Deep learning for NAND read bias decision

Abstract

The NAND flash, which is used as a storage device in various application fields, it is essential to improve the performance of solution algorithms for high memory reliability. As the memory density increases, changes in physical characteristics and distortion of threshold voltage distribution occur, making it increasingly difficult to accurately read stored data bits from memory. In accordance with the development of deep learning technology, we aim to improve the reliability of memory by applying an AI algorithm in the process of responding to threshold voltage distribution deterioration. This paper proposes a deep learning-based read bias voltage estimation algorithm to accurately read data bits stored in NAND flash in response to the threshold voltage distribution distortion. The proposed algorithm has a two-stage decision structure that determines the read bias voltage after reconstructing the threshold voltage distribution based on sparse read trial information. After extracting the characteristics of the distribution that can express the shape of the entire distribution, we proposed a method of restoring the distribution by combining it with the read trial information. In addition, it was confirmed that the proposed algorithm well exceeds the performance of conventional non-learning techniques using datasets collected from actual NAND. It is expected that the proposed algorithms, which predict read bias voltages quickly and accurately with only a few read trials, can significantly improve memory reliability when applied to actual NAND systems.