3D-NAND Flash Solid-State Drive (SSD) for Deep Neural Network Weight Storage of IoT Edge Devices with 700x Data-Retention Lifetime Extention

摘要

3D-TLC (triple-level cell) NAND flash-based solid-state drive (SSD) for deep neural network (DNN) weight storage is proposed. The data-retention lifetime of 3D-TLC NAND flash memory is extended by 700-times to achieve over 10-year lifetime of IoT edge devices such as automobiles and infrastructures. Proposed SSD combines reliability enhancement techniques for 3D-TLC NAND flash memories with unique characteristics of DNN weights, which have values of near 0. This paper proposes two techniques for SSD controller. The 1st proposal, One-State Error Recovery, removes all of 1-state errors of important bits in DNN weights even when error-correcting code (ECC) cannot correct errors. The 2nd proposal, DNN Weight Data Mapping, assigns frequently used "0" to the highly reliable VTH-state of memory cells. Due to error tolerance of DNN weights, acceptable bit error rate (BER) increases by 9.8-times.