40Gbs Optical Transport Milks DSP Technology

1
40Gb/s Optical Transport Milks DSP Technology

• John Sitch
• Nortel Optical Systems
• sitch_at_nortel.com

2
Contents

• 40Gb/s challenge
• Electronic signal processing
• Smart algorithms make it possible
• Data converters are key
• Technology compromises
• Challenges summary

3
40Gb/s challenge why not do 40 like we did 10
i.e. NRZ?

• Chromatic dispersion tolerance 90ps/nm or lt 5
  miles
• If we compensate, we need to get within 90ps/nm
• PMD (polarization mode dispersion) - reach
  depends on quality of fiber. E.g. to reach
  1000km, we need fiber with 0.1ps/sq.rt.(km).
• Spectral occupancy apart from efficiency
  considerations, many networks are using 50GHz
  channel spacing ROADMs (reconfigurable optical
  add-drop muxes), having only -16GHz for the
  signal, so we can transit only 3 ROADMs before
  the signal gets too degraded.
• High analog bandwidth leads to gold brick V
  connector packaging expensive electronics
  expensive optics

4
Electronic Signal Processing

• Light is best at carrying data
• It goes long distances without very much
  happening
• Electronics is what you need to make things
  happen
• By using electronic signal processing we can
  replace expensive optics with inexpensive ASICs
• That sounds like motherhood so why is it news?
• Technology, Technique and Need have all come
  together

5
Technology, Technique and Need

• 130 90 nm CMOS are the first technologies able
  to handle 10 40 Gb DSP with realistic power
  dissipation
• Wireless experience
• has taught us DSP albeit at lower data rates
  (and very much lower carrier frequencies!)
• also a wide variety of spectrally efficient
  modulation schemes e.g. PSK, QAM
• Post-bubble winter is finally over network
  traffic is growing, as is router port rate.

6
Make the algorithm fit the need
Example chromatic dispersion

• Chromatic dispersion in which delay is a
  function of frequency spreads each symbol in
  time.
• CD grows as (baud rate)2 distance
• MLSE (maximum likelihood sequence estimator) can
  fix it, but grows as 2N (N number of symbols -
  N5 is tough)
• CD is a linear problem if carrier phase is
  considered, so we can use a linear filter,
  growing as N2 handy when N 100 ( to take
  carrier phase into account, we need a transmitter
  or receiver with a complex signal path).

7
Converter Resolution Penalties Strong FEC
brings Good News
ROSNR Penalty (dB)
DAC resolution - Tx compensation, intensity
detection (linear CD filter)
ADC resolution - intensity detection, Rx
compensation (non-linear MLSE)
2
4
6
D/A or A/D resolution (effective bits both 2
samples/baud)
In both of these examples were assuming strong
FEC (high raw BER), hence low SNR, so going to
more than 5 effective bits brings diminishing
returns
8
ADC options (Weve done both)
Flash input fanned out to 2N
comparator-latches, thermometer output
Time interleaved input fed to M samplers, each
with a lower rate ADC
9
Converter Tradeoffs

• At data rates of 10Gbs and above, resolution is
  expensive
• Fill the converter
• But not too full, as clipping is harmful
• Good AGC level control algorithm needed
• Oversampling brings benefits
• Higher time resolution in filters
• Lifts sampling phase restrictions
• Use extra bandwidth to avoid aliasing noise,
  increase effective resolution

10
Fit the technology to the job (Nortel products)

• Bipolar gives us the best analog, but BiCMOS
  logic lags CMOS
• 10G Tx equalizer uses 130nm BiCMOS for the best
  DACs
• 40G DP-QPSK Rx equalizer has 4, 6-bit, 23Gs/s
  ADCs in 90nm CMOS (see paper 30.3, this
  conference)

DAC1
5 M gates EQ
DAC2
ADC1
14 M gates EQ
ADC2
ADC3
ADC4
11
Challenges

• Limited CMOS analog performance yes Virginia,
  you can have 1B transistors, but none of them
  work worth a damn.
• Power supplies distribution you want a quiet
  1V while the logics drawing as much current as
  your starter motor, with sharp edges, to boot.
• Data converters, given the first two bullets. You
  want state-of-the-art performance with minimal
  space, power cost ( did I mention
  auto-calibration?)
• DSP algorithms digital design its still
  mostly uncharted territory, with lots of Matlab
  lengthy design optimization.

12
Summary

• Electronic signal processing leads to more
  flexible, better performing cost-effective
  optical networks.
• Understanding optical effects helps us to design
  appropriate algorithms and circuits. Plenty of
  scope for silicon innovation!
• Todays technology efficiently handles 10 40Gb
  signals 100Gb tomorrow!

13
(No Transcript)