Voxel-CIM: An Efficient Compute-in-Memory Accelerator for Voxel-based Point Cloud Neural Networks

TL;DR

Voxel-CIM is a novel compute-in-memory accelerator designed for voxel-based point cloud neural networks, significantly reducing data movement and improving energy efficiency and speed in 3D perception tasks.

Contribution

The paper introduces Voxel-CIM, a compute-in-memory accelerator with innovative data reuse and processing strategies for efficient voxel-based neural network inference.

Findings

Achieves 4.5-7.0x higher energy efficiency (10.8 TOPS/w)

Provides 2.4-5.4x speedup in detection tasks

Offers 1.2-8.1x speedup in segmentation tasks

Abstract

The 3D point cloud perception has emerged as a fundamental role for a wide range of applications. In particular, with the rapid development of neural networks, the voxel-based networks attract great attention due to their excellent performance. Various accelerator designs have been proposed to improve the hardware performance of voxel-based networks, especially to speed up the map search process. However, several challenges still exist including: (1) massive off-chip data access volume caused by map search operations, notably for high resolution and dense distribution cases, (2) frequent data movement for data-intensive convolution operations, (3) imbalanced workload caused by irregular sparsity of point data. To address the above challenges, we propose Voxel-CIM, an efficient Compute-in-Memory based accelerator for voxel-based neural network processing. To reduce off-chip memory access for map search, a depth-encoding-based output major search approach is introduced to maximize data reuse, achieving stable $O(N)$-level data access volume in various situations. Voxel-CIM also employs the in-memory computing paradigm and designs innovative weight mapping strategies to efficiently process Sparse 3D convolutions and 2D convolutions. Implemented on 22 nm technology and evaluated on representative benchmarks, the Voxel-CIM achieves averagely 4.5~7.0$\times$ higher energy efficiency (10.8 TOPS/w), and 2.4~5.4$\times$ speed up in detection task and 1.2~8.1$\times$ speed up in segmentation task compared to the state-of-the-art point cloud accelerators and powerful GPUs.