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

Re: [BP] Informative Channel



Hello, Charles,

I agree your "nits" points.

1. You are right for 4/pi (2dB) factor in a case of square wave at
   TX output.
   In actual case, for example there is 0.5pF stray capacitance
   including TX output transistors, ESD protectors, etc.,
   3rd harmonic will be -13dB referring to 5GHz sine wave, and the
   factor between composite p-p vs sine wave is almost 1dB.
   This is the reason why I ignored the factor previously. But I
   agree that still 1dB difference exits.

2. I agree that RX equalization need to boost at least 5GHz.

4. Length of ..0101.. can be 31~32bit at PRBS31. In the case,
   its spectra distributes approximate 250MHz width. I'm not sure
   that it is wide enough to cover dip and peak of frequency ripple.
   I understand 5dB is too match. please teach me how many dB is 
   reasonable.
   I guess 0101010 (8bit length) spreads approximate 1GHz, and ok to
   cover dip and peak.


I am still wondering BER at lower S/N case.


M. Sugawara

>sugawara,
>
>     Good over all analysis, i will pick a few nits below, but i agree
>with the general overall results.
>
>     In answer to your question, yes.  If we want our silicon to work
>at 10Gb/s over realistic backplanes, we will have to make our receivers
>work at quite small levels, even with cross talk which will not be much
>smaller.  This may mean a pre-amp, DC offset cancellation, etc.  Early
>versions have high power requirements and command a premium price but
>most likely, if KR is successful, competitive forces will bring these
>down.
>
>      Nits:
>
>      1.  At the end of the channel the alternating 10 pattern will
>          have its higher harmonics removed.  The harmonics actually
>          reduce the p-p amplitude so the remaining sine wave will be
>          about 2dB (4/pi) higher than you predict.
>      2.  Rx equalization is likely to enhance the alternating 10
>          pattern some.
>      3.  At the Nyquist rate, 5.15625GHz the equation 69.2 is down
>          26dB :-(.
>      4.  Random data will not sustain an alternating 10 pattern very
>          long, so there will be no spectral line to fall into a dip
>          in the transfer characteristic.  The finite length of the
>          alternating region should spread the spectrum enough to
>          allow it to be spread across a region which includes both
>          dips and peaks.
>      5.  The compliance test channel defined in annex 72A will be
>          chosen by the silicon vendor to make his silicon look as
>          good as possible while still staying below the spec line.
>          You should be able to avoid ISI loss worse than the value
>          shown as "typical", i certainly intend to make my test
>          channel as close to the spec limit as possible.
>
>                    charles
>
>
>-- 
>|--------------------------------------------------------------------|
>|       Charles Moore
>|       Agilent Technologies
>|       ASIC Products Division
>|       charles_moore@agilent.com
>|       (970) 288-4561
>|--------------------------------------------------------------------|
>
>> Hello, all,
>> 
>> As silicon designer point of view, RX input level is very sensitive
>> issue. If it is too low, we have to install pre-amp, DC offset
>> chancellor, etc. and they consume more power and silicon area.
>> 
>> 
>> Let's try simplified consideration;
>> ..01010101.. signal somewhere PRBS consists of 5GHz sine wave
>> (and small 3rd harmonics...). Pre-emphasis should be set most strong
>> level, the other word it should be full swing 1Vppd typ at TX side.
>> 
>> In a case of equation 69.2 per draft 0.91, -25.3dB at 5GHz,
>> ..01010101.. signal of 1Vppd TX makes 54mVppd or 27mVpp single-ended.
>> 
>> However
>> 1. TX minimum amplitude is spec'ed 0.8Vppd, so any RX should receive
>>    the signal from the TX (effective -2dB).
>> 2. Return loss max is 5dB at equation 72-2 per draft 0.91. The
>>    mismatch makes 1~2dB signal level loss at each of TX-channel and
>>    channel-RX. Let's assume total -3dB min.
>> 3. Frequency ripple may be allowed -5dB min at 5GHz in the discussion
>>    on Jun 1. 
>> Assumption of above 3 effects are -10dB min !!
>> Then the really worst case of eye height at RX input will be 18mVppd
>> or 9mVpp single-ended by ..01010101.. signal !!
>> It doesn't include crosstalks, noise and jitter.
>> 
>> Should RX designers really consider such small level ??
>> 
>> 
>> Another values;
>> Figure 72A-2 in draft 0.91 shows -28.86dB "typical" ISI loss at
>> 5.156GHz. When we use the number, RX input will be 28.8mVppd or
>> 14.4mVpp single-ended from 1Vppd TX.
>> If we add above 3 effects, it will be 9.6mVppd or 4.8mVpp single-
>> ended.
>> If we consider maximum ISI loss rather than "typical", 
>> Ooooom...
>> 
>> Anyway these kind of small values of RX input may be smaller than
>> peaks of noise and crosstalk.
>> 
>> 
>> M. Sugawara
>> NEC Electronics America