R&D: Ternary-State Vertical NAND Flash Memory for Improvement of Density and Accuracy of Quantized Neural Networks (QNNs)

IEEE Transactions on Electron Devices has published an article written by Jin Ho Chang; Jae Seung Woo; Department of Electrical and Computer Engineering and the Inter-University Semiconductor Research Center (ISRC), Seoul National University, Gwanak-gu, Seoul, Republic of Korea, Suk-Kang Sung; Advanced Flash Technology Team, Samsung Electronics Company, Hwasung, Republic of Korea, Ki-Whan Song; Flash Design Team, Samsung Electronics Company, Hwasung, Republic of Korea, and Woo Young Choi, Department of Electrical and Computer Engineering and the Inter-University Semiconductor Research Center (ISRC), Seoul National University, Gwanak-gu, Seoul, Republic of Korea.

Abstract: “Novel ternary-state vertical NAND (VNAND) flash memory is proposed for high-density and high-accuracy quantized neural networks (QNNs) for the first time in this study. The proposed ternary-state VNAND features an additional current saturation region in transfer curves to implement ternary weights for QNNs using only a single memory cell. It achieves the inference accuracy of 96.06% and 81.04% for the MNIST and CIFAR-10 datasets, respectively, while maintaining high memory density. It is confirmed that the QNN accuracy of the proposed ternary-state VNAND is robust to variations in the threshold voltage ( Vth) and channel hole diameter ( dch) .“