A 3D Sparse Convolution Accelerator With Unified Map Search Engine for Voxel-Based Point Cloud Neural Network

Voxel-based point cloud neural networks have been applied as an effective algorithm to resolve object 3 -dimenation (3D) information. Sparse convolution (SCONV), as the backbone of voxel-based point cloud neural networks, enhances computational efficiency and reduces memory usage in GPUs. However, challenges still remain in real-time, end-to-end 3D perception tasks, including complex map search computations and suboptimal data reuse, which need increased storage and high power consumption. To address these issues, we propose a 3D sparse convolution accelerator integrated with a unified map search engine for voxel-based point cloud neural network. This accelerator leverages a holistic design approach to minimize the computational complexity of map search through the Unified Map Search Engine. Moreover, we introduce a Computing Core equipped with a reuse lookup table, significantly improving onchip data reuse and minimizing power consumption by reducing data movements. To validate the scheme, the proposed accelerator implemented in a 28 nm CMOS achieves 5.92 TOPS/W peak efficiency in simulation, improving by $2.4 x$ compared to the state-of-the-art work.