Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

RE: [10GBT-Modeling] RE: [10GBT-Cabling] [10GBASE-T] a channel capacity estim...



George, All

The primary purpose of the ongoing discussion was to decide if the PAM10
design could exist on 100m CAT5 cable. Following the discussion I don't feel
on the solid ground as tens dBs in SNR grow and fall mostly in cancellation
ratio parameters. The Alien NEXT level and possible cancellation are not
based with enough data. 
The implications of the Echo, NEXT, FEXT cancellation ratio presented by
Solarflare also are not clear to me. The cancellation ratio will be impacted
by  coefficients resolution in digital domain, by jitter and other
impairments in analog domain. Does the proposed cancellation ratio demand
reasonably achievable analog and digital parameters?

Meanwhile to promote the primary purpose I would like to refer to the
document distributed by William Jones few weeks ago (attached). 
If I understand correctly it describes the SNR after the equalizer on 100m
CAT5 with ground noise only. The SNR for -140dBm/Hz ground noise (no Echo,
NEXT, FEXT) is roughly 28dB. Assuming coded signal SNR for BER 10^-10 as
25dB, it remains only 3 dB margin for Echo,NEXT,FEXT, Alien NEXT and
implementation impairments.
Again if I understand correctly the graph, it seems that there will be
negative margin considering all noises exist.

Regards,

Boris Fakterman - Intel Communications Group, Israel
Tel: 972-4-865-6470, Fax: 972-4-865-5999
mailto:boris.fakterman@intel.com
-----Original Message-----
From: George Zimmerman [mailto:gzimmerman@solarflare.com]
Sent: Wednesday, February 26, 2003 3:24 AM
To: stds-802-3-10GBT@ieee.org; stds-802-3-10GBT-Modeling@ieee.org
Subject: RE: [10GBT-Modeling] RE: [10GBT-Cabling] [10GBASE-T] a channel
capacity estim...


Xiaopeng & Zeev -
Respectfully, no matter how many times you say it, you continue to
grossly distort the proposals put forward.  I have a plot generated by
Zeev's version of your code to show the input noise levels, which can be
compared with the measurement-based models in the November tutorial. I
tried to attach just the relevant graphs from the tutorial, but the
reflector bounced it for size - you'll have to go to the web site.

A few of the significant differences, some mentioned by Bill in an
earlier email are:
1) Use of smooth limit lines vs. a measurement-based model (such as was
used for 1000BASE-T) scaled to worst-case.  This is NOT "a couple of
dB", more like 4-6 dB, and, more importantly, changes the relationship
for the required cancellation.
2) Do NOT use the "required cancellation" numbers we gave for
"achievable cancellation", and it is inappropriate to use them with
different line models.  When there is more crosstalk, as has been said
before, it is often the case that more cancellation is achievable.  This
is often true because the root cause of the crosstalk has changed so
that it involves a shorter time delay with stronger coupling.
3) As more noise sources are accounted for the "background" must be
reduced.  We used -143dBm/Hz in the November tutorial & support that (or
less, based on measurements) when Alien crosstalk is accounted for
separately, as was done in the capacity calculations in the tutorial.
(worth 3 dB)
4) the Alien NEXT model is overly pessimistic, this is a discussion in
the modeling group. Not just the limit line, but data shows (see the
November presentation, not from us, but from Sterling & Avaya) that
actual Alien NEXT is significantly below (10 dB at least) the limit in
the higher frequencies.
5) Zeev has incorrectly used 0 dB alien NEXT reduction in his code under
"SolarFlare cancellation".  We clearly show 10 dB relative to our model.
If the Alien NEXT model is different, more cancellation is likely
possible.  I can't say without seeing a cable & the model.  You can't
just adjust the model keep the cancellation fixed, they are related. (10
dB improvement)

So, we're seeing more than 20 dB pessimism here.  I'd scarcely say "a
couple dB".  It's a pretty gross misrepresentation.  What we need to do
is wait for code using proper models.

George Zimmerman
gzimmerman@solarflare.com
tel: (949) 581-6830 ext. 2500
cell: (310) 920-3860

-----Original Message-----
From: xichen@marvell.com [mailto:xichen@marvell.com] 
Sent: Tuesday, February 25, 2003 2:25 PM
To: Ze'ev Roth
Cc: stds-802-3-10GBT-Cabling@ieee.org;
stds-802-3-10GBT-Modeling@ieee.org; stds-802-3-10GBT@ieee.org
Subject: RE: [10GBT-Modeling] RE: [10GBT-Cabling] [10GBASE-T] a channel
ca pacity estim...


Ze'ev,

Thank you for your message.  Your observation is right.  I thought that
the
-140dBm/Hz background noise level is a double-sided PSD when I got it
from
the document.  Either reducing the background noise by 3dB or increasing
the input signal PSD by 3dB should fix the problem.   You have also
provided the capacity results after the modification.  They basically
tell
us the same story we have been facing.

Of course the smooth limit line model used in the program will be
replaced
by the scaled, selected, measured channel data when they are offically
available.  Only couples of dB SNR improvement to performance based on
the
channel limit model should be expected.   Once we obtain more accurate
results on the channel capacity, we will be able to to assess our
achievable targets for the 10GBT standard.

Regards,

Xiaopeng







"Ze'ev Roth" <zeevr@mysticom.com> on 02/25/2003 05:50:48 AM

To:    "'xichen@marvell.com'" <xichen@marvell.com>, CDimi80749@aol.com
cc:    stds-802-3-10GBT-Cabling@ieee.org,
       stds-802-3-10GBT-Modeling@ieee.org, stds-802-3-10GBT@ieee.org

Subject:    RE: [10GBT-Modeling] RE: [10GBT-Cabling] [10GBASE-T] a
channel
       ca   pacity  estim...


Xiaopeng hi,
Very good work.

In order to probe into this deeper,  I  initially simplified your
simulation
to having only a single simple impairment - background noise (i.e., I
removed from your simulation all other impairment: NEXT, FEXT, ANEXT,
ECHO).
The resulting capacity was 15.29Gbit/sec.
This simplification allows me to  compare your results with my program's
results. Running my routine on same parameters I got capacity of
17Gbit/sec.
So clearly there was a discrepancy in the results.
Previously I've cross checked my routine on simple problems and compared
to
textbook results, as well as put it to scrutiny with a several
colleagues,
so I am quite confident it yields correct results.

Therefore, I dug a bit into your equations (in the Matlab file), I think
there is a slight problem with the definition of spectral density (it
doesn't account for double sided) - there is a subtlety in capacity
equations (the usual 3dB problem) and I think you may have fallen into
this
pit. And indeed when I add 3 dB to the noise floor in my simulation I
get
capacity of 15.327. The difference from 15.29 can probably be attributed
to
a different frequency grid and that I used an older version of the
insertion
loss limit equation which is marginally different than the one you've
used.

I've taken the liberty to modify your code (I started out with the
latest
version you've sent) to account for the double sided density (one can
easily
switch between the original code and my correction) and attached it
herein.
I've added comments showing were the single sided - double sided
spectral
density switch occurred in my opinion. I've also  added a sanity check
option for simple AWGN channel case.

Running the modified program I got the following results:
Cable=CAT-5E   Cancellation=Marvell             Capacity= 8.89 Gb/sec
Cable=CAT-6     Cancellation=Marvell            Capacity=11.36 Gb/sec

I've also added the option to use Solarflare's figures for impairements'
DSP-improvement as presented in Kauai.
Running the modified program under these assumptions yields:
Cable=CAT-5E      Cancellation=SOLARFLARE        Capacity=5.57 Gb/sec
Cable=CAT-6       Cancellation=SOLARFLARE        Capacity=7.26 Gb/sec

Summarizing, although there was a small flaw in the original M file,
your
basic conclusions seem to hold water and moreover using Solarflare's
assumptions regarding DSP cancellation performance yield that neither
CAT5E
nor CAT6 can support 10Gb/sec for 100m cable length.

Regards,
Ze'ev
Mysticom





Attached is a plot of the optimal DFE SNR versus white noise level for two
baud rates.  The PSD is PAM with a launch power of 9.5 dBm.  The channel
model is the 11801 model provided by Chris.

Could someone validate this with their own calculations? thanks

Bill


 <<isomodelperf.doc>> 

isomodelperf.doc