SpCNA: An FPGA-based Accelerator for Point Cloud Convolutional Neural Networks

3D convolutional neural networks (CNNs) are commonly used to process and analyze point clouds for object detection. However, the submanifold sparse convolutions and traditional sparse convolutions, which play important roles in 3D CNNs on point cloud, often encounter performance issues when accelerated by existing convolutional neural network accelerators due to the unstructured sparsity of sparse convolution. In this paper, we present SpCNA, an FPGA-based accelerator for sparse CNNs on 3D point clouds. To avoid computations that involve a large number of zeros in point cloud, it utilizes hash tables to map input-weight-output pairs. The computational performance of our design achieves 11.00 GFLOP/s. It can provide 1.2× and 1.98× energy efficiency improvement compared to Nvidia Xavier SoC and Nvidia 2080ti GPU, respectively.