Sanjay,
If you were to review the following presentation
(pg.12 & 17)
There you'll see:
1There are two distinct margins specified; a noise
margin and a nonlinearity margin.
2Two cases have been studied; one for 0m and
another for 100m
Please also review:
and compare the inputoutput nonlinear graphs given
on pg.5 & 8. Notice how
by PGA helping to normalize the inputoutput curve
the absolute nonlinearity
values for all three line codes drops however their
relative values remain more
or less the same with PAM5 tolerating the highest
amount of nonlinearity.
Regards,
Joseph N. Babanezhad
Plato Labs.
 Original Message 
Sent: Wednesday, May 19, 2004 1:33
PM
Subject: [SPAM] Re: [10GBT] [SPAM] Re:
[10GBT] [SPAM] [10GBT] symbol rate
Joseph,
You say:
The criteria for our linearity
analysis has been with 0m cable (loopback) where the RX signal
is the largest. Based on this
criteria AFE's nonlinearity should be such that it does not cause
any errors in the input
pattern.
I
think there are two extreme situations to cover:
a)
0m cable length. There should be no errors there. In this situation, the
signal level is going to be high so nonlinearity can take up most of the
overall "noise" budget
b)
100m cable length with worst case alien next. Here there signal level is low.
Now the overall noise budget must be shared between nonlinearity and other
impairments. In this situation the absolute voltage levels seen will be lower
than in (a) because the remote transmitter will be significantly
attenuated.
Regards,
Sanjay
cell (650)
7047686
office (408)
6532235
George,
Our cancellation numbers all along have been the
same. See for instance the following (pg.9):
The criteria for our linearity analysis has been
with 0m cable (loopback) where the RX signal
is the largest. Based on this criteria AFE's
nonlinearity should be such that it does not cause
any errors in the input pattern.
Regards,
Joseph N. Babanezhad
Plato Labs.
 Original Message 
Sent: Wednesday, May 19, 2004 9:07
AM
Subject: [SPAM] Re: [10GBT] [SPAM]
[10GBT] symbol rate
Joseph
–
The main value of
the program you used is that it gets beyond capacity limitations. The
constellation sizes noted in the printout ( PAM4) are for uncoded.
PAM5 at 1250 Mbaud is equivalent to coded PAM4. In order to get the
margin for the PAM5 system, you can take the PAM4 DFESNR (which is the
Optimal DFE SNR for 1.25 Gbaud PAM, regardless of the number of levels), and
subtract from it the SNR that your coded system requires.
Equivalently the
PAM_2p5 DFESNR is the optimal DFE SNR for a 1 Gbaud system (regardless of
the levels), and the PAM8 DFESNR is for an 833.333 Mbaud system (3
bits/baud/pair at 10 Gbps) regardless of the number of levels.
For coded margins, these SNRs can be compared to the SNRs listed
for the coded systems, properly adjusted for shaping gain or precoding
loss.
For reference, you
mentioned a 1.56 Gbaud PAM4 system, which I presume has coding overhead, it
would be equivalently (10Gb/sec / 4 pairs / 1.56 Gbaud = 1.60 bits /
baud).
I hope this
helps.
The point of all
this is that it is wellknown that increasing the bandwidth of a DFE, DFSE,
or TH Precoded system does NOT always give you the benefit of the increase
in the channel capacity by including more bandwidth. You CAN signal
too fast for the channel. The point of diminishing return is usually
where SNR(f) (SNR at a given frequency f) approaches zero (crossover of
signal and noise) for all frequencies > f_o. For PAM systems, f_o
will be the nyquist rate, half the baud rate. Signalling beyond this
rate generally adds less to the SNR in marginal capacity than it loses due
to the additional noise and shorter baud interval. See slide 25
(1^{st} backup) in my presentation at the last meeting for the basic
math, or go directly to the reference: J. Salz, “Optimum meansquare
decision feedback equalization”, Bell System technical Journal, pp.
13411373, Oct. 1973.
Joseph, on another
topic,I also see that you have had to increase the echo, next and fext
cancellation numbers significantly over the default values. This seems
to fly in the face of your claimed improved nonlinearity tolerance.
While the received equalized signal in the absence of interference will be
more tolerant to nonlinearity, nonlinearities following the point where the
reference for next/fext and echo cancellation are taken will will show up as
residual echo, next, and FEXT, and hence a good proxy for the linearity
requirements is the maximum degree of cancellation required. The
signal that has to be cancelled (echo, next or fext) will have a near
Gaussian PAR in reality, so the linearity of the canceller or cancellation
signal is largely independent of the number of levels used in the line
code. As such, you seem to have required some 50 dB of linearity for
FEXT cancellation alone, (was 30 dB), and upped the echo cancellation to 65
dB. Did you really require those increases?
george
Original
Message From:
stds802310gbt@IEEE.ORG [mailto:stds802310gbt@IEEE.ORG] On Behalf Of Joseph
Babanezhad Sent:
Wednesday, May 19,
2004 3:25
AM To:
STDS802310GBT@LISTSERV.IEEE.ORG Subject: Re: [10GBT] [SPAM] [10GBT]
symbol rate
In May of 1998 at the CICC
conference while waiting to present my paper
I was listening to Mehdi
Hatamian of Broadcom, one of the movers &
shakers
of IEEE 1000BASET standard,
give his tutorial presentation on 802.3ab standard
draft. There was one thing that
he kept repeating it over and over ... and over
again;
"Remember the most important
things for Ethernet are power, power and power"
If this was relevant to
1000BASET it definitely is more relative to
10GBASET.
With this in mind let me address
your comments:
1Please do not confuse PAM4
with PAM5
PAM4 baudrate=1.56 GB/s
Nyquistfrequency=780 MHz
PAM5 baudrate=1.25 GB/s
Nyquistfrequency=625 MHz
2The complete list of PAM5
advantages are:
bMore tolerant to AFE
nonlinearity
cSignificantly lower
power
3As far as channel's higher IL
& ANEXT at frequencies beyond 500 MHz are
concerned the following are the
capacity simulation results using SolarFlare's
provided program from the
website:
Launch Power : 7 dBm (2Vpp
PAM5)
nextcanc=50;
echocanc=65;
fextcanc=50;
solarsep_varlen7a(10.5,650,4,55,6,1,6,2)
solarsep_varlen7a(4.5,650,4,100,6,1,6,2)
 Original Message 
Sent:
Tuesday, May 18, 2004 9:14 PM
Subject:
[SPAM] [10GBT] symbol rate
Looking through some of the
prior presentations proposing 4PAM and 5PAM I see that the
significant benefits claimed for 5PAM or 4PAM
are
a) Lower SNR
requirement.
b) Lower linearity
requirement
Going from 5PAM to 10PAM
raises the SNR required for the same BER by 6dB.
The increase in SNR required
is indeed painful, however the presentations proposing the lower PAM do
not take into account the fact that
a) The higher symbol rate will
result in higher net attenuation since the transmit spectrum extends to
higher frequencies where the attenuation is higher
b) The alien cross talk is
higher at higher frequencies.
Both a and b result in the
available SNR being lower than what you get at symbol rates below
1Gsym/sec for the Channel models #1 and #3.
For short cable lengths or
lower attenuations, this effect is less severe.
Bottom line is that beyond
about 1Gsym/sec, the theoretically achievable system margins drop off
sharply.
The higher PAM (8, 10,
12 etc) will require higher SNR and this implies higher linearity
requirements at the transmitter, lower noise in the receiver than if you
targeted a shorter distance using 5PAM however I don't think this is a
choice we have given our distance objectives.
Please note that this is my
personal opinion and not a directive as editor.
I leave this as a qualitative
argument because quantitative arguments have been made earlier but don't
seem to have been accepted by some. I hope this
helps.
Regards,
Sanjay
Kasturia
sanjay@teranetics.com
cell (650)
7047686
office (408)
6532235
fax (408)
8448187
Teranetics
Inc.
2953 Bunker Hill Lane, Suite
204
Santa Clara, CA
95054
