A 1ms High-Speed Vision Chip with 3D-Stacked 140GOPS Column-Parallel PEs for Spatio-Temporal Image Processing

📋 论文概要

本文提出了一种采用3D堆叠技术的列并行处理单元（PE）视觉芯片，用于高速时空图像处理。通过位串行ALU和固定指令选择信号，实现了在108MHz频率下140GOPS的性能，处理延迟低至1ms。

💡 主要创新点

• 采用3D堆叠架构集成列并行PE阵列，提升处理吞吐量并降低延迟。
• 设计位串行ALU支持算术、逻辑和移位操作，通过固定指令选择信号实现单周期多操作。
• 4位PE在108MHz下达到140GOPS，适用于高速视觉应用。

🏷 关键词

📄 原文摘要

Masatsugu Kobayashi1, Sayaka Shida1, Masaki Odahara2, Kenichi Takamiya2, Yasuaki Hisamatsu2, Shizunori Matsumoto2, Leo Miyashita3, Yoshihiro Watanabe3, Takashi Izawa1, Yoshinori Muramatsu1, Masatoshi Ishikawa3 Sony Semiconductor Solutions, Atsugi, Japan Sony LSI Design, Atsugi, Japan 3 University of Tokyo, Bunkyo, Japan 1 bit-serial ALU supports arithmetic operations, logical operations, and shift operations. To achieve processing with a few cycles, all selection signals to the PE are fixed at one instruction. The 4b PEs achieve 140GOPS at a 108MHz peak frequency because a 1b primitive operation takes 1cycle and 1304 PE × 108MHz = 140GOPS. The power dissipation is 0.23mW/GOPS since the 4b PEs power consumption is 32mW. The PE inputs can be selected from neighbor-pixel data, processed data in neighboring PEs, and data in the individual PE’s work memory to realize highly flexible spatial operation. A maximum of four different programs

👥 作者与机构

Tomohiro Yamazaki1, Hironobu Katayama1, Shuji Uehara1, Atsushi Nose1,