A Sparse Convolution Neural Network Accelerator for 3D/4D Point-Cloud Image Recognition on Low Power Mobile Device with Hopping-Index Rule Book for Efficient Coordinate Management

Abstract

This work presents the first 3D/4D sparse CNN (SCNN) accelerator for point cloud image recognition on low power devices. A special hopping-index rule book method and efficient data search technique were developed to mitigate the overhead of coordinate management for SCNN. A 65nm test chip for 3D/4D images was demonstrated with 7.09–13.6 TOPS/W power efficiency and state-of-the-art frame rate.