DesignCon 2012

Salz SNR Modification for High Speed Serial

Link System Performance Estimation

Henry Wong, Gennum Corp.

Henry.Wong@gennum.com

Xiaoqing Dong, Huawei Technology

Dongxiaoqing82@huawei.com

Francois Tremblay, Gennum Corp.,

Francois.Tremblay@gennum.com

Geoffrey Zhang, Huawei Technology

Geoff.Zhang@huawei.com
Abstract
Salz SNR has been widely adopted as a mathematical SNR figure-of-merit, over bandwidth-
limited channels, over the years. Its application to 10/40/100G Ethernet, over electrical
backplane and copper channels, is a direct result of its use of the MMSE-DFE criterion that
many today’s SerDes devices commonly share. However, this theoretical SNR maintains to have
a gap, from a practical link SNR standpoint. This contribution is to develop a methodology to
narrow this gap, so that the modified Salz SNR has the following main advantage: Provide a
more reliable SNR prediction for better link budgeting, targeted for each channel and each
SerDes configuration.

The proposed methodology is built on the original Salz SNR fundamentals, but addresses each
of its original assumptions one-by-one, through the calculation of a list of power penalties. We
use today backplane channels, based on various state-of-the-art SerDes configurations, to
demonstrate the practicality of such needed SNR modification.

Biography
Henry Wong is a Principal Engineer at Gennum Corp. He has over 17 years of industrial
experience in transceiver design ranging from applications in high-speed backplane, multi-mode
fiber, to ADSL, cable, and wireless broadband. Prior to joining Gennum, he has worked as a
technical leader in Nortel, Cadence, and a series of start-ups. In the past 7 years with Gennum
he focuses on high-speed CDR and equalizers design, system architecture, and system
modeling. He holds B.A.Sc. and Ph.D. degrees both in Electrical Engineering, and has offered
graduate courses and seminars in universities.

Xiaoqing Dong joined Huawei Technologies in 2006 as a signal integrity research engineer,
where she works on high speed active SI simulation and measurement technology. She received
her bachelor and master degrees in communications and information system from Harbin
Institute of Technology, China.

Francois Tremblay is Director of Design Engineering for Gennum Corp. responsible for the
development of 8-25Gbps module retimer and backplane products. Prior to Gennum, Francois
held positions from manager to VP at various semiconductor companies such as Nortel,
Cadence, Catena Networks, Ciena and LTRIM technology. His expertise in communications
ranges from Discrete Multitones modulation on twisted pair to QAM on cable modem to
28Gbps NRZ for module and backplane applications. Francois holds a BSc degree in Electrical
Engineering from the University of Ottawa.

Geoffrey Zhang has been with Huawei Technologies since 2009. He is currently the CTO for the
Interconnect Department. Prior to joining Huawei he worked for Texas Instruments, Lucent
Technologies, Agere Systems, and LSI Corporation. He has worked on various products including
data converters, timing devices, read channel chips, and SerDes related applications. He
received his Bachelor and Master degrees both from the Department of Electrical Engineering,
Zhejiang University, China, and Ph.D. in microwave engineering and signal processing from Iowa
State University.
1. Introduction
In today’s standardizations for 40G/100G electrical backplane (BP) and copper cable applications, Salz
SNR [1] is often used as a means to calculate end-to-end transmission SNR margin for link-budgeting
purpose. Salz SNR is also used in 8G Fiber Channel [2] as one methodology for signal quality
specification. And in the past, Salz SNR was also often applied in IEEE 10GBase-T Ethernet [3] for SNR
margin study. It is seen that this SNR figure-of-merit has a popular usage in the past, in today, and likely
continues to be in the future, for calculating a theoretical SNR value to help us gauge transmission link
margin.

Such popularity can be clearly understood from:

a) Salz SNR is derived originally for a bandwidth-limited telephony channel (300 Hz to 3.4 kHz) for
data transmission over voice band. And today, although the data rate is drastically different –
10Gbps to 25Gbps per channel, the commonality of having bandwidth limitation issue still
exists.

b) Salz SNR is derived as optimal, when the receiver is configured with a linear filter (feedforward)
and a feedback filter (decision-feedback equalizer). And by adjusting these filters, the decision-
slicer error is minimized in a minimum-mean-squared-error (MMSE) sense. This same
configuration that Salz has used is largely a common configuration, in today’s equalizers for
high-speed applications.

This Salz’s MMSE-DFE criterion is highly applicable in today’s applications.

Despite this, it is well recognized that Salz SNR is only a theoretical SNR, and there exists a gap to bridge
from such theoretical value, to a more practical SNR value, for its successful use in today’s SerDes
(serializer/de-serializer) link – under practical transmitter/channel/receiver, that has limited number of
taps, limited receiver sensitivity, presence of colored noise, etc. If this gap is not addressed properly,
inaccurate margin allocation results, leading to an impact on link design.

There are other useful techniques as proposed recently, so that SNR at slicer can be derived from a
different manner. Our current contribution, however, is different, in a way that, it is built on the
fundamental theoretical basis of Salz SNR, as a starting point. We then tackle Salz’s mathematical
assumptions one by one, so that each power penalty is built-up, to address each specific channel
response, and each target equalizer(s) configuration. In short, we propose a method to convert from a
theoretical mathematical figure-of-merit, to a more practical SNR figure-of-merit.

2. Salz SNR
Salz SNR is derived for bandwidth-limited channels based on the use of the MMSE (minimum mean-
squared error) criterion. The problem is to minimize the decision-slicing error ek , in a mean-squared
error sense, as below:
 MSE  E  vk  aˆk  
2
Eq. 1
 

where vk is an input to the slicer, and aˆ k is the slicer result, i.e., a slicing estimate of the received
symbol. In this section, we attempt to follow [1] in all mathematical symbols to avoid confusion.

The block diagram of the signal model in Salz SNR is illustrated as follows:

Figure 1. Block diagram of Salz SNR signal model.

where S (t ) is for transmit-pulse shaping (e.g., rectangular pulse convolves with a pulse-shaping filter),
h(t ) represents the channel, w(t ) is a continuous-time feedforward filter, and bn is the causal
feedback filter.

There are four main components in the Salz’s model in carrying out RX equalization:

1. Matched filter
2. Anti-causal feedforward filter
3. A baud-rate UI-interval sampler
4. Feedback filter

In this MMSE derivation, it was found that the optimal feedforward filter w(t ) is represented by a
composition of a matched filter f (t ) and an anti-causal feedforward filter g (t ) . The matched filter is to
match its impulse response, to the pulse response of the channel after convolving with the TX pulse-
shaping filter, in the following way [4]:

 (t )  S (t )  h(t )
Eq. 2
f (t )   (t )

where  (t ) = pulse response after the channel, and f (t ) is the matched filter to match with this pulse
response.

In frequency-domain representation, and by taking the UI-sampler into account, the TX pulse shaping,
channel, matched filtering, and UI-sampler, has this following representation. Here, at this moment, we
have not taken the anti-causal filter into account. We obtain [1, 4]:
 2
1   2 n 
S pp ( )   P   
2
 Eq. 3
T n  T 

where T = symbol period, and the magnitude-squared spectrum is due to the frequency-domain
representation of the convolution of  (t ) and  (t ) in matched filtering. The summation of this
magnitude-squared spectrum of P( ) is due to baud-rate sampling. The aliased spectrum is kept added
to the original magnitude-squared spectrum because of discrete-time sampling of a continuous-time
signal.

The matched filter, although matching to the RX pulse shape, may not imply there is no ISI. The anti-
causal filter is to then cancel the precursor ISI of the convoluted pulse response after matched filtering.
The way of precursor cancellation is not zero-forcing, but rather balancing out with any noise
amplification by this feedforward filter, so that the error deviation from the slicing threshold is
minimized in an MSE (mean-squared error) sense.

The MSE at the slicer error, due to the use of this formulation, was found to be [1]:

 T  T 
MSE   a2 exp   ln Y ( )  1 d  Eq. 4
 2  T

where  a2 is the variance of signal or average signal energy. Y ( ) is defined as

2
1 
 2 n 
Y ( )  '
NoT

n 
P  
 T 
 Eq. 5

where

No
N o'  Eq. 6
2 a2

and N o 2 , the variance of RX input noise, is the double-sided power spectral density at Rx input,
assumed white and Gaussian.

Re-arranging Eq. 4, we can obtain the Salz SNR as:

 2  T  T 
Salz SNR   a   exp   T ln  Y ( )  1 d   Eq. 7
 MSE   2 
where Y ( ) has been defined in Eq. 5.
3. MMSE-DFE SNR
There has been a good amount of quality research in the past on MMSE-DFE. One focused and unified
research effort is due to Cioffi, Dudevoir, Eyuboglu, and Forney [4]. One main conclusion was that
MMSE-DFE, as derived based on linear-estimation theory, arrives at the same result as what Salz SNR
represents.

Our purpose of focusing on this MMSE-DFE research development, in this Salz SNR modification context,
is that such formulation of MMSE-DFE helps us bring the concept of SNR, and then the concept of
modified SNR, in a more straight-forward manner. Note that, such MMSE-DFE technique has been used
by a variety of others for performance evaluations in different disciplines of digital wireline and wireless
communications. Here, we formally introduce its use as a way to modify the ideal Salz SNR. And then,
we make use of examples to show its use, and compare with the ideal Salz SNR results.

The signal model of the MMSE-DFE has been described in [4]. By using the same mathematical variables
as what the reference has used, we have:

Figure 2. Feedfoward and feedback equalization satisfying MMSE-DFE criterion

The signal model in Figure 2 has largely the same structure as the model has in Figure 1, but they differ
in a couple of main areas. In the following we discuss the consistency and differences:

1. Both models have the same matched filter to match with the pulse response arises from the
transmit filter and the channel
2. Both models have the same decision slicer, and feedback filter
3. The model in Figure 2, however, has the sampler immediately after the matched filter
4. Because of (3), the model in Figure 2 addresses the feedforward filter as a discrete-time filter

The motivation of (3) and therefore (4) above is that the MMSE-DFE derivation as treated in [4] assumes
the use of a multi-tap FFE. Such treatment does not impact the main theme of the paper, and so for
convenience we follow such placement of the sampler. Note that, the sampler can be a fractional-UI
sampler, e.g., by means of using a fractional-spaced FFE. In the following, for simplicity, we assume the
use of a baud-rate UI sampler, and, therefore, assume a baud-rate discrete-time feedforward filter.
The MMSE-DFE in general has the same error definition as in Salz SNR, and therefore has the same
MMSE mathematical criterion. By assuming all decisions are correct decisions [1, 4], we get xk  xˆk ,
leading to

ek  xk  zk
Eq. 8
2  min E[ xk  zk ]
2
MMSE  DFE

Following [4, 5], the front-end of the signal model can be interpreted mathematically as a MS-WMF
(mean-squared, whitening matched filter). It means that, the error sequence at the slicer, is a whitened
sequence after feedforward and feedback filtering (can be of infinite durations), and, as a result, the
formulation can be conveniently analyzed as an approximation of an AWGN channel model. But of
course, the white noise sequence here is not necessarily Gaussian, but rather depends on the probability
distribution of the input signal.

The MMSE-DFE decision-slicer SNR is:

1   Sx 
  log  N
2
log (MMSE-DFE SNR)  Sqq ( )  1 d Eq. 9
2 
o 

where the logarithm can have any common base. S x is the average signal energy, and

  2 m 
2
1 
Sqq ( )   Q
2
Eq. 10
T m 2 T

where Eqs. (10) and (3) have the same meaning, only a change of variables, representing the magnitude-
squared spectrum due to matched filtering. The summation again is due to the discrete-time sampling of
a continuous-time signal at the UI sampler.

One can find the equivalence of Eqs. (9), (10), with Eqs. (3), (4), (5), (6), indicating that the MMSE-DFE
SNR is indeed equivalent to the Salz SNR.

4. Modified SNR for Practical SerDes Devices

Our purpose is to derive SNR estimation at the same decision slicer that we discussed in the previous
sections, but take the practical limits into account, under the current backplane applications
environment. The practical limits refer to the implementation limitation of current start-of-the-art
SerDes devices. In this paper, we target on equalization, but the general concept is applicable to other
non-idealities as we will show.

When we make such SNR modification in Section 5, we refer heavily to the MMSE-DFE SNR concept, and
its mathematical treatment. Without loss of generality, we understand that this modification is made in
the same way as to the Salz SNR, but based on the concept as used in MMSE-DFE. We then later on,
once again, refer to this MMSE-DFE treatment in simulations, in order to demonstrate with examples.

Here is a list of practical limitations of current SerDes devices:

a) Matched filter may not exist in the RX front end

In some implementations, RX filter in the front end may only serve as a bandwidth-limitation
filter. Even when Nyquist filter (i.e., filter that satisfies the Nyquist Criterion) is used as an RX
filter, the channel is not known a priori, as such, the RX Nyquist filter can only match with the TX
Nyquist filter, such as the use of square-root raised-cosine in both ends. In this case, the RX filter
does not match with the convolution of the TX filter and the channel. Because of this, the
matched filtering that are discussed in Eqs. (2) and (3) are not commonly used in practice.
Therefore, the matched-filter SNR bound [4, 5] is difficult to achieve, even in the absence of
channel ISI.

b) Feedforward filter is of finite length, and may not be entirely anti-causal

The feedforward filter resulting from Salz’s MMSE derivation in [1], and the same feedforward
filter resulting from MMSE-DFE derivation in [4] come to the same conclusion that the optimal
feedforward filter is a combination of a matched filter and an anti-causal (one-sided) filter.
In practice, the feedforward filter may not be entirely responsible for precursor cancellation.
E.g., an CTLE (continuous-time linear equalizer) cancels both precursor and postcursor, but not
necessarily focus on precursor cancellation. If an FFE is used instead, the main tap may also not
be placed as the last tap in the FFE tapped-delay line, so that the FFE can shorten the postcursor
ISI tail, in order not to use too many extra DFE taps. The result of this is that, there can be a bias
toward more noise amplification, and less precursor ISI cancellation. As such, any residual
precursor ISI needs to be part of the degradation, of the MMSE slicer SNR calculation.

c) Feedback filter is of finite length

The feedback filter is certainly causal as used in Salz SNR [1] and in MMSE-DFE [4], as well as in
the implementation. But the length in practical limitation is limited and cannot cancel all the far
reflections and the long postcursor ISI tail as often encountered in backplane channels. As such,
any residual ISI needs to be part of the degradation, of the MMSE slicer SNR calculation.

d) Consideration of RX sensitivity

The SNR as calculated in the original ideal Salz SNR equation is a calculation of signal-to-noise
ratio, with a consideration of the magnitude spectrum of the matched filtering being matched to
the channel and TX. However, only a ratio is considered, not the absolute RX voltage level.
Because of this, RX sensitivity of typical SerDes devices is not taken into account.

There is a good way to account for such and that is putting an AGC in the RX front end as a
 common practice. The AGC can be a stand-alone block or integrated with the feedforward filter.
In any case, there is a penalty associated with this AGC gain which amplifies the input voltage
noise, being random noise and/or crosstalk noise. As such, this noise penalty needs to be part of
the degradation, of the MMSE slicer SNR calculation.

e) DFE error propagation

The assumption of having correct decisions, to feedback to the filter, to generate the correct
postcursor ISI cancellation, is made in the Salz SNR derivation, and in the same way as in the
MMSE-DFE derivation. If this assumption is not made, the derivation becomes mathematically
complex and cumbersome. In practice, depending on the feedforward filtering, and the channel
of concern, the DFE taps can be of large magnitude and require the use of many taps in a row.
The probability of error propagation therefore increases. As such, one can allocate a penalty
number as part of the degradation, of the MMSE slicer SNR calculation, perhaps with statistical
approximation based on the postcursor strength on a case-by-case basis.

5. Approaches to SNR Modification

Section 4 has discussed several practical limitations that the SerDes devices experience, when compared
to the idealized Salz SNR conditions. In order to account for these limitations, we calculate power
penalty for each limitation, from which we factor these power penalties into the SNR calculation. Such
degradation of the ideal Salz SNR can give a more realistic assessment of today’s SerDes devices
performance.

Authors in the past have calculated various kinds of power penalty and used the penalty to degrade the
ideal SNR, as discussed in [9, 10, 11, 12, 13, 14]. Therefore, the concept of SNR modification by using
power penalty is not new. What we demonstrate here is one formal way to account for penalty. And
then, in Section 6, we will demonstrate the use of these penalties to backplane channels. We then
compare with Salz SNR, from which to indicate what improvements this proposed technique has made,
in terms of predicting the practical BP channel performance.

We start with the concept of slicer SNR:

The decision-slicer SNR is defined to be the ratio of average signal energy to the mean-square of slicing
error, as

E  xk2 
SNR  Eq. 11
E  xk  zk 
2
 
based on the input and output as defined Figure 2 and in Eq. 8.
The numerator represents the strength of signal. Such strength in a typical receiver depends on the AGC
gain and the feedforward filter gain. The denominator typically represents a deviation from the signal
nominal amplitude. The deviation, in the ideal Salz SNR formulation, is due to both the residual
precursor ISI and any noise, that may exist after the equalization. Apparently, the denominator
represents a variance of noise, and the numerator can be said as the varince of signal.

In non-ideal situations, as discussed in Section 3, there are a variety of “noise” variances that can make
the signal at the slicer deviate from the slicing threshold. This can be represented as a series of variances
to sum, and is then represented as a total of “noise” variances. Some of the variances of particualr
concern are:

1. Residual precusor ISI variance,  pre

2
isi

 pre isi   x  qm
2 2
Eq. 12
m 0

where qm is the mth precursor ISI cursor, and  x is the signal energy.

In practice, the precursor ISI can be handled by a variety of filtering, through:

a. The precursor taps of Tx deemphasis filter

b. The precursor taps of the RX FFE
c. The phase & magnitude responses of the RX CTLE

Unless zero-forcing is the criterion, which is not our focus here, there will be remaining ISI due
to balancing noise amplification and ISI reduction in an MMSE sense. In addition, usage of
limited number of taps or limited number of poles and zeros will leave some ISI uncanceled
inevitably. In some cases, in practice, the feedforward filter, e.g., an FFE, is not entirely anti-
causal, so that the postcursor tap can shorten the tail, to help reduce practical DFE size. In doing
so, this can also lead to more amplfication of noise, which we will address in the following
subsections.

2. Residual postcursor ISI variance,  post

2
isi

 post isi   x  qm
2 2
Eq. 13
m0

where qm is the mth postcursor ISI cursor, and  x is the signal energy.

Similar to the precursor ISI, the postcursor ISI can be handled by a variety of filtering, through:
a. The postcursor taps of Tx deemphasis filter
b. The postcursor taps of the RX FFE
c. The phase & magnitude responses of the RX CTLE
 d. The filter taps of the RX DFE

As opposed to the ideal MMSE-DFE formulation, where number of taps is of infinite duration,
under finte-length equalization, residsual postcursor tail or far reflections remain uncanceled.
This penalty can be substantial under long BP channel or under highly reflective channels. The
variance of this can be significant.

3. Crosstalk noise after the feedforward filter gain G

 xtalk
2
 slicer   xtalk G feedfwd
2
Eq. 14

Crosstalk noise is non-white. And most likely far from Gaussian if we focus only on one or two
aggressors. And in the past, the formulation of crostalk noise into a power penalty usually
receives some level of averaging [2, 3]. Here, we also follow such power-averaging calculation,
by calculating the total variance of the RX input crosstalk noise, where such input crosstalk noise
is expressed as a power sum.

The RX crosstalk noise is then altered by an AGC and the feedforward filter, before the noise
reaches the decision slicer. In some implementations, the AGC is embedded in the feedforward
filter. In any case, there will be a power gain (or attenuation) depending on the system design.
For convenience, the feedforward filter discussed here includes an AGC, so the gain that is
required to bring the signal level to the slicer threshold, also applies to the crosstalk noise. A
drawback of this approach to calculate crosstalk noise power penalty is that, the frequency
variation of crosstalk noise, before and after the feedforward filtering, that may impact the
signal in different ways, is not revealed. However, when we compare with the Salz SNR
definition, it is also a calculation of power-spectrum averaging. We do not deviate from this
aspect.

4. Random noise after the feedforward filter gain

 rn2 slicer   rn2 G feedfwd Eq. 15

A very similar treatment to the crosstalk noise, the average random noise power is obtained at
the RX input, and then modified by the feedforward filter gain. The random noise is assumed
white and Gaussian.

One important aspect of this penalty is that, one can also interpret it as an equalization penalty.
More specifically, it is related to the insertion loss of the channel, that the equalizer needs to
compensate for, in order to bring the average signal energy up, to match with the slicing
threshold. For example, when the insertion loss is high, it causes received signal energy small,
that the AGC in the feedforward path needs to amplify the signal, such that the signal energy is
 restored to be compatible with the slicing threshold. This amplification of signal for insertion-
loss equalization, therefore becomes a penalty for noise amplification. Although one can
interpret this as an equalization penalty, we make it specific here to name this noise penalty.

5. Other degradations

There are apparently other penalties such as DFE error propagation, crosstalk noise frequency-
dependency, CDR timing jitter, DCD, non-linearity, offset, etc., that need to be considered in
order to impair the SNR further to better reflect the link margin. One can formulate “noise”
variance for each, but this requires some sort of statistical significance in this formulation. In this
paper, we focus on the 4 penalties that we just discussed.

By considering now the four penalties that we just discuss, Eq. 8 can be modified as:

E  xk2 
SNR  Eq. 16
 2
pre isi   post
2
isi   xtalk  slicer   rn  slicer 
2 2

Here the denominator shows a sum of variances in an RMS manner, assuming no correlation among
these disturbances.

6. Relating Salz SNR Modification to Finite-Length MMSE-DFE

At this point we come to a convergence. Before we discuss the convergence point, let’s step back to
sequence our approach so as to help understand this approach of SNR modification in a clearer way.

We start describing the Salz SNR definition of an ideal system. Under the MMSE criterion, the optimal
feedforward filter was found to be a matched filter followed by an anti-causal filter. The structure also
includes a baud-rate sampler together with an infinite-length DFE, coupled with a decision slicer. Based
on the MMSE-DFE research by [4] it was found that this is equivalent to the MMSE-DFE solution. The
MMSE-DFE solution starts with the slicer-SNR definition as given in Eq. 11. We make use of this SNR
definition as our basis for Salz SNR modification. In the modification, we took the approach of
generating a list of power penalties. The power penalties, although not specified explicitly in Section 4,
involve both time- and frequency-domain calculations. Example of this calculation will be shown in
Section 7. Following this, we can easily see that, the power penalties that we consider are due to finite-
length effect of the equalizer, and the noise amplification of the equalizer. Such effects fit nicely into the
finite-length MMSE-DFE framework [6, 7, 8]. Because of this, this framework is used in our calculations.

The MMSE-optimal feedforward and feedback filter taps, assuming the use of FFE and DFE as equalizer,
can be calculated as:
 1
  1 
w  1 P  PP *  PJ  J * P * 

*
Rnn 
 SNR 
Eq. 17
1
  1 
b  1 P  PP *  PJ  J * P * 

*
Rnn  PJ 
 SNR 

where w = feedforward FFE taps, b = feedback DFE taps, SNR = RX input SNR. Also, 1* represents a
vector of ones and zeros for positioning, * represents matrix transpose (as matrix is real-valued here for
NRZ), P = channel Toeplitz matrix, which is obtained from the pulse response, due to a convolution of
the TX pulse response with the channel impulse response.

Making use the structure above and the power penalties we listed, we calculate the Modified SNR in this
way:

1. Calculate RX input SNR

a. Based on RX signal energy, crosstalk noise power, random noise power

2. Calculate required AGC gain, finite feedforward filter, and finite feedback filter
a. Based on MMSE criterion
b. Based on RX signal energy and required slicing threshold, RX input SNR

3. Calculate power penalties

a. Based on residual precursor ISI, residual postcursor ISI, both in time domain
b. Based on AGC gain, and RX input noise power, in frequency domain

4. Calculate Modified SNR

a. From the results above

Section 7 gives this calculation by using examples. We then compare with the Salz SNR to show the
degradation. Subsequently, we show the simulation results from a collection of SerDes devices, to
demonstrate that the modified SNR indeed reflect the performance appropriately, because of the
inclusion of the penalties. Whereas when the penalties are not included, ideal SNR can opposite
predictions of performance.

7. Examples and Comparisons with SerDes Simulation Results

a. Calculate Modified SNR for 2 channels

We use an example in this section to illustrate the Modified SNR results. Figure 3 below shows 2
backplane channels. Case 1 has somewhat larger loss than Case 2 has, but Case 1 is somewhat less
reflective than that in Case 2. Using data rate = 8.5 Gbps as an example, both channels are very similar
and the purpose of using this example is to show how the penalty can be accounted for and be
expressed in the modified SNR values, to demonstrate their difference, even when the 2 channels are
very close.

A small averaged noise power representing the actual crosstalk-noise power sum is used in the following
calculation. A small noise power is used in this example so that it does not dominate the total noise
variance in Eq. 16. However, the same calculation can apply when the averaged noise power is large.

Figure 3. Two similar channels with slight differences in loss and reflections. Top: insertion loss and crosstalk; bottom: pulse
responses; left: Case 1 (lager-loss, less-reflective channel); right: Case 2 (smaller-loss, more-reflective channel).

In the derivation below we assume that the feedforward filer is just an FFE equalizer, not a matched
filter, and is calculated based on the finite-length MMSE structure as described in Section 6. In
calculating the FFE taps, one practical tap arrangement is to use not exactly an anti-causal filter, but
leave one postcursor tap to help shorten the postcursor ISI tail, to help relieve the requirement of the
DFE. The following shows:

1. An example of 5-tap FFE being applied to Cases 1 and 2

2. 3 precursor taps, 1 main tap, and 1 postcursor tap
3. UI-sampled pulse response before and after the FFE filtering
4. AGC gain as embedded in FFE
 Figure 4. MMSE-derived FFE taps for channels in Case 1 (left) and Case 2 (right).

A few observations can be made from Figure 4:

(1) The FFE gain G feedfwd that is defined as

N 5

 w (i)
i 1
2
Eq. 18

is more in Case 1, than that in Case 2, due to the more insertion loss as seen in Case 1. The embedded
AGC gain inside the FFE is applied to bring the signal energy to a pre-defined slicing threshold. This same
FFE gain therefore also applies to the input noise to amplify the noise.

(2) The precursor ISI is largely suppressed, but a small residual ISI is still present, as can be seen in Figure
5.

(3) The postcursor ISI tail is also shortened. The undershoot is expected to be removed by the DFE, but
the remaining reflections in Case 2 depend on the extent of the DFE.
Figure 5. A zoom-in to the pulse response, after FFE application to the respective channel pulse response, in Case 1 & Case 2.

The feedback filer is a causal feedback filter, and is calculated also based on the finite-length MMSE
structure as described in Section 6. In essence, it is responsible for removing the N = 5 postcursor ISI
cursors in this example.

Figure 6. Pulse responses for Case 1 and Case 2, after both FFE and DFE have been applied.
 Figure 7 shows explicitly the residual ISI left after FFE and DFE equalizations. These ISI cursors will be
used in Eqs. (9) and (10) for calculating residual IS power to degrade the SNR.

Figure 7. Residual ISI after FFE and DFE for channels in Case 1 (left) and Case 2 (right). The main cursor is removed from the
plots. Postcursors appear on and after arrow.

We now present the calculation of the Modified SNR as presented in Eq. 16 of Section 5. By considering:

1. Variance of precursor ISI

2. Variance of postcursor ISI
3. Variance of crosstalk noise power
4. Variance of random noise power

To make the comparison more useful, we compare the modified SNR for the two channels: Case 1 and
Case 2, for various lengths of the equalizer. Table 1 gives this comparison in detail.
Table 1. Modified SNR results for Case 1 (larger loss & less reflective), and Case 2 (smaller loss & more reflective).

Run number # of FFE taps # of DFE taps Modified SNR for Case 1 Modified SNR for Case 2
(dB) (dB)

1 5 5 28.25 dB 25.09 dB

2 5 15 31.97 dB 26.86 dB

3 5 25 36.22 dB 36.52 dB

4 5 35 37.00 dB 38.88 dB

where RX input SNR for Case 1 = 43dB, calculated as:

E  xRX
2

input 
SNR RX input  Eq. 19
 2
xtalk  RX input   rn2  RX input 
 E  xRX input    250mVpd 

2 2

 xtalk  RX input  1.78mVrms 

2 2

For simplicity, random noise  rn2  RX input  0 is assumed in this example. Note that xRX input represents
the signal amplitude (commonly the main cursor height is used), rather than the DC swing of the signal.

RX input SNR for Case 2 = 44.5dB, calculated the same way Case 1 is calculated, when

E  xRX input    300mVpd 


2 2

 xtalk  RX input  1.78mVrms 

2 2

The difference of signal energy between Case 1 and Case 2 is due to a difference of loss between the
two channels.

We can interpret the results in Table 1 in this way. Assuming that practical SerDes devices have DFE taps
ranging from 5 to 15, Case 1 has Modified SNR 3 to 5dB more than that in Case 2. This shows that, even
though there is more insertion loss in Case 1, Case 2, as a more reflective channel, where the reflections
are far enough, that the impact of reflections are captured accurately in the Modified SNR, giving
actually a smaller SNR value in Case 2.

Although the noise amplification is more in Case 1, the RX input noise is considered small in this
example. Thus, the loss of SNR in noise amplification in Case 1, is insignificant when compared to the
loss of SNR due to uncanceled reflections in Case 2.

As number of DFE taps increases (25 to 35 taps), the Modified SNR reflects the improvement in Case 2,
when most far reflections are canceled out. Since Case 2 has less insertion loss, it requires less AGC gain,
so it is the higher noise amplification in Case 1 that becomes inferior, even though the noise is small.

Note that when DFE taps amount to 35, as a theoretical discussion, the Modified SNR for Case 2 is close
to 2dB more than that in Case 1. This reflects more or less the difference in loss of the two channels, in
terms of their signal-energy ratio at the input.

One might wonder, using Case 1 as an example, there is still a difference between the RX input SNR =
43dB and the long-length equalizer that gives 37dB. This attributes to the (1) penalty due to noise
amplification, and (2) penalty due to the remaining ISI.

b. Calculate ideal Salz SNR for the same 2 channels

To illustrate one main point of the paper, we now calculate the ideal Salz SNR by means of ICR, based on
[2]:

( )
∑ [ ] Eq. 20
for the same channels in Case 1 and Case 2, as shown in Figure 8, by using the same crosstalk noise
aggressors. We see that because of the infinite-filter-length assumption here, the ideal Salz SNR shows
very high SNR ratio. Referring to Table 1, not surprisingly, this is the case when the number of filter taps is
high. As we show before, when the number of DFE filter taps > 25, the Modified SNR in Case 2 is actually
better than that in Case 1, which is consistent with the Salz SNR calculations. However, when various
penalties are included, the Modified SNR in fact shows the opposite. This is not surprising, but to show
that when performing link budgeting, the Modified SNR predicts link performance better under the use
of practical SerDes devices that are being employed in today’s backplane links.

Figure 8. Calculation of ideal Salz SNR for the same channels in Case 1 and Case 2.

To verify the Modified SNR method, SerDes models or link simulators from 6 different chip vendors are
picked to do time-domain simulations. Basic architectures of these SerDes models are listed here, and
eye measurements under certain BER levels are picked for comparison.
 Table 2. Simulation data from different vendor’s simulators/models for comparison between Case1&Case2, Case3&Case4.

1
Vendor Basic Architecture Measurement 8.5Gbps 10.3125Gbps

TX RX Case1 Case2 Case3 Case4

A 2-tap FFE (1 post tap) CTLE + Analog DFE V-Eye@1e-17 147.44mV 113.39mV 109.76mV 120.67 mV

H-Eye@1e-17 0.6UI 0.495UI 0.465 UI 0.465 UI

2
B 4-tap FFE(1 main tap, VGA + 2-stage peaking V-Eye@1e-12 NA NA 557 mV 410 mV

1 pre tap, 2 post taps) filter + 5-tap DFE H-Eye@1e-12 NA NA 0.31 UI 0.31 UI

C 3-tap FFE(1 main tap, VGA + 5 tap DFE V-Eye@1e-17 111mV 139mV 120mV 141mV

1 pre and 1 post tap) H-Eye@1e-17 0.29UI 0.29 UI 0.31UI 0.30 UI

D 3 tap FFE(1 main tap, 1 VGA + Peaking filter + V-Eye@1e-17 102 mV 94 mV 113 mV 68 mV

pre and 1 post tap) 15 tap DFE H-Eye@1e-17 0.5 UI 0.44 UI 0.52 UI 0.46 UI

E 3 tap FFE(1 main tap, 1 VGA + 14-tap DFE (8 V-Eye@1e-17 76.487 mV 75.823 mV 52 mV 49.5 mV

pre and 1 post tap) fixed taps+ 6 floating

taps)

F 4-tap FFE(1 main tap, VGA + CTLE + 5-tap DFE V-Eye@1e-17 135 mV 71 mV 60 mV 62 mV

1 pre tap, 2 post taps, H-Eye@1e-17 0.43 UI 0.36UI 0.29UI 0.29UI

but post2 is not used)

The green color highlights the better result in eye height and eye width in a comparison group.
Generally, with these models and simulators, if we count the number of vendor results in green color,
then we can see that, using channel and crosstalk of Case 1, we obtain better results, than using the
channel with the same crosstalk of Case 2. (The data rate considered here is 8.5Gbps as used before in
previous sections.) This result is consistent with the results as presented in Table 1, where it shows that
when practical SerDes devices are used for equalization, the channel with somewhat more loss actually
can give better SNR, when various equalization penalties are taken into account for the Modified SNR
calculation.

Pick vendor A as an example. Vendor A has 3 taps of FFE, 14 taps of DFE, and a VGA. It gives ~147mV of
eye at BER =1E-17 in Case 1, and gives only ~113mV of eye at the same BER in Case 2. The penalty of
Case 2, relative to Case 1, can therefore be calculated as 2.3dB. The largest penalty in Case 2, relative to
Case 1, is due to Vendor F, giving 5.6dB. Referring to the first two rows of Table 1, the calculated penalty
of Case 2, relative to Case 1, range from 3.2 to 5.1dB. This kind of trend and range of values, appear to
be quite consistent with the SerDes simulation results based on vendors simulators.

For a straight-forward SNR margin calculation, one can compare the results in Table 1 with SNR = 18.6dB
which is the minimum SNR required to achieve BER = 1E-17 under an AWGN channel. One can find that
the SNR margin ranges from 6.5 to 10dB, if 5 taps of FFE and DFE are used, for the channels and
crosstalk noise as considered in Case 1 and Case 2. However, when using the ideal Salz SNR, the margin
is found to be much larger, which amounts to 25 to 26.5dB under the same two channels with the same
1
The study of 10.3125Gbps will be discussed in the presentation due to limited space;
2
Vendor B does not support 8.5Gbps and the targeted BER can only be set to 1E-12.
crosstalk noise. For the purpose of link budgeting the Modified SNR is therefore expected to provide a
more realistic prediction of the performance.

8. Conclusions
We base the Salz SNR fundamentals as a starting point, by first studying its assumptions, from which to
define a list of power penalties which can exist in current state-of-the-art practical SerDes devices
currently employed in backplane applications. Based on well-known research results on the equivalence
of Salz SNR and MMSE-DFE, we then discuss our Modified SNR calculations. Such SNR-modification
methodology may not be unique and similar ways may have possibly been explored by researchers in
the past. Our contribution has been to strike a more formal way to link from mathematical
fundamentals to practical equalizers and noise conditions.

The Modified SNR consists of signal energy at the decision slicer in the numerator, and potentially can
include all kinds of power penalties in the denominator. In the paper, we demonstrate the use of four
power penalties, where the ISI is derived in time domain and the noise are in frequency domain. In
calculating the optimal MMSE equalizer we reference the well-developed finite-length MMSE-DFE
structure, so that the MMSE equalizer in the paper can be conveniently derived from this structure.

The modification of the SNR in this way is specific in the sense that it targets a backplane channel and
targets a given SerDes equalization configuration. Such tight-coupled modification helps bridge the gap
between a theoretical SNR margin and a practical SerDes SNR margin. Consequently, link budgeting
becomes more confident and system design more robust. Various SerDes devices have been simulated
to support the methodology. The proposed methodology could prevent costly mistakes in concluding
one channel is more optimally designed based on Salz SNR while the opposite is true when a SerDes IP is
tested, if it is available.

9. References
[1] J. Salz, “Optimum mean-square decision feedback equalization,” Bell System Technical Journal, vol.
52, no. 8, pp. 1341-1374.

[2] Fiber Channel – Methodologies for Signal Quality Specification – MSQS. T11.2/Project 1734-DT/Rev
3.1, September 13, 2010.

[3] IEEE 10GBase-T 802.3an-2006.

[4] J. Cioffi, G. Dudevoir, M. Eyuboglu, G. Forney, Jr., “MMSE decision-feedback equalizers and coding –
Part 1: Equalization results”, IEEE Transactions on Communications, vol. 43, no. 10, pp. 2582 – 2594, Oct.
1995.

[5] E. A. Lee and D. G. Messerschmitt, Digital Communications, Boston, MA: Kluwer, 1988.

[6] N. Al-Dhahir, J. Cioffi, “MMSE decision-feedback equalizers: Finite-length results”, IEEE Transactions
on Information Theory, vol. 41, no. 4, pp. 961-975, July 1995.
[7] N. Al-Dhahir, J. Cioffi, “Finite-length vs. infinite-length MMSE-DFE: The connection”, in Proceedings of
Asilomar Conference on Signals, Systems, and Computers, vol. 1, pp. 677-681, Nov. 1993.

[8] N. Al-Dhahir, J. Cioffi, “Fast algorithms for the computation of the FIR MMSE-DFE”, in Proceedings of
IEEE International Conference on Acoustics, Speech, and Signal Processing, vol. 4, pp. 533-536, March
1992.

[9] A. Ran, “Time-domain SNR analysis for contributed channels”, IEEE 802.3 Interim Meeting, 100Gbps
Backplane and Copper Cable Study Group, May 2011.

[10] A. Healey, “Link budget for 40GBase-CR4 and 100GBase-CR10”, IEEE P802.3ba Task Force Meeting,
Jan. 2009.

[11] C. Moore, A. Healey, “A method for evaluating channels”, IEEE 100Gb/s Backplane and Copper
Study Group, March 2011.

[12] J. Ewen, M. Meghelli, T. Beukema, “25Gb/s signaling for 100G backplanes: Channel loss vs.
equalization”, IEEE 802.3 100G Backplane and Copper Cable Study Group, March 2011.

[13] H. Takatori, “40/100G copper feasibility”, IEEE 802.3 Higher-Speed Study Group Meeting, Nov. 2007.

[14] H. Takatori, “40/100G copper feasibility – 2”, IEEE 802.3 Higher-Speed Study Group Meeting, March
2008.