3840x Reliability Enhanced Robust NAND flash Optimized to Store Weight Data for Object Detection and Semantic Segmentation of Self-driving Car at High Temperature

This paper proposes reliability enhancement techniques in harsh environment to store weight data of machine learning (ML)-based applications, object detection and semantic segmentation for self-driving cars. Proposed techniques consist of robust NAND and Optimized Huffman Coding Compression (OHCC). Proposed robust NAND drastically decreases bit-error rate (BER) in extremely high temperature such as 210degC. Therefore, proposed techniques reduce miss recognition caused by weight data error and contribute to safety self-driving cars. Besides, proposed OHCC modulates weight data of ML-based image recognition applications by utilizing weight data characteristics that concerned around ‘0'. Consequently, proposed techniques extend data-retention (D. R.) time by 3,840 times for object detection and 2,550 times for semantic segmentation, respectively compared with conventional 3D triple-level-cell (TLC) NAND flash. In addition, proposed techniques achieve to decrease read access time and data-overhead by 39% and 94%, respectively.