Energy-efficient capsule network processor for real-time point cloud segmentation

Abstract

An energy-efficient capsule network processor is proposed for real-time 3D point cloud segmentation. The proposed processor adopts the heterogeneous core architecture pipelined in channel group level to achieve 35.4% latency reduction. Furthermore, the processor integrates a dynamic route skipping controller that predicts unimportant operations and skips them to reduce the external memory access by 39.1%. At last, the squash activation function unit exploits the look-up table (LUT) based computing with L2-norm approximation, which decreases capsule network inference time while only consuming around 1% area and power of the entire processor. The processor is simulated with the 65nm CMOS technology, showing 94.3mW power consumption and 1.15 TOPS/W energy-efficiency.