CV-CIM: A 28nm XOR-Derived Similarity-Aware Computation-in-Memory for Cost-Volume Construction

📋 论文概要

本文提出了一种基于XOR推导的相似性感知计算内存架构（CV-CIM），用于实时成本体积构建，解决了立体视觉处理中内存带宽和运算量的瓶颈问题。该设计在28nm工艺上实现了高能效的相似性计算。

💡 主要创新点

• 提出XOR推导的相似性计算方法，简化了成本体积计算中的像素匹配操作。
• 采用相似性感知的计算内存架构，利用数据局部性减少内存访问，提升带宽效率。
• 在28nm工艺下实现，支持实时立体视觉应用（>30fps）并达到0.254Tb/s带宽和391GOPs运算能力。

🏷 关键词

📄 原文摘要

pixels in paired images, is a fundamental kernel of stereo vision processing and has been directly used in robotic, autopilot, and AR/VR applications. However, the large parameter size and consecutive data accesses of real-time cost-volume construction (>30fps) exerts a high demand on the memory bandwidth (0.254Tb/s) and operation (391GOPs). A promising candidate to resolve the memory bottleneck is computation-in-memory (CIM), which provides computing parallelism based on a structured array [1-7]. However, to deploy CIM applications one needs to overcome three challenges: 1) Considering various image properties and frame rate scenarios, cost functions are pattern-dependent and could be, in essence, classified as distances between pixels. CIM only supports restrictedvector Boolean logic, but distance operations require c

👥 作者与机构

Zhiheng Yue, Yang Wang, Huizheng Wang, Yabing Wang, Ruiqi Guo,

Limei Tang, Leibo Liu, Shaojun Wei, Yang Hu, Shouyi Yin Tsinghua University, Beijing, China Cost-volume construction, which accurately computes the similarities between