A 22nm 2Mb ReRAM Compute-in-Memory Macro with 121-28TOPS/W for Multibit MAC Computing for Tiny AI Edge Devices

📋 论文概要

本文提出了一款基于22nm工艺的2Mb ReRAM存内计算宏，支持多比特乘累加（MAC）运算，实现了121-28TOPS/W的高能效，适用于微型AI边缘设备。解决了传统非易失性存内计算仅支持二进制输入/权重的问题，大幅提升了能效和延迟。

💡 主要创新点

• 首次在ReRAM存内计算中实现多比特（非二进制）输入和权重的MAC运算，支持灵活的比特精度配置。
• 通过电路和架构优化，实现了宽动态范围的能效（121-28TOPS/W），在2Mb容量下保持高性能。

🏷 关键词

📄 原文摘要

Hui-Yao Kao, Jing-Hong Wang, Ta-Wei Liu, Shih-Ying Wei, Sheng-Po Huang, Wei-Chen Wei, Yi-Ren Chen, Tzu-Hsiang Hsu, Yen-Kai Chen, Yun-Chen Lo, Tai-Hsing Wen, Chung-Chuan Lo, Ren-Shuo Liu, Chih-Cheng Hsieh, Kea-Tiong Tang, Meng-Fan Chang National Tsing Hua University, Hsinchu, Taiwan Nonvolatile computing-in-memory (nvCIM) can improve the latency (tAC) and energy-efficiency (EFMAC) of tiny AI edge devices performing multiply-andaccumulate (MAC) computing after system wake-up. Prior nvCIMs have proven effective for binary input (IN) and weight (W), and 3b output (OUT) [1], 1-8-1b IN-W-OUT [2], and 2-3-4b IN-W-OUT [3] neural networks; however, the higher precision (4-4b IN-W) for MAC operations is needed for multi-bit CNNs to achieved high-inference accuracy [4]. As Fig.15.4.1 shows, improving the precision of nvCIM macros involves various challenges. (1) A large number of activated WLs provides a wide range of BL current (IBL) resulting in an inaccurate

👥 作者与机构

Cheng-Xin Xue, Tsung-Yuan Huang, Je-Syu Liu, Ting-Wei Chang,