An Adaptive Read Control Voltage Scheme for Reliability Enhancement of Flash-Based In-Memory Computing Architecture for Neural Network

The storage reliability is critical for flash memory based computing in-memory (CIM) architecture for Convolutional Neural Network (CNN). In this paper, we constructed a CIM scheme based on the Nor Flash array (NFA). We conducted simulations to investigate the impact of threshold voltage distribution and drift of Flash memory cells on the recognition accuracy for various CNN architectures based on the CIM schemes. Based on the reliability study, we proposed a novel compensation scheme to effectively mitigate the impact of threshold voltage drift and evaluated the effectiveness of the proposed scheme by recognition accuracy evaluation.