Over the last couple of years, the industry has brought autonomous driving from merely a futuristic concept to a number of realistic prototypes. Despite these remarkable advancements, building autonomous driving systems is still challenging due to the unique performance requirements, constraints and computational characteristics. The computation and data transferring patterns in autonomous driving systems are fundamentally different from their counterparts in traditional computation problems. The general purpose processing platforms such as CPUs and GPUs, which are not originally designed for the autonomous driving tasks, are incapable of running them efficiently. This particular need for customizability has made FPGAs suitable platforms for autonomous driving. In this talk, we share our first hand experience of FPGA architecture exploration in autonomous driving systems at Pony AI. We describe two FPGA-powered architectures in depth - a sensor synchronization module and a domain specific accelerator. The FPGA architectures have enabled accurate sensor synchronization and substantially improved the computation power efficiency in our system.
Sr. Systems Architect