A Deep Dive into the Design Space of a Dynamically Reconfigurable Cryogenic Spiking Neuron

TL;DR

This paper analyzes the design space of a superconducting memristor-based cryogenic spiking neuron, demonstrating enhanced reconfigurability, tunable spike frequency and amplitude, and robustness through systematic sensitivity and Monte Carlo analyses.

Contribution

It introduces a novel reconfigurable cryogenic neuron design using SM and SNW, with improved tunability and robustness over previous fixed-resistance designs.

Findings

Reconfigurability improved by ~70% with SM resistance variation.

Spike frequency modulated up to ~3.5 times via bias current.

Spike amplitude achieved at ~1V and ~1.8V, with robustness to variations.

Abstract

Spiking neural network offers the most bio-realistic approach to mimic the parallelism and compactness of the human brain. A spiking neuron is the central component of an SNN which generates information-encoded spikes. We present a comprehensive design space analysis of the superconducting memristor (SM)-based electrically reconfigurable cryogenic neuron. A superconducting nanowire (SNW) connected in parallel with an SM function as a dual-frequency oscillator and two of these oscillators can be coupled to design a dynamically tunable spiking neuron. The same neuron topology was previously proposed where a fixed resistance was used in parallel with the SNW. Replacing the fixed resistance with the SM provides an additional tuning knob with four distinct combinations of SM resistances, which improves the reconfigurability by up to ~70%. Utilizing an external bias current (Ibias), the spike frequency can be modulated up to ~3.5 times. Two distinct spike amplitudes (~1V and ~1.8 V) are also achieved. Here, we perform a systematic sensitivity analysis and show that the reconfigurability can be further tuned by choosing a higher input current strength. By performing a 500-point Monte Carlo variation analysis, we find that the spike amplitude is more variation robust than spike frequency and the variation robustness can be further improved by choosing a higher Ibias. Our study provides valuable insights for further exploration of materials and circuit level modification of the neuron that will be useful for system-level incorporation of the neuron circuit