Radiation Tolerance of 3-D NAND Flash Based Neuromorphic Computing System

Abstract

In this paper we show the effectiveness of a multi-level-cell 3-D NAND flash chip as a weight storage device for a neuromorphic computing system under radiation environment. We find that the error-correction codes can be avoided for storing model weights in 3-D NAND for enabling low-power computing applications without sacrificing much accuracy (radiation dose <10k rad). Additionally, radiation induced BER data shows layer-to-layer variations, which can be utilized in favor of improving neural network’s accuracy.