Thermal Instability Compensation of Synaptic 3D Flash Memory-Based Hardware Neural Networks With Adaptive Read Bias

Ambient temperature variations caused by extensive operations in hardware neural networks (HNNs) are a critical issue that significantly degrades performance. In this work, HNNs based on synaptic 3D flash memory with high area and energy efficiency are proposed. The impact of ambient temperature on synaptic weights composed of pairs of synaptic devices was investigated. Our proposed thermal instability compensation method using adaptive read bias restores distorted outputs with a low hardware burden. The effective weight modulation with adaptive read bias successfully recovers HNN performance to near baseline even at a high ambient temperature of 100°C.