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 estimation program for your evaluation


Strictly speaking, Chris provided no information on alien xtalk. Based on what he provided yields a capacity of 17.4 Gbps (using your program with the aforementioned corrections).  Of course alien xtalk can't be ignored and must be dealt with accordingly.     As a point of clarity, what you are using to model alien xtalk is the pair-to-pair NEXT limit line out of 568b lowered by 3dB.  We provided the NEXT limit for comparison to our measurements only.  We can hardly draw any conclusions from the heuristic use of a NEXT limit line applied as an alien xtalk limit line.  

As to whether scaled measurements will yield capacity results substantially different then the smooth envelop associated with the measurements, I have this to say.  First, by definition, the envelop exceeds the measurement at all frequencies.  Therefore, a measurable contribution to capacity will occur at every frequency where the measurement does not touch the limit line.  I suggest reducing the frequency step size of the integration before concluding whether or not there is an appreciable difference in capacity. Secondly, from our experience and from talking with a number of individuals who are actively taking cable measurements,  high frequency extrapolation of these limit lines are rather pessimistic.



-----Original Message-----
From: []
Sent: Thursday, February 20, 2003 1:18 PM
To: William Jones
Subject: RE: [10GBT-Modeling] RE: [10GBT-Cabling] [10GBASE-T] a channel
capacity estimation program for your evaluation


Sorry I have to clarify my opinions which may not be the same as yours.
According to the calculation of the program which is based on the ISO
channel model provided by Chris, the target we are talking about (10Gbps
over 100m CAT-5E/6 cable) is even theoretically infeasible (channel
capacity is <10Gbps).   And I don't think that there will be an appreciable
difference in capacity when the measured data (while scaled to the channel
model limit)  is used.

Another thing I want to clarify is about the Shannon limit.  The Shannon
limit is only theoretically achievable without limit on the complexity of
system.  With the currently proposed approaches (traditional
modulation/coding schemes), I don't think there is any way for us to make
it close to the channel capacity.  Nevertheless, the implementation
constraints will only make the really achievable performance worse.



"William Jones" <> on 02/20/2003 11:57:35 AM

To:    <>, <>
cc:    <>,

Subject:    RE: [10GBT-Modeling] RE: [10GBT-Cabling] [10GBASE-T] a channel
       capacity estimation program for your evaluation


I'm glad you feel now that this is only a difficult problem and not one
that is impossible.  We have always said that it was a hard problem.  To
your point that these calculations are theoretical limits that may be
practically challenging to realize in implementations is precisely why we
presented cycle based time domain simulation results as well as the
capacity results.  This modeling approach captures the entire effect of the
channel and crosstalk including phase information which is lost in the
frequency domain calculations and which is important for practical
considerations.  These simulations included finite length filters,
adaptation jitter and clock recovery effects among other things thereby
accounting for real implementation effects and supporting the
implementation feasibility.

To the point about the transmit PSD, after correcting for the programming
errors, I got an additional 760Mbps in capacity with a flat PSD bandlimited
to 400MHz.  I do agree that alien xtalk is a limiting noise source but
still dispute the level.  If I lower the alien xtalk level by 6 dB aligning
it closer with our measured data but still using the ISO cable models I get
a capacity of 10.0Gbps on Cat5e.

To Chris' point about using measured data scaled to limit lines, I think
this is very important.  There is an appreciable difference in capacity
when the calculations are used with smooth curves for the xtalk as compared
with actual measured data.  There is a lot of physically based structure
in, for example NEXT, as seen in any of the frequency response curves which
is further reflected in the impulse responses.  We have seen SNR
improvements on the order of 3 to 5 dB when using the measurement data as
opposed to the envelop.



William W. Jones, Ph.D.
Director of Systems Engineering
SolarFlare Communications, Inc.
949-581-6830, ext. 2550
mobile: 619-405-2445
fax: 949-581-4695

-----Original Message-----
From: []
Sent: Thursday, February 20, 2003 10:23 AM
Cc:;; William Jones
Subject: Re: [10GBT-Modeling] RE: [10GBT-Cabling] [10GBASE-T] a channel
capacity estimation program for your evaluation


First, thank you for pointing out the mistake I made on the return loss.  I
have corrected it in the attached program.  As you said, it does not affect
the result much.

Also you are correct that the input signal PSD should not be calculated
using my simplified approach.  Instead, the best thing we can do is to
launch as much as signal power in this band while meeting the FCC
restriction.  However, by my observation, the data I used in the program is
very close to the best thing we can do.  So you can think the approach is
reasonably near-optimal.

The difference between Solarflare's data and the ISO 11801 cable models was
a topic in the last Jan meeting.  Should we follow the standard model or
some specific measurements is still a problem.

I did not give a careful calculation of the channel capacity based on
Solarflare's channel model.  But I guess you are right.  The major factor
that causes the difference between the results based on the standard
channel model and Solarflare's model is the strength of the residual alien
NEXT since it is the dominant noise source.  In my view point, suppressing
alien NEXT is very difficult, especially when there are many disturbers.

Also I would like to emphasize that the calculated Shannon capacity is the
theoretical limit.  It is fairly reliable when the frequency-domain channel
model is reliable.  However, as all of us have known, it is not easy to
approach this limit at all.  In my opinion, even the Shannon capacity can
be 14Gbps based on Solarflare's calculation, to achieve effective 10Gbps
channel throughput will be extremely difficult.


Marvell Semiconductor Inc.
(Please check the updated i3e.m)

(See attached file: i3e.m)

"Joseph Chou" <> on 02/19/2003
07:28:01 PM

Please respond to <>

Sent by:

To:    <>, <>,
cc:    "'William Jones'" <>

Subject:    [10GBT-Modeling] RE: [10GBT-Cabling] [10GBASE-T] a channel
       capacity estimation program for your evaluation

Hi Xiaopeng,

I am new in this reflector. I looked at your program. Here are some

1. The conclusion that you pointed "the Alien NEXT dominates the
capacity" is surely right.
2. The Echo (Return loss) used in i3e.m has wrong sign. The return loss
should be negative (loss) when adding to the signals. However, it does
not affect the result much.
3. The Input Signal Si in i3e.m should be calculated based on the
bandwidth. In the theoretical estimation of channel capacity, the actual
baud rate due to coding or modulation is not in the picture.
4. The Alien NEXT model used in i3e.m is about 5 to 10 dB higher than
the one used by SolarFlare's presentation of last November.
5. The SolarFlare's presentation also has less Insertion Loss (5 to 10
dB) than what specified in the ISO 11801 Class D.

The above observations conclude why you and Bill have different result
of channel capacity: 7GH vs. 14GH for Cat5e cable.

Best Regards,

-Joseph Chou

-----Original Message-----
[] On Behalf Of
Sent: Wednesday, February 19, 2003 2:13 PM
Cc: William Jones
Subject: [10GBT-Cabling] [10GBASE-T] a channel capacity estimation
program for your evaluation

Hi, Bill and 10GBASE-T SG members,

Sorry for the so late response.  I was on a quite long vacation past
several week.

I have finished a Matlab program to distribute in the 10GBase-T SG for
the evaluation of the bottomline channel performance.  The channel
simulation is based on the models (return loss, FEXT, NEXT, insertion
loss) provided by Chris and a presentation (Alien NEXT) last Novermber.
Without consideration of any implementation and modulation/coding, the
Shannon channel capacity can be calculated only based on the
channel model.   Please read the comments in the program for details

Unfortunately, based on the available channel models, even with a quite
optimistic belief in the achievable DSP power, the Shannon capacity
tells me that both CAT-5E and CAT-6 cables cannot support 10Gbps
throughtput over a distance of 100 meters.  Please read the comments in
the program for details.

Surely, the calculation is completely based on the available channel
models.  I know more work is needed on setting up a channel model
accepted by all of us.  But unless some dramatic improvement can be
achieved in the alien NEXT suppression, the program tells us pretty much
the story.

If you have any question about the program itself, you are more than
welcome to post on the group reflector for discussion.

Best regards,


(See attached file: i3e.m)

"William Jones" <> on 01/22/2003
11:18:48 PM

Sent by:

To:    <>

Subject:    RE: [10GBASE-T] channel model


How are the calculations coming?  Are you using the models Chris sent
out last week?


-----Original Message-----
From: William Jones
Sent: Wednesday, January 15, 2003 4:08 PM
Subject: [10GBASE-T] channel model(part 2)


Sure.  Once the common channel model (I only need the frequency domain
characteristics) is available, I can begin to work on it.



Would you be willing to sign up for technical feasibility based on an
optimal DFE calculation?



I has no problem to use TIA or ISO based channel and noise model once it
is standardized.  My point is that even using the frequency-domain model
we can still give a quite good estimation of the practically achievable
distance at a throughtput of 10Gbps over CAT-5E or CAT-6 cables.




We believe, as George discussed in his part of the tutorial in Kauai,
that with an adequate amount of crosstalk cancellation, the throughput
is achievable.  The question then becomes can this level of cancellation
be practically achieved, hence, my interest in time domain models.

Until we get the models from the channel modeling Ad Hoc, we could argue
this point endlessly.  Until then, I still believe we should not use DSL
models, but, rather something from TIA 568b or ISO 11801.