Data Pattern Memory Variation Aware Fine-Grained ECC Optimized by Neural Network for 3D-TLC NAND Flash Memories with 2.0x Data-Retention Time Extension and 30 Parity Overhead Reduction

摘要

This paper proposes fine-grained ECC optimized by Neural Network (NN) for 3D-TLC NAND flash memories. Among 48 combinations of 8 reliability enhancement techniques (RETs) and 6 BCH ECCs, proposed NN automatically selects the optimal combination of RET and BCH ECC. Errors with complicated data pattern dependence and memory variations are adaptively compensated. In addition, the design trade-off between the high reliability and parity overhead (cost) is resolved. As a result, both high reliability and small parity overhead are realized simultaneously. The acceptable data-retention time increases by 2.0-times and the data overhead decreases by 30%.