Transient Joule Heating in PrMnO3 RRAM enables ReLU type Neuron

Abstract

Hardware realization of scalable neurons and synapses are essential for the implementation of large scale Spiking neural network (SNN). In this paper, first, we propose, a novel transient Joule heating based the leaky-integrate and fire neuron (LIF) in scalable PrMnO3 (PMO) RRAM device experimentally. The Joule-heating based thermal runaway is utilized to achieve rectified linear unit (ReLU) voltage dependence of spiking frequency similar to a typical LIF neuron. Second, the Jouleheating hypothesis in PMO is validated by TCAD DC and transient simulations. PMO is extremely thermally resistive semiconductor (300x cf. Si) and hence enables low energy thermal dynamics. The excellent energy, area performance with a synapse in the same material system and thermal engineering makes PMO neuron attractive. Finally, PMO neuron shows software equivalent learning accuracy in SNN on the Fischer’s Iris dataset.