Paul Kwon

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2023-215

August 11, 2023

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-215.pdf

An ever-increasing demand for high data rate wireline links must be met to support the continual scaling of computing and communication systems. Typical serial link architectures use feedback-based channel equalization schemes, which can be challenging and even infeasible to realize for data rates beyond 100 GBaud/s. To alleviate this feedback-induced latency bottleneck, this thesis explores feedforward equalizers inspired by the maximum likelihood sequence estimation (MLSE) algorithm. Targeting short-reach, die-to-die links, a 1-tap fully feedforward MLSE architecture is shown to achieve comparable error statistics with the conventional 1-tap decision feedback equalizer (DFE).

A 160 Gb/s NRZ receiver implementing the 1-tap MLSE equalizer is taped out in a 16 nm FinFET process to evaluate the future promise of the proposed approach. The feedforward MLSE is time-interleaved to achieve the desired throughput. Current integration techniques provide energy-efficient analog latches used in the MLSE datapath. The datapath is designed using the Berkeley Analog Generator (BAG) framework. Both the top level datapath and its subblocks are created using parameterizable circuit schematic and layout generators. Measurement and design scripts identify the impact of inter-subblock routing parasitics, facilitating agile subblock design iteration with top level floorplanning in mind.

As the channel characteristics are unknown a priori, the coefficient settings of the MLSE must be adapted in real time. An adaptation scheme for the feedforward MLSE equalizer is proposed with separate loops for each interleaved MLSE slice to account for process variation. The receiver is taped out along with a corresponding 160 Gb/s NRZ transmitter and tested over an on-package loopback channel of 8.5 mm, which incurs 3 dB of loss at the Nyquist frequency. The receiver is simulated to operate at an energy efficiency of 2.08 pJ/bit, with the datapath itself consuming 0.72 pJ/bit.

Advisors: Elad Alon and Vladimir Stojanovic

---

BibTeX citation:

@phdthesis{Kwon:EECS-2023-215,
    Author= {Kwon, Paul},
    Title= {Feedforward MLSE Equalization for High Speed Serial Links},
    School= {EECS Department, University of California, Berkeley},
    Year= {2023},
    Month= {Aug},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-215.html},
    Number= {UCB/EECS-2023-215},
    Abstract= {An ever-increasing demand for high data rate wireline links must be met to support the continual scaling of computing and communication systems. Typical serial link architectures use feedback-based channel equalization schemes, which can be challenging and even infeasible to realize for data rates beyond 100 GBaud/s. To alleviate this feedback-induced latency bottleneck, this thesis explores feedforward equalizers inspired by the maximum likelihood sequence estimation (MLSE) algorithm. Targeting short-reach, die-to-die links, a 1-tap fully feedforward MLSE architecture is shown to achieve comparable error statistics with the conventional 1-tap decision feedback equalizer (DFE).

A 160 Gb/s NRZ receiver implementing the 1-tap MLSE equalizer is taped out in a 16 nm FinFET process to evaluate the future promise of the proposed approach. The feedforward MLSE is time-interleaved to achieve the desired throughput. Current integration techniques provide energy-efficient analog latches used in the MLSE datapath.  The datapath is designed using the Berkeley Analog Generator (BAG) framework. Both the top level datapath and its subblocks are created using parameterizable circuit schematic and layout generators. Measurement and design scripts identify the impact of inter-subblock routing parasitics, facilitating agile subblock design iteration with top level floorplanning in mind.

As the channel characteristics are unknown a priori, the coefficient settings of the MLSE must be adapted in real time. An adaptation scheme for the feedforward MLSE equalizer is proposed with separate loops for each interleaved MLSE slice to account for process variation. The receiver is taped out along with a corresponding 160 Gb/s NRZ transmitter and tested over an on-package loopback channel of 8.5 mm, which incurs 3 dB of loss at the Nyquist frequency. The receiver is simulated to operate at an energy efficiency of 2.08 pJ/bit, with the datapath itself consuming 0.72 pJ/bit.},
}

EndNote citation:

%0 Thesis
%A Kwon, Paul 
%T Feedforward MLSE Equalization for High Speed Serial Links
%I EECS Department, University of California, Berkeley
%D 2023
%8 August 11
%@ UCB/EECS-2023-215
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-215.html
%F Kwon:EECS-2023-215