Development of a Neuromorphic Network Using BioSFQ Circuits

TL;DR

This paper presents a novel neuromorphic network built from bioSFQ superconductor circuits, demonstrating a 3x3 array of artificial neurons capable of analog functions, aiming to leverage superconductor electronics for energy-efficient computing.

Contribution

It introduces a new neuromorphic circuit design using bioSFQ superconductor components, combining neural network architecture with superconductor technology.

Findings

Successful fabrication of bioSFQ neuron array

Demonstrated analog memory and read/write interface

Close resemblance to a three-layer perceptron

Abstract

Superconductor electronics (SCE) appear promising for low energy applications. However, the achieved and projected circuit densities are insufficient for direct competition with CMOS technology. Original algorithms and nontraditional architectures are required for realizing SCE energy advantages for computing. Neuromorphic computing (NMC) is a commonly discussed deviation from conventional CMOS digital solutions. Instead of mimicking a conventional network of artificial neurons, we compose a network from the previously demonstrated single flux quantum (SFQ) electronics components which we termed bioSFQ. We present a design and operation of a new neuromorphic circuit containing a 3x3 array of bioSFQ cells - superconductor artificial neurons - capable of performing various analog functions and based on Josephson junction comparators with complementary outputs. The resultant asynchronous network closely resembles a three-layer perceptron. We also present superconductor analog memory and the memory Read/Write interface implemented with the neural network. The circuits were fabricated in the SFQ5ee process at MIT Lincoln Laboratory.