FPGA Latency Optimization for LiDAR Point Clouds: From Wall Street to Autonomous Driving

Lately, I've been working on an FPGA acceleration solution for LiDAR. The latency on this path reminds me of my days tuning high-frequency trading engines on Wall Street. Back then, every nanosecond was money. Now, in autonomous driving, every nanosecond could be life or death. People usually use GPUs for point cloud processing, but GPU latency jitter is too high, making timing convergence difficult. The advantage of FPGAs lies in deterministic latency and pipelined processing, allowing true hard-real-time performance for point cloud segmentation and feature extraction. I previously tested a solution using pure FPGA for preprocessing 64-line LiDAR point clouds, keeping latency stable under 2.3 microseconds with roughly 75% resource utilization. If used for emergency braking, this could be several orders of magnitude faster than GPU solutions. Any peers working on similar things? Let's chat about the pitfalls in timing constraints and resource allocation.