Re: [10GMMF] TP3 Meeting minutes, November 23rd
Stefano,
Thank you for the kind words. In response
to your comments,
1) The 12 tap FFE I used in my simulations is
T/2 spaced corresponding to a span of ~600ps.
In many implementations the span of the FFE is the
limiting factor. This is at the low end of your
estimates (6-9 taps of T spaced FFE corresponds
to a span of 600ps - 900ps).
Hence, I very much disagree that 12 T/2 spaced taps
represents a long filter from a practical design.
The results we publised in Nov '03 CFI and the FDDI
experimental results presented in king_1_0904 based
on an Oct 2003 EDC design had FFE + DFE taps that
were representative.
Furthermore, I very much agree with Lew that we use
the number of taps essentially on the high end of what
the consensus feels is practical so we are not too
pessimistic about coverage. The penalty would be
even smaller if I used a span of 900ps in the FFE
as per your suggestion.
2) It is well known in the literature that T/2
spaced FFE is matched to the incoming signal
spectrum. Under conventional LMS adaptation a
T/2 spaced FFE takes on the form of a Matched
Filter followed by an Equalizer.
Hope this helps.
Best Regards,
Sudeep
-----Original Message-----
From: owner-stds-802-3-10gmmf@IEEE.ORG
[mailto:owner-stds-802-3-10gmmf@IEEE.ORG]On Behalf Of
Bottacchi.external@INFINEON.COM
Sent: Thursday, December 02, 2004 1:29 AM
To: STDS-802-3-10GMMF@listserv.ieee.org
Subject: Re: [10GMMF] TP3 Meeting minutes, November 23rd
Importance: High
Sudeep,
That was an excellent work, since now we have a quantitative difference
between ideal-infinite length and still ideal but finite-length
equalizers. Assuming 12 taps (T/2) FFE + 5 taps DFE you computed about
6.1dBo penalty assuming residual ISI as Gaussian unbound noise. I agree
on this approach. Just few comments:
1 - 12 taps is a quite long filter for practical design at 10G and
at my knowledge, sampled IC designs have FFE with shorter length,
usually between 5 to 9 taps and T-spaced. How would be the extra penalty
in this case for a shorter length? Might it be 0.5dBo more? In that case
we would account for 6.5~6.6 dBo for ideal finite length FFE + 5 taps
DFE.
2 - Moving to real finite length we have to account at least for two
more things:
a - EDC receiver is not matched to incoming signal spectrum
and noise bandwidth would not anymore minimum.
b - Noise figure of EDC input stage is not accounted at all
in PID calculations. How would be optical sensitivity degradation due to
EDC input stage noise?
c - Real design must account for circuit impairments,
crosstalk, skew, duty-cycle, not-ideal tap design and temperature
degradation. How this would be accounted for?
I guess that 1dBo for 2a-c is quite optimistic. In conclusion, could we
account for about 8dBo penalty for "real" EDC based link including 99%
coverage? In our recent test we measured about 7.2dBo penalty between
BTB and optical OC-48 link emulation. Similar penalty has been observed
in the case of real 300meters MMF (Infineon Siecor benchmark fiber). I
guess 8dBo should represent a reasonable engineering margin to get link
connected for 300 meters operation including temperature and aging
effects. This would be revaluated in link budget calculations.
Best regards
Stefano
-----Original Message-----
From: owner-stds-802-3-10gmmf@IEEE.ORG
[mailto:owner-stds-802-3-10gmmf@IEEE.ORG] On Behalf Of Sudeep Bhoja
Sent: Donnerstag, 2. Dezember 2004 04:01
To: STDS-802-3-10GMMF@listserv.ieee.org
Subject: Re: [10GMMF] TP3 Meeting minutes, November 23rd
Abhijit, Lew and all,
I agree that there are many ways to implement EDC's and
hence we should strive to choose a relatively implementation
independent EDC metric. PIE-D as a metric has served us
well on this front.
However, we do need to justify that finite length EDC's
fits within the 1.5dB allocated implementation penalty.
I have simulated the OM1 Monte Carlo set with 2 connectors
that are Rayleigh distributed and compared PIE-D with
a 12 tap (T/2) spaced FFE and a 5 tap DFE. Please see
attached my simulation result for standard 1GE offset launch.
My summary:
1) PIE-D for 99% coverage is 5.62dB. This is in excellent agreement to
John Ewen's results in ewen_2_1104.
2) The additional implementation penalty of the
simulated finite length EDC is 0.45dB, well within the
allocated implementation penalty.
Furthermore, in the simulation of the finite length EDC,
I have lumped uncancelled ISI as unbounded Gaussian noise. This
simplifies the penalty calculations but is pessimistic.
A more accurate calculation would only result in a number smaller than
0.45dB
Hence I think that PIE-D and the allocated implementation
penalty are adequate to close the link budget for 300m
of OM1 link with 2 connectors.
Best Regards,
Sudeep
-----Original Message-----
From: owner-stds-802-3-10gmmf@IEEE.ORG
[mailto:owner-stds-802-3-10gmmf@IEEE.ORG]On Behalf Of Abhijit Shanbhag
Sent: Wednesday, December 01, 2004 11:51 AM
To: STDS-802-3-10GMMF@listserv.ieee.org
Subject: Re: [10GMMF] TP3 Meeting minutes, November 23rd
Lew, Martin-
Along the lines that we have been discussing in the conf calls, I'd
suggest treating the link budget and compliance testing slightly
differently.
Regarding link budget considerations, I believe there was a clear
consensus across module, IC and component vendors within this group,
several months back, on defining an "equalizer metric or bound" which
was relatively implementation independent, so as not to constrain
implementation choices & innovation, and to pad it with an
"implementation penalty" corresponding to worst-case scenarios &
finiteness aspects. We have the PIE-D metric, which may be somewhat
constraining in subsuming a DFE-based architecture. This has worked out
well so far. Note that a lot of the empirical data presented by at least
one IC vendor during the IEEE meetings was using an unspecified
equalizer but the PIE-D was always used as a reference. There may be
different ways of achieving performance close to PIE-D, within a certain
range, which may move away from the classical and strict x-tap FFE with
z tap-spacing and y-tap FB. I do not expect it to be easy to come to a
general consensus towards select! ing x,y,z - which may delay things.
Regarding compliance testing, as has been empirically (with h/w) shown
within the IEEE meetings from Jan'04 and along the lines of Petre's Sept
contribution, that an FIR filter (possibly cascaded with a LPF filter)
can emulate the MMF channels closely. I'd agree with Lew's & Petre's
suggestion to use an FIR filter with 4 peaks (taps) across a maximum 3
UI (TBD) with a certain PIE-D (~4.5 dB as been discussed in the last few
calls) to set the reference compliance channel but allow for any
implementation resulting in a waveform within a certain MSE from this
reference. Some of the above numbers might need further validation.
Regards-
Abhijit
------------------------------------
Scintera Networks
Abhijit G Shanbhag
ashanbhag@scinteranetworks.com
4340 Stevens Creek Blvd.,
Suite 260.
San Jose, CA 95129
tel: 408-557-2810 x18
fax: 408-557-2812
mobile: 408-893-8069
------------------------------------
-----Original Message-----
From: Lew Aronson [mailto:lew.aronson@FINISAR.COM]
Sent: Wednesday, December 01, 2004 10:21 AM
To: STDS-802-3-10GMMF@listserv.ieee.org
Subject: Re: [10GMMF] TP3 Meeting minutes, November 23rd
I have long commented that a finite metric was important since the PIE
metrics were not nearly as sensitive to long impulse responses, and we
risk lumping in very long impulse responses into the compliance test.
The problem, of course, is what parameters (number of taps etc) to
choose for the finite equalizer metric. If we make a somewhat aribrary
choice which is less than what can practically be achieved, we can get
results which are much too pessimistic.
I believe we have seen some presentations (sorry I can't recall the
specific ones at present) on finite equalizer calculations, and
generally, they seem to show that the extra penalty is very small as
long as the number of taps in the equalizer is appropriate to the length
of the impulse responses, and that the number of taps for the cases
considered did not seem unrealistic.
I would suggest that if we do adopt a finite equalizer metric, it be
with a relatively large number of taps, essentially on the high end of
what the consensus feels is practical. This would not imply that you
had to implement such a long equalizer, but would prevent us from coming
to an overly pessimistic conclusion of the coverage, yet still prevent
us from selecting overly long impulse response cases for the TP3
stressed sensitivity test.
It would be very helpful to start getting suggestions on what the
parameters of this ideal finite equalizer would be even ahead of the
decision to start using the metric.
Lew
Lew Aronson (lew.aronson@finisar.com)
Finisar Corporation
1308 Moffett Park Drive
Sunnyvale, CA 94089-1133
408-542-4215 (PH)
408-543-0083 (FAX)
-----Original Message-----
From: Lobel, Martin [mailto:martin.lobel@INTEL.COM]
Sent: Wednesday, December 01, 2004 3:18 AM
To: STDS-802-3-10GMMF@listserv.ieee.org
Subject: Re: [10GMMF] TP3 Meeting minutes, November 23rd
All,
I completely agree with the statement that finite-length MMSE-DFE
calculations are indeed the way to go for an accurate analysis of the
problem. The infinite-length equalizer (PIE metric) has, however, the
strong benefit of very simple computation for fast estimation of the
challenge (penalty) of the problem.
I believe that there is no way around an analysis based on finite-length
equalizer in order to verify the link budget and Intel is working on
simulation results to share with IEEE. As a matter of fact, Intel
proposed back at the March meeting in Orlando (see
http://www.ieee802.org/3/10GMMFSG/public/mar04/lobel_1_0304.pdf) to
focus on finite equalizers in order to establish a link to equalizers
with realistic filter complexity and thereby include up front the
penalty associated with finite vs infinite equalizers.
Regards,
Martin
-----Original Message-----
From: owner-stds-802-3-10gmmf@IEEE.ORG
[mailto:owner-stds-802-3-10gmmf@IEEE.ORG] On Behalf Of Lars E. Thon
Sent: 30. november 2004 19:06
To: STDS-802-3-10GMMF@listserv.ieee.org
Subject: Re: [10GMMF] TP3 Meeting minutes, November 23rd
Dear Robert, and other contributors to this thread,
I believe that using finite-length MMSE-DFE calculations is indeed a
more accurate method for closing the the link budget.
For an exposition on the topic, please read
http://grouper.ieee.org/groups/802/3/aq/public/jul04/thon_2_0704.pdf
which I presented in Portland in July 2004. Other contributors,
including Intel and recently Georgia Tech, have also touched upon this
subject, either directly or indirectly.
I agree with Sudeep that MMSE-DFE (whether "ideal" or finite length) is
a better measure of the required penalty term of the link budget than is
a ZF-DFE. The latter ignores noise, while the former attempts to balance
noise and the pattern-specific ISI.
Neither MMSE-DFE nor ZFE-MMSE is the "optimal" DFE (Minimum BER or
MBER-DFE). However, I think an analytic method for calculating the
MBER-DFE in the general case is beyond the reach of current
communication theory, let alone a practical goal for an adaptation
algorithm.
Hence, I think it is a good choice to use MMSE-DFE for performance
evaluation and budget closure. The ideal/PIE-D/infinite version is good
because of its simplicity, whereas one should also contemplate using a
finite length version for more accurate analysis.
Lars
Lingle, Jr, Robert (Robert) wrote:
> I am less knowledgeable than many on this list, but I am trying to
> understand the difference between the ideal, infinite case and the
real,
> finite case. As a step in that direction, we tried in San Antonio
> presentation to take a step in that direction by looking at ideal,
finite
> case.
>
> What we saw, but do not fully understand, is that it seemed that the
higher
> the MSE PIE-D, the lareger the difference between PIE-D and any finite
> equalizer could be. Some questioned my conclusions afterward, but
none
> refuted them conclusively.
>
> The implications are the following: suppose we set a limit that 99% of
> fibers must pass PIE-D < 5.5, and leave a 1 dB implementation penalty.
Then
> what we really need to know is the following: do fibers with PIE-D
between
> 4.5 and 5.5, which nominally seem to pass, actually have
implementation
> penalties (hardware and equalizer design) that keep them below the 6.5
dB
> limit? If my logic is flawed, please help me correct it.
>
> do we have to calculate finite equalizer cases as well?
>
> Robert Lingle, Jr
> Fiber Design and Development
> OFS R&D, Atlanta, GA
>
>
> -----Original Message-----
> From: Bottacchi.external@INFINEON.COM
> [mailto:Bottacchi.external@INFINEON.COM]
> Sent: Tuesday, November 30, 2004 10:23 AM
> To: STDS-802-3-10GMMF@listserv.ieee.org
> Subject: Re: [10GMMF] TP3 Meeting minutes, November 23rd
> Importance: High
>
>
> John,
>
> It is an interesting proposal, at least theoretically. Assuming
infinite
> length FFE, I guess zero forcing solution would be possible (in
> principle). This means no residual ISI at sampling instant (not just
> minimum error modulus like MMSE). As a consequence, noise enhancement
is
> strongly expected to be the limiting factor for EDC performances. I am
> quite curious to see how much would be the optical penalty since I
> calculated some month ago the PID-L for the zero forcing (linear)
> equalizer in term of noise bandwidth enhancement( I sent a ppt copy to
> Sudeep for comparison with PIE-D/L reported metrics). Does zero
forcing
> PIE-D expected to be different from noise bandwidth enhancement due to
> full frequency compensation for a given output ISI free spectrum
> (raised-cosine for example)?
>
> Thank you for posting this issue...
>
> Best regards
>
> Stefano
>
> -----Original Message-----
> From: owner-stds-802-3-10gmmf@IEEE.ORG
> [mailto:owner-stds-802-3-10gmmf@IEEE.ORG] On Behalf Of Abbott, John S
Dr
> Sent: Dienstag, 30. November 2004 15:48
> To: STDS-802-3-10GMMF@listserv.ieee.org
> Subject: Re: [10GMMF] TP3 Meeting minutes, November 23rd
>
>
> In regard to Lew's and Sudeep's points, are there conditions where we
> should calculate the zero forcing PIE-D as well as the MSE PIE-D? At
> the San Antonio meeting there seemed to be comments supporting both.
>
> Regards,
>
> John Abbott
>
> -----Original Message-----
> From: Sudeep Bhoja [mailto:sbhoja@BIGBEARNETWORKS.COM]
> Sent: Monday, November 29, 2004 5:39 PM
> To: STDS-802-3-10GMMF@listserv.ieee.org
> Subject: Re: [10GMMF] TP3 Meeting minutes, November 23rd
>
>
> Lew,
>
> This in response to your Comment #1 below.
>
> The PIE-D equations from bhoja_1_0704 targets an infinite length DFE
> that minimizes mean square error (MMSE). MMSE Equalizers perform
better
> than Zero forcing Equalizers. Conventional adaptation algorithms such
as
> LMS, minimize mean squared error.
>
> If we agree on an MMSE based infinite length DFE, there are two
> variables that enter into the calculation:
>
> 1) The first variable is easy and follows directly from the link
budget.
>
> sigma^2 -> This is the Electrical noise floor at the input
> of the ideal EDC and is easily derived from the link budget. We had
> previously set sigma^2 = 10^(-17-2*6)/10 since we had allocated 6dB of
> total dispersion budget.
>
> However since the connector loss was updated to 1.5dB from 2dB Page 5
in
> lawton_1_1104 allocated 6.5dB to the total dispersion budget.
>
> Hence we need to update sigma^2 = 10^(-17-2*6.5)/10 = 10^(-30/10) in
our
> PIE-D calculations.
>
>
> 2) The rise time used in deriving the fiber pulse response in
> bhoja_1_0104 was set to 47.1ps (20-80% Gaussian). This number was
chosen
> from -LR.
>
> For the purpose of the TP3 stressed tests, we only need to represent
the
> rise time of the test setup of Fig 68.6 in D0.2. For this purpose
47.1ps
> is probably an adequate rise time.
>
> Best Regards,
>
> Sudeep
>
>
> -----Original Message-----
> From: owner-stds-802-3-10gmmf@IEEE.ORG
> [mailto:owner-stds-802-3-10gmmf@IEEE.ORG]On Behalf Of Lew Aronson
> Sent: Monday, November 29, 2004 12:47 AM
> To: STDS-802-3-10GMMF@listserv.ieee.org
> Subject: Re: [10GMMF] TP3 Meeting minutes, November 23rd
>
>
> Some comments:
>
> 1) I think it is very important that all are aligned on the PIE-D
> calculation. I would be interested in a discussion on the variables
> mentioned below, what differences exist now between different task
force
> members algorithms and the likely impact of changes of each of these
> parameters.
>
>
> Lew
>
>
>
> -----Original Message-----
> From: Michael Lawton [mailto:mike_lawton@AGILENT.COM]
> Sent: Wednesday, November 24, 2004 6:10 AM
> To: STDS-802-3-10GMMF@listserv.ieee.org
> Subject: [10GMMF] TP3 Meeting minutes, November 23rd
>
>
>
> Dear TP3ers,
>
> Here are my notes from yesterdays call.
>
>
>
> Key issues which were raised:-
> PIE-D has variables associated with it (rise time, sigma^2, ZF
> vs MMSE calculation) - how do we handle that?
>
> Any comments/corrections please get back with me.
>
> Best Regards
>
> Mike
--
----
Lars E. Thon <lars@aeluros.com>
Aeluros Inc., 201 San Antonio Circle, Suite 172
Mountain View, CA 94040-1254
650-917-4113(w) 650-917-7394(f) 408-439-5914(c)