VoxCAD: A 0.82-to-81.0mW Intelligent 3D-Perception dToF SoC with Sector-Wise Voxelization and High-Density Tri-Mode eDRAM CIM Macro

📋 论文概要

本文提出VoxCAD，一种用于端到端3D感知的智能低功耗dToF SoC，通过集成LiDAR dToF传感与2D-ROI引导的点云构建、粗-细扇区体素化以及可重构三模eDRAM CIM宏，解决了传感器传输能量和延迟高、片上内存大以及3D变换器面积效率低的问题。

💡 主要创新点

• 提出LiDAR dToF传感与2D-ROI引导的点云构建，减少传感器传输能量和延迟。
• 采用粗-细扇区体素化与中央计算单元，降低片上内存大小和EMA数量。
• 设计可重构三模eDRAM CIM宏，采用混合1T1C/2T1C单元，以高面积效率和存储密度支持3D变换器。

🏷 关键词

📄 原文摘要

*Equally Credited Authors (ECAs) 1 Abstract VoxCAD, an intelligent low-power dToF SoC for end-to-end 3D perception applications, is presented with three features: 1) LiDAR dToF sensing-integrated 2D-ROI guided point cloud construction for reducing sensor transfer energy and latency; 2) coarse-fine sector-wise voxelization with a central-computation unit for reducing on-chip memory size and EMAs; and 3) reconfigurable tri-mode eDRAM CIM macros with hybrid 1T1C/2T1C cells for supporting 3D transformers at high area efficiency and storage density. The rapid growth of the electric-vehicle (EV) and smart robotics market drives demand for full self-driving (FSD), advanced driver-assistance systems (ADAS) and autonomous navigation. The foundation of such systems lies in end-to-end environment perception from 2D/3D sensors and efficient execution of AI tasks. Figure 7.2.1 presents a conventional 3D

👥 作者与机构

Haoyang Sang*1, Zhao Wang*1, Longzhen He1, Guangshu Zhao1, Wenao Xie2, Bo Wang3, Rui P. Martins1, Man-Kay Law1

University of Macau, Macau, China, 2KAIST, Daejeon, Korea, 3Hamad Bin Khalifa University, Doha, Qatar