23.4 Nebula: A 28nm 109.8TOPS/W 3D PNN Accelerator Featuring Adaptive Partition, Multi-Skipping, and Block-Wise Aggregation

Three-dimensional (3D) point clouds are increasingly deployed across various emerging fields, such as autonomous driving, robots, drones, and virtual reality (VR) [1]–[6]. Point-based point-cloud neural networks (PNNs) [3]–[6] have demonstrated superior performance in point-cloud analysis, compared to both sparse 3D convolution-based networks [7], [8] and graph-based convolutional neural networks [9], [10]. Due to the high computational complexity, low parallelism, and frequent irregular external memory accesses, deploying PNNs in hardware is a great challenge. PNN hardware accelerators have been developed [11]–[20]. However, three key challenges remain unsolved in these accelerators, as illustrated in Fig. 23.4.1. 1) The inherent farthest point sampling (FPS) features serial computation and suffers from quadratic growth in inference latency with rising point counts. The existing uniform block-wise FPS techniques [13], [21] fail to achieve a well-balanced block segmentation, due to a typically non-uniform point distribution. 2) A large amount of redundant operations exist for both discarded points (DPs) and retained points (RPs) in FPS. These operations exist in the sampling operations of RPs ① as well as grouping ② convolution ③, and aggregation ④ for DPs, introducing unnecessary energy and latency costs. 3) The irregular memory accesses in the aggregation operation cause significant latency penalties. Channel-wise aggregation in [11] relieves irregularity, yet is unsuitable for large-scale point clouds, as the external memory access of features and the neighbor index table (NIT) is quadratically increased due to the iterative loading of features or the NIT.