Re: 64/66 control code mapping
Dear Rich,
> Rich Taborek <rtaborek@xxxxxxxxxxx> writes
> I realized yesterday while analyzing SLP as a XAUI code that your
> Albuquerque presentation of PCB trace analysis of 8B/10B @ 3.125 Gbaud
> vs. scrambles @ 2.5 Gbaud shows better eyes for 8B/10B than
> scrambling for 5 out of 6 cases.
>
> Ignoring all slides at >3.125 Gbaud since they are not relevant to
> XAUI trace s, the 8B/10B eye is superior to the slower scrambled eye
> in:
I don't agree. You've picked a very biased definition of "superior".
Usually an eye margin is defined in terms of both time and voltage. You
seem to be focussing solely on voltage and completely disregarding the
time axis.
Lets look at each slide:
slide EyedeltaT EyedeltaV Description
----- -------- --------- -----------
15 295ps 355mV DC 8b/10b 3.125G test setup
14 373ps 347mV DC scrambled 2.5G test setup
10 297ps 355mV AC 8b/10b 3.125G test setup
9 372ps 311mV AC scrambled 2.5G test setup
28 367ps 313mV DC scrambled 2.5G 5.5" diff pair
29 281ps 315mV DC 8b/10b 3.125G 5.5" diff pair
20 289ps 311mV AC 8b/10b 3.125G 5.5" diff pair
19 359ps 279mV AC scrambled 2.5G 5.5" diff pair
25 278ps 321mV AC 8b/10b 3.125G 11" diff pair + 2 via
24 344ps 311mV AC scrambled 2.5G 11" diff pair + 2 via
27 191ps 165mV AC 8b/10b 3.125G 22" diff pair + 4 via
26 242ps 203mV AC scrambled 2.5G 22" diff pair + 4 via
> The only slides which show the scrambled eye to be superior is for the
> case of a 22" differential pair and 4 vias in slides 26,27,33.
My reading is that the scrambled code is superior in the time opening
for ALL the cases and has superior voltage margin in the worst case
condition.
In the voltage axis, both codes have quite useable eye openings until
the worst-case interconnect. Once we reach the long PCB trace, it is
clear that 8b/10b has fallen off of a cliff.
So, I get a completely opposite reading from the data, Rich. Scrambling
is only slightly degraded in the voltage axis due to BLW - and this is
for the short distance traces. For electrical links, there is no large
gaussian noise term, so openings bigger than ~50mV generally have
unmeasurable BER. By this reasoning, both codes show extremely wide
voltage margins.
You can see that the difference is only due to capacitive coupling by
comparing slides 20/19 with 28/28. In the latter case, when the links
are DC-coupled, there is only a trival 2mV difference between the two
codes.
The presentation doesn't say what the AC coupling frequency was. The
BLW penalty can be reduced to zero if two chips are DC coupled on the
PCB. It can also be reduced arbitrarily by increasing the capacitor
size.
> In a nutshell, the "eye" analysis seems to be either inconclusive or
> questionable since 8B/10B seems to "fall off a cliff" with respect to
> scrambling somewhere between 11"/2 vias and 22"/4 vias. Where is the
> cliff and what accounts for it? This is why I'm having great
> difficulty with your suggestion to "slow down XAUI a bit".
The cliff is can be explained by noting that a PCB interconnect has an
impulse response with a certain half-width in time. When the bit time
is larger than the impulse response width, you see essentially no
amplitude degradation at the mid-eye.
Once the bit time gets shorter than the channel impulse response width,
010, or 101 patterns will have reduced amplitude.
Because 8b/10b has a narrower eye it will hit this "wall" 25% earlier
than a scrambled code. This is certainly the effect in slide 27 as
can be seen by the distinctive "double trace" in the eye diagram. The
inner of the two traces is caused by the reduced amplitude of the 010 and
101 data patterns.
So, my conclusion is that this data clearly shows that scrambling
provides a considerable margin improvement for the worst-case link, and
has negligable penalty for short links.
As I said in Kauai, I think it would be an improvement to replace 8b/10b
with a higher efficiency code. However, it is not yet clear whether it
is worth the pain. To help us determine that, we need to keep our eyes
and ears open and let the group dispassionately evaluate the pros and
cons of the various options. It will certainly be painful to consider
dropping our old friend: 8b/10b.
My belief is that ~4-5 Gb/s is near the ultimate skin-loss limit for 0.5
Meter NRZ signalling with 4 mil differential copper traces. You can
push this by equalizing, reducing trace length, or by reducing layout
density by widening the traces.
It is not clear whether 3.125G is pushing too aggressively for low cost
implementations. For this we will have to trust the experience and
experimentation of the various 10GbE participants over the next few
meetings.
Best regards,
--
Rick Walker