LiDAR point cloud data processing in ADAS, AD, robotics, and surveillance applications requires ultra-low latency 3D perception, which is an immense challenge for a compact embedded system. With end-to-end hardware and software co-optimization, the adaptable Zynq® UltraScale+™ MPSoC platform can run multi-class 3D detection and segmentation models using the state-of-the-art PointPillars and SalsaNext models in real-time. Vitis™ AI supports the quantization, compilation, and deployment of the PointPillars model and runs efficiently on Zynq UltraScale+ MPSoC platform that contains integrated DPU cores.