Spiking Neural Network Equalizer with Fast and Low Power Decoding for IM/DD Optical Communication

Neuromorphic computing based on spiking neural networks (SNN) realized in CMOS mixed-signal circuits promises lower power consumption than conventional digital computing. This makes SNNs an interesting technology for low-footprint optical transceiver ASICs. Arnold et al. proposed a non-linear SNN equalizer and demapper architecture outperforming a linear digital equalizer for a simulated IM/DD link. In this work, the spike decoding layer providing the demapper decision is optimized with respect to throughput and power. Four decoding methods are compared, namely max-over-time membrane (MOTM), end-of-time membrane (EOTM), time-to-first-spike (TTFS), and spike rate decoding. Optimized EOTM decoding is found to provide the fastest decision and the lowest spike rate, improving throughput and power consumption by factors 4 and 10, respectively, compared to previously used MOTM decoding.