A Low-power 3D Point Clouds Matching Processor with 1D-CNN Prediction and CAM-based In-memory kNN Searching
Jeongmin Shin, Hoichang Jeong, Seungbin Kim, Keonhee Park, Sang-Ho Lee, Kyuho Lee · 2024
This paper presents a computing-CAM-in-memory (C2IM) system designed for energy-efficient k-nearest neighbor (kNN) operations within 3D point clouds, which is an essential process for autonomous driving applications. In mobile processors, a new paradigm is required to conduct 3D point cloud kNN searching, under the limited hardware resources. Therefore, three features are proposed: i) Prediction using a dilated 1D-CNN, which enables voxel-based partitioning by minimizing the external memory accesses from O(n2) to O(n). ii) Vertex clustering, which restructures groups of points into evenly distributed clusters based on the underlying data distribution and effectively reduces the number of points for comparison by 49.8%. iii) Data movement minimization through in-memory kNN with content addressable memory (CAM). Designed with 28 nm CMOS technology, the proposed C2IM achieves reductions of 2.07× in memory footprint and 206.68× in power consumption, compared to state-of-the-art (SOTA) architectures.