods ago they occurred. Recalling the sum of infinite geometric
S m ,n
progression, we get zm ,n  . Once the state variables have been found for any symbol in
1  EmN
a periodic pattern, there is no need to find partial sums for the other symbols. We can apply a
recursive formula zm ,n  Em zm ,n 1  xn 1 and find all other values in a period. Then, the
M
respective values of BLW noise are yn   Re{Km zm,n } .
m1

Next, if the period of the input pattern exceeds the length of BLW response, we can truncate
the summation when finding partial sum S m ,n . Also, there is no need in this case for
cyclostationary correction. Then, we’ll have to go through one full period with recursive
update, possibly using aggregation of symbols together, as we did in time domain analysis.

(a) (b)
Figure 7. (a) BLW noise caused by two periods of PRBS31Q (red) on top of BLW noise from random PAM-4
pattern (blue); (b) almost Gaussian histogram from random PAM-4 pattern (blue) and asymmetrical from
PRBS31Q (red)

We applied this method to find BLW noise caused by two periods of PRBS31Q pattern, for
the same channel and symbol interval, about 4B symbols total. The waveforms are shown in
Figure 7a. The response from PRBS31Q demonstrates abnormally larger undershoots, down
to -10.93 mV. Random PAM-4 produces nearly Gaussian distribution, with  x  0.9153 mV
(Figure 7b). It is smaller than that found statistically for NRZ input by factor 5 9 , which
agrees with (10). The effects of ISI and BLW were computed with separate solvers working
with different rates: BLW noise was updated once per 1K symbols using the “aggregated”
input.
SER plot and its cross-sections found from 4B symbol time domain simulation is shown in
Figure 8. The effect of BLW noise on the SER profile is consistent with its histogram in
Figure 7. Note however that the BLW noise shown in Figure 7 is what we miss in the eye
diagram when ignoring BLW. The BLW error has opposite polarity. That’s why large
negative undershoots in Figure 7 make the eye shrink in the positive direction.
 (a) (b) (c)
Figure 8. (a) SER built from PRBS31Q test pattern; (b) vertical cross-section of SER at xUI  0 with
uncorrelated input and BLW noise ignored (red), same with BLW considered (green), and with PRBS31Q and
BLW noise considered (blue); (c) is the same zoomed around the central eye

X. Important factors affecting BLW noise

The value of DC blocking capacitor

By reducing the value of the DC blocking capacitor we increase BLW cutoff frequency, and
hence, the power of the BLW noise. Both random uncorrelated and PRBS31Q patterns
behave very similar to Figure 7a, but demonstrate larger magnitude. Figure 9 shows the peak
values for random uncorrelated PAM-4 pattern and PRBS31Q. General trend agrees with
1
theoretical prediction ( ~ ), shown by dashed green. However, there is a certain
Cdc
deviation from this rule, especially for PRBS31Q. There are many reasons for that: BLW
response is not exactly exponential, the pattern is not random uncorrelated and its LF
spectrum may have low-frequency frequency resonances interacting with BLW response, etc.

Figure 9. Peak values of BLW noise (red – positive, blue – negative) for random PAM-4 and PRBS31Q

Transmit jitter
Transmit jitter modifies symbol length and therefore modulates the average running imparity.
For example, a DC balanced pattern affected by pulse interval modulation (PIM) may create
considerable BLW noise. The effect is hard to predict because it depends on BLW transfer
function and two independent processes: input pattern and its pulse interval modulation
caused by jitter. From a BLW perspective, random or deterministic jitter is especially harmful
when correlated with averaged imparity of the pattern. The “averaging window” can be as
small as one symbol or may be as long as the BLW response duration. Consider a simple
clock pattern as an example: one percent of duty cycle distortion results in a systematic BLW
component that is 2% of As H 0 , or the signal level. Irregular patterns create a more
complicated dependence of BLW noise on Tx jitter. Figure 9 illustrates the effect of duty
cycle and sine jitter on BLW for PRBS31Q and random PAM-4 inputs. As we see, jitter may
occasionally increase or decrease BLW noise over time.

(a) (b)
Figure 9. (a) BLW caused by PRBS31Q without jitter, with DCD and sine jitters; (b) same but for random
uncorrelated PAM-4 pattern

Channel equalization
Linear equalization (FFE and CTLE) works in two different ways. On one hand, it reduces
level separation by reducing signal transfer at low frequency. On the other, it lowers the
BLW transfer function approximately in the same proportion. The sum effect on BLW-
related error is not clear. Two other factors could be reducing ISI but increasing crosstalk.
Non-linear equalization (DFE) appears neutral to BLW.

Figure 10. Top row (a-c): no equalization, while BLW cutoff frequency is 100kHz, 1MHz, and 10MHz. Bottom
row (d-f): same as above by with identical FFE/CTLE equalization

We performed a series of simulations while increasing BLW cutoff frequency (hence, BLW
noise), with and without linear equalization. As we see from Figure 10, the eye diagrams
becomes progressively closed (left to right), due to BLW noise increase. Equalization
 improves the eye, but doesn’t cope with BLW noise. On the other hand, equalization doesn’t
increase BLW noise, as we see from comparing the two plots on the right.

XI. Correlations with measurement results

Experiment Design
The purspose of the measurements is to correlate the effects of BLW on the global
performance of the PAM4 signaling from the point of of the receiver. Effects of BLW are
examined in the context of global attributes that qualify PAM4 signaling from the receiver
angle.
The following industry standard qualifiers are assessed:
1. BER qualified eye openings, horizontal and vertical for all 3 PAM4 eyes.
2. Data Dependent Noise (DDN)
3. Signal levels
4. Level linearity of PAM4 signals
5. Eye closure penalty

Test Patterns
The ideal test pattern would be a PRBS31 that stresses most the BWL effects. The reality is
that no acquisition system can acquire such long patterns with enough resolution, so instead
the SSPRQ (Short Stress Pattern Random Quaternary) test pattern, with length set to
PRBS16, 2^16-1 precisely, is used. The SSPRQ is comprised of 4 segments, each based on
key stressors from PRBS31. It is a stressful pattern, but short enough to allow for advanced
analysis of jitter, noise, and BER with Equalization.
To compare results with less stressful patterns, measurements are performed with PRBS15
and PRBS13, which have the benefit of higher throughput acquisition, but effective in
assessing the intersymbol interference due to limited bandwdith.

Data Source
In order to isolate the effects of BWL from bandwidth limitations of the target DUT, we
chose to perform the measurements at 26.5625Gbaud, PAM4 data. The PAM4 pulse
modulation was selected because of the critical loss of 9dB signal-to-noise ratio when
moving from NRZ to PAM4.

The values of DC blocking capacitor

The acquisition and analysis was performed on a data stream that included a series of DC
blocking capacitors between the signal generator and the front end of the acquisition module.
Testing was performed with:
• No DC blocking capacitor inserted
 • 30nF, 220nF, 1uF capacitors inserted in the signal path.

Signal Source and Acquisition System

Signal source is a PAM4 pattern generator that provides all test pattern configurations.
The acquisition system is a 60GHz bandwidth remote head module.

Eye Measurements
The PAM4 Eye openings, horizontal and vertical are evaluated at the receiver slicer position,
as specified by IEEE standard. Since PAM4 signaling is typically using Forward Error
Correction (FEC), the target BER for these Eye measurements is 5e-5.

Figure 11 PDF Eye Modeling Receiver Slicer

In following tables record the results of global jitter, noise and eye measurements for no
blocking capacitor, and 3 different DC blocking capacitors.
Figure 12 No DC blocking capacitor

Figure 13 DC blocking capacitor – 30nF

Figure 14 DC blocking capacitor – 220nF

Figure 15 DC blocking capacitor – 1uF

By inserting a 3nF capacitor the horizontal and vertical eye openings at 5e-5 BER drop as
much as by a factor of 2.
Total Noise increases due to BLW from about 30mV to 39mV for the top PAM4 eye. The
plot below is a validation of the plot in figure 9, showing noise increase in function of DC
block capacitor value. This is a reverse, showing the Vertical Eye opening in function of the
DC block value.

Figure 16 Vertical Eye opening function of DC block.

A deeper look into the noise analysis for the acquired data, shows clearly that, as expected, the contribution to
the increase of Total Noise is due to the Data Dependent Noise, which points to the BLW as the cause for the
increase of the deterministic noise.

Figure 17 Data Dependent Noise function of DC block.

Global PAM4 measurements: Linearity Mismatch, Vertical Eye Closure, Level

Deviations
IEEE and OIF Standards have defined a number of global PAM4 measurements to assess the
distortion of different sources, and that drive the overall BER performance of a serial link.
The question to answer is which of these measurements are affected by the BLW.
The Level Mismatch is a measure of the linearity, defined as the level deviation from an ideal
level distribution. Ideal value for Level Mismatch Ratio is 1. As shown below, and as
expected, the DC coupling capacitor does not affect the level symmetry.
Figure 18 Level Mismatch ratio function of DC block.

A third measurement to consider is Vertical Eye Closure penalty (VECP). Similar with the
TDECQ (Transmitter and Dispersion Eye Closure Quaternary) used by the optical system,
VECP identifies the smaller eye opening relative to an ideal given a signal amplitude. Ideal
value is ~5dB for 3 equal PAM4 eyes. Correlating with the increase in the noise due to BLW,
VECP will penalize the receiver.

Figure 19 Vertical Eye Closure Penalty function of DC block.

Relevance of Test Patterns

In order to isolate the effects of Baseline Wander from effects of Intersymbol Interference
and the effects of uncorrelated jitter and noise impairements, the testing was performed by
using a series of standard PRBS patterns along the testing with the SSPRQ pattern.
The relevance of using the long runs for each symbol level is illustrated in figure 20, which
shows the vertical eye measurement results function of DC block for SSPRQ and PRBS13.
While SSPRQ was designed for compliance testing for optical links, it proves to be the
required test pattern for isolating the effects of BLW.
XII. Conclusion

In this paper we considered Baseline Wander, an important impairment that should be

considered in SERDES design process. We analyzed the cause of BLW, its difference from
ISI, and proposed its analysis by using two independent characteristics of a channel, at low
and high frequency. This can also be performed using two step responses which cover
different time regions and complement each other. A separate low-frequency BLW
characterization allows its fast and accurate analysis, both in time and statistical domains. We
considered a number of input pattern types: random uncorrelated, non-deterministic and
correlated with known correlation function, and deterministic periodic.
With multi-level modulation, and tighter signal margins in state of the art designs, BLW can
no longer be ignored. We believe that the methods proposed in this paper would benefit
different simulation and measurement approaches, including StatEye-type, IBIS-AMI,
COM/PCIE/USB/JCOM compliance methodology, as well as BLW measurement techniques.

References

[1] IEEE Standard for Ethernet, IEEE Std. 802.3bj-2014, Annex 93A
[2] A. P. Clark, “Principles of digital data transmission”, Willey, 1983
[3] S. Haykin, “Digital communications”, John Wiley & Sons, 1988
[4] J. Proakis, “Digital communications”, 4th edition, McGraw-Hill, 2000
[5] F. Waldhauer, “Quantized feedback in an experimental 280 Mb/s digital repeater for
coaxial transmission”, IEEE Trans. on Communication, v.22, Jan. 1974
[6] N. Sommer, L. Lusky, M. Miller, “Analysis of the probability distribution of the
baseline wander effect for baseband PAM transmission with application to gigabit
Ethernet”, Proc. of the 2004 11th IEEE International Conference on Electronics,
Circuits and Systems, ICECS 2004.
[7] N. Na, D. Dreps, J. Hejase, “DC wander effect of DC blocking capacitors on PCI
Gen3 signal integrity”, 2013 IEEE 63rd Electronic Components and Technology
Conference.
[8] P. Anslow, “Baseline wander with FEC”, IEEE P802.3bs Task Force, 2017
[9] P. Anslow, “SSPRQ test pattern”, IEEE P802.3bs Task Force, 2016
[10] B. Gustavsen, A. Semlyen, “Rational approximation of frequency domain responses
by vector fitting”, IEEE Trans. Power Delivery, 1999, vol. 14
[11] B. Gustavsen, C. Heitz, “Modal vector fitting: a tool for generating rational models of
high accuracy with arbitrary terminal conditions”, IEEE Trans. Adv. Packag., 2008,
vol. 31
[12] V. Dmitriev-Zdorov, V. Heyfitch, G. Pratt, “Fast time-domain simulation of 200+
port S-parameter models”, DesignCon 2006.