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

Re: [10GBT] Validity of comparing PSDs to determine electromagnetic emissions


If you believe that dependent parity checks make such a huge impact on the
performance of an  LDPC code under soft-decision decoding, then why bother
adding independent parity checks and reducing the code rate? Why don't we
just keep adding dependent parity checks ad infinitum until the desired BER
performance is met? Certainly, we can do this with any (N, N-2) code, since
there are an infinite combination of dependent checks that can be created
with just two independent parity check vectors.

In any case, I intend to show simulation results for 1E-12 performance of
the (2048,1723) code at the September interim, thus obviating the need for
choosing risky LDPC codes and ensconcing them permanently in our 10GBASE-T

Sailesh Rao.

>From: Scott Powell <>
>Subject: Re: [10GBT] Validity of comparing PSDs to determine
>electromagnetic emissions
>Date: Wed, 8 Sep 2004 11:47:11 -0700
>   In regards to the possible issue you raised below:
> > ** Sailesh Rao wrote:
> >
> > Furthermore, I've found another issue with the PAM12 proposal.
> > I leave you now with a description of this Issue #6.
> >
>your statements on the code construction are correct but your "bet" on the
>performance impact is only a good bet in the context of hard-decision
>decoding.  Decoder implementations will certainly use soft-decision
>decoding.  After adding 64 check nodes, the graph of the LDPC code is
>changed.  Every bit node has additional extrinsic information (two) from
>added check nodes in every iteration step.  There are 1024 bit nodes.  This
>means that in every iteration step, there are an additional 2048 extrinsic
>information passed from check nodes.  At this time, as stated before, we
>have hunches on LDPC code performance (including the presence/absence of an
>error floor or slope change) but these expectations can only be validated
>through *simulations* proving BER=1e-12 is achievable/unachievable.  Other
>codes may be found and shown to be superior to LDPC(1024,833) but so far,
>is the only code with presented simulation results that demonstrate 1e-12
>BER is achievable (seki_1_0704.pdf).
>   - Scott
>Dr. Scott Powell
>Senior Manager, Ethernet PHYs
>Broadcom Corp.
>-----Original Message-----
>From: [] On
>Of sailesh rao
>Sent: Tuesday, August 31, 2004 5:14 PM
>Subject: Re: [10GBT] Validity of comparing PSDs to determine
>Hi All,
>Regulatory EMI tests are not conducted with worst-case alien NEXT
>environments set up in the range.
>Since we are comparing two TH-precoded (pre-equalized) PAM systems, and not
>grossly dissimilar systems, it is also sufficient to compare the amplitude
>of the pulse responses at the receiver input to understand which PAM system
>will be more robust in the EMI susceptibility tests. Therefore, there is no
>question that PAM8 will be 2.0dB to 3.9dB better than PAM12 in the 0-100m
>range. This advantage is critical in the installed base, where this
>regulatory test will be especially difficult to pass. In the installed
>this advantage ranges from 3.2dB to 3.9dB.
>With respect to the PAM12 constellation, please note that changing the
>constellation is tantamount to doing a heart bypass surgery in a data
>communications system. With respect to the holes, to draft an analogy,
>cheese is Swiss cheese any way we slice it, and it won't suddenly become
>American. After the two PAM12 hole repositioining schemes that I have
>examined only in the past few days, isn't it time we converged on the PAM8
>proposal instead of looking for other ways to patch up the holes in the
>PAM12 constellation?
>Furthermore, I've found another issue with the PAM12 proposal. I leave you
>now with a description of this Issue #6.
>Sailesh Rao.
>Issue #6: LDPC Code
>The (1024,833) LDPC code in the PAM12 proposal is constructed using
>Djurdjevic's method by placing blocks of parity check matrices in a
>systematic manner. In the course of this construction, as more and more
>blocks are used, the rate of the code decreases and the lower bound on the
>Hamming distance increases.
>The parity check matrix for the (1024,833) LDPC code used in the July PAM12
>proposal contains the parity check matrix for the (1024,845) LDPC code that
>the PAM12 proponents were using in the May presentation. Indeed, the parity
>check matrix for the (1024,833) code is obtained by adding 64 rows to the
>parity check matrix for the (1024,845) code. However, of the 64 rows added,
>52 of them are dependent (do-nothing) rows and only 12 of them are
>independent rows. This is why the number of information bits in the code
>only reduced from 845 to 833.
>It was reported in seki_1_0704.pdf that the (1024,845) code has a BER slope
>change in the 1E-9 BER range, which made it unacceptable. However, it was
>also reported that the (1024,833) code has no such BER slope change until
>the 1E-12 BER.
>There is only so much that 12 parity check equations can do, that the
>previous 179 parity check equations could not. Therefore, as an engineer, I
>wouldn't bet on the BER slope change for the (1024,833) code to occur well
>beyond the 1E-12 BER point.
>Please note that a 1E-12 BER in a 10GBASE-T system corresponds to 1 error
>every 1.6minutes. If we work very hard on the receiver and squeeze out a
>SNR margin to this 1E-12 BER, it would be truly tragic if that implied 1
>error every hour or so for our customers, due to this property of the LDPC
>code used in the system.
> >From: George Zimmerman <gzimmerman@SOLARFLARE.COM>
> >Reply-To: "IEEE P802.3an" <>
> >To:
> >Subject: [10GBT] Validity of comparing PSDs to determine
> >electromagnetic emissions
> >Date: Tue, 31 Aug 2004 14:33:24 -0700
> >
> >Scott -
> >Thank you for pointing out a simple fact that has made for amusing
> >reading while I've been busy with other things. Much of the discussion
> >here of PAM-12 vs. PAM-8 is well within the bounds of measurement
> >tolerances, and, as will be discussed below, can be flip-flopped by
> >design trades in these measurement tolerances.  This covers more than
> >just EMI, so here are some of the summary points:
> >
> >1) EMI emissions - We're way beyond 20 log (f) approximations.  We've
> >seen measured results in numerous presentations by multiple sources
> >(see cohen_1_0903.pdf for methodology and also a presentation by Scott
> >Powell).  Scott's right.  Differences in the PSDs are TINY when
> >compared with differences test setups, cabling configurations, and
> >measurement equipment can make - if a design is so close that the small
> >difference in the PSD is going to make it fail, it will fail for a
> >dozen larger reasons.  I'd be interested in hearing from Thuyen Dinh
> >which proposal looks easier to build magnetics for, keeping in mind
> >both emissions and insertion loss.
> >
> >2) Coded margin - we're asking for accuracy way beyond assuming the
> >same coding gain, or even running AWGN simulations on non-TH Precoded
> >constellations.  In my experience this can cloud things by 1 dB.
> >
> >3) Component and noise budgeting accuracy - The solarsep simulation
> >code with Salz-bounds provides an optimal DFE basis which can be used
> >as a line/environment limited bound on how well you can do.  It does
> >NOT provide a basis for evaluating sensitivity to noise which enters
> >the receiver at different points in the receive chain, hence it should
> >not be used by varying the AWGN level to try to measure sensitivity to
> >receiver noise, or for accurately tweaking the transmit and receive
> >filtering.  It can be adjusted for that, but for that you need to
> >become architecture-specific to a particular vendor, and add additional
> >linear transfer functions (such as magnetics, location of gain stages,
> >sources of noise and hybrid structure).  Net-net, the degradation from
> >optimal becomes significantly vendor-architecture dependent and
> >therefore less capable of multiple vendors tracking the results.
> >
> >4) EMI susceptibility - we're also way beyond looking at susceptibility
> >just as the partition separation of a sine wave on a single pair in the
> >absence of the consider ANEXT noise (e.g., Crane's test), because in
> >the presence of noise the cumulative coded errors will cause errors
> >first. Real EMI happens on all 4 pairs at once (to differing levels),
> >generally in the presence of noise.  In this situation, generally, the
> >code breaks down way before the partition difference (at least for
> >1e-12 BER).  As such, the EMI susceptibility must be measured in the
> >presence of full ANEXT.  This will require LDPC code simulations in the
> >presence of EMI ingress and ANEXT.  It must also respect the difference
> >in receive gain required by the PAM 12 and PAM 8 systems.  The answer
> >is not obvious, but I'm pretty sure I haven't seen it discussed on this
> >forum amidst all of the heated arguments, AND, it really requires
> >selection of a code to get it done.
> >
> >Long and short -
> >We need to get this matter decided already.  However, in order to
> >finalize this decision, we need not only PAM levels on the table, but
> >code and modulation rates nailed.  We also need to consider impact on
> >non-silicon components (cabling and magnetics, both of which are
> >stressed for bandwidth).  There are only a small number of points that
> >have been made that I think are unambiguous:
> >
> >1) The PAM-8 baud rate is faster, so for a fixed cancellation ratio, it
> >will require more digital processing.  If it, in fact, has higher
> >margins, these may be traded for implementation loss to ease the
> >burden.
> >
> >2) The PAM-12 constellation with a hole in it is inefficient.  This
> >fact is responsible for more than one of Sailesh's complaints.  The
> >cross constellation just moves the inefficiency around to a more
> >familiar form, but still suffers precoding losses.  A more efficient
> >mapping can be adapted.
> >
> >3) The framing proposed in the PAM-12 is complex and possibly can be
> >improved.
> >
> >4) The choice of transmit filter will govern the emissions more than
> >the number of PAM levels.
> >
> >5) The sensitivity of margin to baud rate is broad enough that the
> >ANEXT-limited performance differences in an optimal signal processing
> >sense are small.
> >
> >-george zimmerman
> >---
> >
> >-----Original Message-----
> >From: stds-802-3-10gbt@IEEE.ORG [mailto:stds-802-3-10gbt@IEEE.ORG] On
> >Behalf Of Scott Powell
> >Sent: Tuesday, August 31, 2004 11:29 AM
> >Subject: Re: [10GBT] Validity of comparing PSDs to determine
> >electromagnetic emissions
> >
> >I agree with Kishore's point that there is not a significant EMI
> >difference between the two proposals.  It's been shown a couple of
> >times on the reflector that one proposal is slightly better over
> >certain frequencies and
> >the same proposal slightly worse over others.  In response to a recent
> >posting on the importance of "slight" PSD differences:
> >
> > > "saliesh rao" wrote
> > >
> > > With regard to issue #1, I disagree that a 1dB difference in the
> >transmit
> >PSD is "well below the
> > > measurements and setup errors" for the emissions profile.  Anyone
> > > who
> >has
> >spent hours and days
> > > meeting the FCC/CISPR emissions profile in the range will know that
> >1dB
> >will make a huge difference
> > > between a pass and fail, since the likelihood of a fail would be
> >exponentially determined by this single dB.
> > >
> >
> >I would respectfully request that you take a look at ANSI standard
> >C63.4 before relying too heavily on the fractional dB emissions you
> >think you are measuring.  This standard document can be ordered from
> >the IEEE and describes the measurement accuracy that certified EMI test
> >ranges must comply to.  As you can see in section on page 20,
> >the "Acceptability
> >Criterion for Emission Test Range" is +/- 4dB.   Your EMI range may not
> >always "volunteer" this information but should if you ask.
> >
> >ANSI C63.4 standard means that if the same 10GBASE-T transceiver (+
> >cabling/connectors) is measured in 2 different *ANSI Qualified* EMI
> >test ranges and results differ but are within +/- 4dB of each other,
> >the measurements are both considered valid - even though they differ by
> >+/- 4dB. Susceptibility accuracy requirements for qualified ranges are
> >even worse.
> >These permitted inaccuracies are in addition to the inherent set-up
> >inaccuracies of finding the "worst case" antenna height, antenna
> >polarity,
> >and cable configuration/orientation for each frequency in the scan.
> >
> >I don't believe the small PSD differences shown so far between the two
> >proposals are significant enough to warrant choosing one over the other
> >based on expected emissions.  Again repeating my original message:  "it
> >has been clearly shown that the PAM-8 and PAM-12 proposals have similar
> >performance (slight edge for PAM-12), but the PAM-12 proposal offers the
> >advantage of a reduced operating frequency resulting in reduced power
> >and
> >reduced implementation difficulties."
> >
> >Regards,
> >   - Scott
> >
> >-----Original Message-----
> >From: [] On
> >Behalf Of Kishore Kota
> >Sent: Tuesday, August 31, 2004 7:21 AM
> >To:
> >Subject: [10GBT] Validity of comparing PSDs to determine electromagnetic
> >emissions
> >
> >
> >I would like to question the validity of comparing the power spectra
> >(PSD)
> >of various proposals to determine electromagnetic emissions. In earlier
> >postings on this reflector, such comparisons have been used to claim
> >superior emissions performance for PAM8. There are several unknowns
> >here:
> >
> >a) As a taskforce we haven't agreed to a model for how differential
> >signal gets converted to common mode, and
> >b) we do not have a model for how the common mode signal then gets
> >converted
> >to electromagnetic emissions.
> >
> >Simply comparing PSDs without applying such models would lead to
> >nonsensical conclusions. For instance, in the lower frequency band
> >1000BASE-T spectrum
> >is atleast 8db higher than either of the 10GBT proposals (because the
> >same
> >transmit power is spread over a smaller frequency band). However, this
> >does
> >not mean 1000BASE-T is tougher than 10GBT for emissions compliance
> >testing.
> >
> >Some comparisons of PAM12 and PAM8 spectra have been posted to the
> >reflector which include a 20logf model. Although this is a crude
> >approximation, it does seem to have some consensus in the group. By
> >playing games with this
> >model and lowpass filters, either proposal can be made to look slightly
> >better than the other. This suggests that there are no significant
> >differences between these proposals as far as emissions are concerned.
> >Obsessing over these differences just keeps us from answering the more
> >difficult question of whether either proposal is sufficient from an EMI
> >perspective.
> >
> >                 regards,
> >                 Kishore
>On the road to retirement? Check out MSN Life Events for advice on how to
>get there!

Express yourself instantly with MSN Messenger! Download today - it's FREE!