Reliability of NAND Flash Memory as a Weight Storage Device of Artificial Neural Network

Abstract

NAND flash memory is a popular choice for storing a large number of model weights of an Artificial Neural Network (ANN) in many Internet of Things devices and edge computing applications. While being used as a weight storage device, the bit error rate of flash memory plays a significant role in the performance of the ANN application. In this paper, we propose two novel weight storage method in NAND flash memory, which will significantly suppress the effects of bit error rate on the ANN’s model weights and its performance. The proposed weight storage methods utilize the NAND flash’s page-to-page variability in favor of storing the model weights on more reliable pages. In order to demonstrate the benefit of the proposed method, we perform different reliability experiments on commercial flash memory chips containing the trained weight values of an ANN application. The experimental evaluation shows that the proposed method outperforms the traditional weight storage method and ensures prediction accuracy more than 90% even with a bit error rate exceeding 1% value.