A 22nm Delta-Sigma Computing-In-Memory (∆ΣCIM) SRAM Macro with Near-Zero-Mean Outputs and LSB-First ADCs Achieving 21.38TOPS/W for 8b-MAC Edge AI Processing

📋 论文概要

提出了一种基于Delta-Sigma调制技术的计算内存(∆ΣCIM)SRAM宏，利用输入特征的渐进变化特性，通过近零均值输出和LSB优先ADC实现低功耗高精度矩阵向量乘法，解决AI边缘设备中数据移动能耗大的问题。

💡 主要创新点

• 首次将Delta-Sigma调制应用于计算内存架构，利用输入信号的时域冗余减少计算能耗。
• 采用近零均值输出和LSB优先ADC设计，在保持高精度的同时显著降低转换功耗。

🏷 关键词

📄 原文摘要

Peking University, Hangzhou, China, 3 Nano Core Chip Electronic Technology, Hangzhou, China *Equally Credited Authors (ECAs) 1 2 In AI-edge devices, the changes of input features are normally progressive or occasional, e.g., abnormal surveillance, hence the reprocessing of unchanged data consumes a tremendously redundant amount of energy. Computing-in-memory (CIM) directly executes matrix-vector multiplications (MVMs) in memory, eliminating costly data movement energy in deep neural networks (DNNs) [2-6]. Prior CIM work only explored the sparsity of DNNs to improve energy efficiency, but the trend of employing non-sparse activation functions, e.g., leaky ReLU, degrade the benefits of leveraging sparsity [1]. Even if sparsity can be exploited, the redundant unchanged input features in analog CIM

👥 作者与机构

Peiyu Chen*1, Meng Wu*1, Wentao Zhao1, Jiajia Cui1, Zhixuan Wang1,2,

Yadong Zhang3, Qijun Wang3, Jiayoon Ru1, Linxiao Shen1, Tianyu Jia1, Yufei Ma1, Le Ye1,2, Ru Huang1 Peking University, Beijing, China Advanced Institute of Information Technology of