Thursday, June 13, 2019

Design of 56 Gb/s NRZ and PAM4 SerDes Transceivers in CMOS Technologies

J. Lee, P. Chiang, P. Peng, L. Chen and C. Weng, "Design of 56 Gb/s NRZ and PAM4 SerDes Transceivers in CMOS Technologies," in IEEE Journal of Solid-State Circuits, vol. 50, no. 9, pp. 2061-2073, Sept. 2015.

Disclaimer: I've only read up to and including section III but that includes the PAM4 design, which I am interested in at the time.

Great, well written paper.  This is written by a university group so I felt like they shared a lot of useful information and were not trying to hide details like it feel like if the group was industry based.


  • They have a great (first order) explanation as to when you want to choose PAM4 over NRZ.  
  • The concept is, the PAM4 eye is 1/3 smaller than the NRZ eye. db(1/3)  = -9.5.  For 56 Gbps, the Nyquist for NRZ is 28 GHz and 14 GHz for PAM4.  Therefore, if the difference in Insertion loss between 28 and 14 G is less than 9.5 dB, you should do NRZ, but if it is larger then PAM4 is a viable option.
Figure 2

  • PAM4 Transmitter
    • "the number of FFE taps is usually limited to 4 in order to minimize the parasitics".
    • May need DLLs to neutralize skews between the multiple data stream, due to the use of clock multiplexing (a half rate clock, quarter rate clock, etc).  
      • skew may be an interesting impairment to model.  This would be a form of DCD
    • Summer with tuneable weighting combiner to adjust the level spacing.
      • These are needed to compensate for a non-linear optical impairment.
      • We could model this behavior in a MATLAB function block but would only be useful if we also had a EML (electroabsortion-modulated) laser block that would create the impairment that the tunable weights would be compensating for.  
        • Unless an AMI model injected this EML laser impairment, I see little point in including a summer with tunable weights in an AMI model.
    • Transmitter includes peaking inductor to increase the edge rate of the transitions.
Figure 3



  • PAM4 Receiver
    • A pure linear CDR.  They say that it naturally handles multiple levels and is some ways easier to implement than a bang-bang CDR.
    • Preamp/Equalizer
      • They do slightly equalize the signal differently depending on if it goes to the middle (0 Volt threshold) slicer or the upper/lower slicers..
    • PAM4 slicers, linear CDR and DFE
      • 3 slicers at the 3 PAM4 threshold values
      • Linear CDR
      • Apply DFE separately to the three slicers
    • Question.  Why is the DFE after the slicer?  I guess I usually combine the slicer with symbol decoder.  Here these functions are separated.  This allows for a soft decision (slicers, or a 2 bit ADC), followed by some DSP (through the DFE) and finally the PAM4 decoder and retimers make the bit decisions.
      • This type of circuit is possible since they have a high precision linear CDR.  Then they can operate on the data samples with high confidence.
Figure 7