A 28nm 192.3TFLOPS/W Accurate/Approximate Dual-Mode-Transpose Digital 6T-SRAM CIM Macro for Floating-Point Edge Training and Inference connection using the 3rd-metal layer and connect the corresponding diagonal in the 4th layer; the row connection is in the 5th layer. This method circumvents the need for numerous MAC circuits and read ports as is the case for previous T-CIM works [7,8], resulting in a reduction in area and power consumption.

📋 论文概要

本文提出了一种28nm工艺下的精确/近似双模式转置数字6T-SRAM存算一体宏单元，支持FP8、BF16、INT4和INT8等多种数据格式，实现了192.3 TFLOPS/W的高能效，旨在解决边缘训练中浮点运算的能效瓶颈。通过引入FP8新型浮点数据格式，相比BF16提升了训练效率。

💡 主要创新点

• 提出精确/近似双模式转置数字6T-SRAM存算一体架构，兼顾精度与能效。
• 支持FP8、BF16、INT4、INT8多种数据格式，特别是FP8新型浮点格式，提升训练效率。
• 实现192.3 TFLOPS/W的极高能效，适用于边缘训练场景。

🏷 关键词

📄 原文摘要

Shidong Lv3, Hao Wu1,2, Cailian Ma1,2, Ming Li1,2, Jinshan Yue1, Xinghua Wang3, Guozhong Xing1, Pui-In Mak4, Xiaoran Li3, Feng Zhang1 Figure 14.5.4 depicts the DCIM architecture supporting FP8, BF16, INT4, and INT8 formats. FP8 is proposed as a novel FP data format to achieve higher training efficiency than BF16

👥 作者与机构

Yiyang Yuan1,2, Bingxin Zhang1,2, Yiming Yang3, Yishan Luo1,2, Qirui Chen3,