A 28nm 72.12TFLOPS/W Hybrid-Domain Outer-Product Based Floating-Point SRAM Computing-in-Memory Macro with Logarithm Bit-Width Residual ADC

📋 论文概要

本文提出了一种基于混合域外积（Hybrid-Domain Outer-Product）的浮点SRAM存算一体宏，采用对数位表示（Logarithm Bit）以提升计算精度与能效。在28nm工艺下实现了72.12 TFLOPS/W的峰值能效，解决了边缘AI应用中浮点运算能效低、精度不足的问题。

💡 主要创新点

• 提出混合域外积计算架构，融合对数域与线性域优势，在保持高精度的同时显著降低功耗。
• 首次在SRAM CIM中实现全浮点外积运算，避免了传统整数CIM的精度损失和量化开销。
• 通过对数位表示优化数据流与存储，减少了位宽和计算复杂度，提升了能效和吞吐率。

🏷 关键词

📄 原文摘要

Qirui Chen3, Meini Tang3, Xi Wei3, Zhixian Hou3, Jialiang Zhu1,2, Hao Wu1,2, Qirui Ren1,2, Guozhong Xing1, Pui-In Mak4, Feng Zhang1 Institute of Microelectronics of the Chinese Academy of Sciences, Beijing, China University of Chinese Academy of Sciences, Beijing, China 3 Beijing Institute of Technology, Beijing, China 4 University of Macau, Macau, China 1 2 SRAM-based computing-in-memory (CIM) is considered crucial to achieving high-energy efficiency (EF) for artificial-intelligence (AI) applications on edge devices. Researchers are currently exploring floating-point (FP) CIM [1,2], as integer (INT) precision CIMs [36] are no longer sufficient for new AI applications, which demand increased accuracy, complexity, and on-chip training. However, both analog and digital FP-CIMs face several significant challenges in realizing FP calculations, due to difficulties associated with handling high-bit precision: including (1) effectively combining the advantages of analog

👥 作者与机构

Yiyang Yuan1,2, Yiming Yang3, Xinghua Wang3, Xiaoran Li3, Cailian Ma1,2,