LIF Neuron Based on a Charge-Powered Ring Oscillator in Weak Inversion Achieving 201 fJ/SOP

Abstract

This letter presents the experimental results of a leaky-integrate-and-fire neuron (LIF) neuron based on time-domain analog circuitry. This kind of neuron is the core of spiking neural network (SNN) used in edge applications. Edge applications require power-efficient neuron designs whose power consumption is extremely low when idle, and low when in dynamic operation. The proposed neuron complies with the aforementioned requisites by transforming the voltage-based threshold of conventional LIF neurons into a time domain threshold on a quadrature oscillator. In conjunction with a charge-sharing integrator, the proposed neuron shows an energy efficiency of 201 fJ/SOP implemented in $0.13\mathbf {\mu m}$ process.