LPNA: A 2.75 ms Low-Latency and 0.40 uJ/point Energy-Efficient LiDAR Point-Cloud Neural Network Accelerator with Cylindrical Bin Partitioning

As the interest in 3D perception for outdoor applications grows, 3D Point-cloud Neural Network (PNN) with LiDAR has been actively developed [1]. Due to the unstructured and unordered properties of point-cloud, prior PNNs applied additional preprocessing, e.g., voxelization, 3D projection, even though such methods lose point data. To achieve high performance in 3D segmentation, recent PNNs directly use point-cloud data as input and learn accurate point-wise features [2]. However, such a network requires a serial process of special functions, e.g., k-Nearest Neighbor (kNN) search and shared multi-layer perception (MLP), due to the properties of point-cloud. The serial process of kNN search and shared MLP causes long processing latency. Also, a large amount of external memory access (EMA) is inevitable in decoder layers for interpolation. Thus, PNN solely takes $\sim 335 \mathrm{~ms}$ on a high-end CPU+ GPU system (Fig.1).