PAM-5 at 5 Gbaud
Oscar,
I have some question marks regarding your presentation in Dallas:
"10 Gb/s PMD using PAM-5 modulation"
by Oscar Agazzi
Broadcom
a) 5 GHz equalizer
You use in your simulations a Decision Feedback Equalizer (DFE) at 5 GHz. You
mention, to support your proposal, that DFEs are also used in Fast Ethernet
and 1000BASE-T. However, the latter DFEs run at 125 MHz (8 nsec baud period).
The DFE that you are proposing must run 40 times faster (200 psec baud
period).
A DFE has a feedback loop (slide # 15 in your presentation) that consists of
at least one adder, a 5-level slicer and the internal delay of one flip-flop.
The serial operations in this feedback loop (addition + slicer + internal
delay of the flip-flop) have to be completed within one baud period, in this
case 200 psec.
There was a very heated debate within the 1000BASE-T Task Force two years ago
whether the DFE could be implemented at 125 MHz. I remember that during these
debates you and Broadcom vehemently sustained that it would be extremely
difficult to implement the feedback loop in 8 nsec. Now you propose to
implement it in 200 psec.
I have doubts whether this DFE could be moved from the world of simulations
into a real implemented system. And in CMOS, as slide # 2 of your
presentation seems to suggest. Even using parallel processing.
For comparison, the architecture I proposed, PAM-5 4-WDM at
1.25 Gbaud, using the 1000BASE-T PCS, (see my presentations
in Kauai and Dallas) has two options:
1) Viterbi decoding, with 6 db coding gain
2) symbol-by-symbol decoding, with 3 db coding gain
There is already a significant amount of previous work
on fast parallel processing of Viterbi decoders that can
be found in the open literature. See, for example, Ref. 5
in my presentation in Kauai:
H. David, G. Fettweis and H. Meyr
"A CMOS IC for Gb/s Viterbi decoding: System design
and VLSI implementation"
IEEE Trans on VLSI Systems, vol 4, pp 17-31, March 96
Specifically, following the detailed guidelines of this Ref,
the complete Viterbi decoder can be implemented using a
312.5 MHz clock (3.2 nsec clock period). This is also a very
handy clock, since we need it anyway in the parallel interface.
These 3.2 nsec are enough to implement the path metrics
update, which is the bottleneck in fast Viterbi decoders.
However, I also suggested to you that we could propose
in the 10 GbE Task Force to use the 3-dB coding option
of this PCS, if you prefer. The 3-dB coding option does not
use Viterbi decoding.
The burdens on the receiver analog front end of your proposal are even more
daunting.
b) 5 GHz ADC
The main claim of your proposal is that it can reach 500 meters of installed
multimode fiber (500 MHz*km bandwidth)
At 5 Gbaud and 1300 nm wavelength the optical eye pattern of PAM-5 is
completely closed even before reaching the 200 meters link length.
At 500 meters the ISI (Inter Symbol Interference) is as bad or worse than the
ISI we get in Fast Ethernet using 100 meters of cat-5 Copper wire. In Fast
Ethernet we needed a true 6-bit (64 levels) ADC for the DFE to be able to
deal with this strong ISI.
Slice # 15 of your presentation shows an ADC.
I think that you will have to use at least a 6-bit ADC in your system. This
also looks extremely difficult to implement at 5 GHz. For example, in the
last International Solid-State Circuits Conference held this month in San
Francisco, the maximum sampling rate achieved by a nominal 6-bit CMOS ADC was
800 Msamples/s (only 5-bit effective using a 200 MHz signal). It was
fabricated in a 0.25 um process.
For comparison, PAM-5 4-WDM at 1.25 Gbaud does not have
any ISI up to 400 meters and uses an 18 level "soft slicer".
This is barely a 4-bit ADC. And it is sampled at 1.25 Gbaud.
All the simulations I presented in Kauai were obtained using
this simple 18-level ADC. And, as I showed in Part IV of the
presentation, 18 levels are enough to reach an actual coding
gain close enough to the ideal.
This should not come as a surprise. It is a well known fact
that Viterbi decoders for binary encoded information (PAM-2)
need very simple "soft-slicers" to get most of the coding
gain of the convolutional code. A "soft-slicer" for PAM-2
coding needs only 8 levels to get a performance near to the
ideal Viterbi decoder. See, for example:
J. A. Heller and I. M. Jacobs
"Viterbi decoding for satellite and space communications"
IEEE Trans on Commun Tech, vol COM-19, pp 835-848,
October 1971
S. B. Wicker
"Error control systems for digital communications and
storage"
Prentice Hall, 1995
(A "hard-slicer" is the standard n-level slicer for PAM-n.
A "soft-slicer" uses more intermediate levels to get more
accurate decisions).
c) Dynamic range of the Receiver Analog Front End
You will need 5 Gbaud Transimpedance Amplifiers (TIA) and AGCs (slice # 15 of
your presentation). What should be the needed dynamic range of these blocks ?
A 6-bit ADC means about 36 dB dynamic range:
20*log(64) = 36 dB
However, you would need to add some margin in your design of the analog front
end. This means, you will need TIAs and AGC at 5 Gbaud with a dynamic range
of about 41-46 dB. This also looks extremely adventurous to propose in CMOS
(and I would add, in any technology).
On the other hand, using PAM-5 at 1.25 Gbaud, and
remembering that:
20*log(18) = 25 dB
we will need TIAs and AGCs at 1.25 Gbaud with a
dynamic range of only 30-35 dB.
All the above place an interrogation mark on the technical viability of the
serial PAM-5 approach at 5 Gbaud.
I doubt if the HSSG members were aware of these technicalities when they
rushed to a strawpoll in Dallas, specially since you did not post your
presentation in the web site before the Dallas meeting for a peer preview.
This did not give the HSSG members a fair chance to take a critical look at
your proposal.
Jaime
Jaime E. Kardontchik
Micro Linear
San Jose, CA 95131
email: kardontchik.jaime@xxxxxxxxxxx