JSNPE: A Digital Superconducting Spiking Neural Processing Element

Abstract

Superconducting electronic devices are considered a compelling candidate for designing neuromorphic hardware due to their similar behavioral characteristics. Previous works have shown high potential for applying them to design spiking neural networks. However, convincing schemes remain lacking. This article proposes a Josephson junction (JJ)–based spiking neural processing element, JSNPE, a digital scheme composed of a neural and synaptic module. The neuron is a concise quasidigital circuit with three JJs. It operates asynchronously and realizes the integrate-and-fire model. The synapse is a digital circuit constructed with superconducting rapid single flux quantum devices. Compared to previous designs, JSNPE facilitates network scalability, exhibits higher computational resolution, and enables easier weight configuration. To validate the design, we conducted inference tasks on the IRIS and MNIST datasets using WRspice simulation, with inference parameters obtained through software-based training. The analysis indicates that the inference network can achieve a classification accuracy of 96.7% on IRIS, while for MNIST, the classification accuracy for digit “0” reached 94%. We also evaluated the performance and energy efficiency of JSNPE, demonstrating that JSNPE could achieve significant improvements in both metrics compared to complementary metal oxide semiconductor designs.