Machine-learning-based read voltage estimation for NAND flash memory systems without knowledge of retention time

Abstract

Since the increment in program/erase (P/E) cycles and retention time distort the threshold voltage distribution of memory cells, read voltages should be updated based on the knowledge of them to maintain low error rates of NAND flash memory systems. However, a flash memory controller is unable to acquire the knowledge of retention time, so it is challenging to estimate the optimal read voltages in general. In this paper, we propose novel NAND flash memory read voltage estimation methods which are inspired by machine-learning. The proposed methods can perform highly-reliable and low-latency read operation without the knowledge of retention time. We first find features from the sensing and decoding process to obtain alternative information on the retention time. By exploiting them as the input metrics, we attempt to employ machine-learning techniques to the read voltage estimation framework in NAND flash memory systems. Since it requires only read of fractional cells in a memory block, a few read trials are required to estimate near-optimal read voltages. Also, we optimize the complexities of the proposed methods by selecting a subset of features to reduce computational complexity while maintaining estimation accuracy. Simulation results show that the error rates of the proposed methods can approach the optimal performance while requiring low-latency.