A team from UCLA won the Best Paper Runner-Up Award at MICRO 2022 for their work: “OverGen: Improving FPGA Usability through Domain-specific Overlay Generation”. This project was a collaborative effort between Tony Nowatzki’s PolyArch lab and Jason Cong’s Vast lab. The mainstream FPGA programming approach of high-level synthesis is highly effective, but it has significant compilation time overhead in hours, and the FPGA reconfiguration time limits applicability to tasks with less dynamic behavior.
Motivated by the need to improve the usability of FPGA programming techniques and to enable broader adoption of powerful FPGA accelerators, the UCLA team developed the open source framework OverGen, which is an end-to-end FPGA acceleration system leveraging prior work in customizable spatial architectures, accelerator compilers, and flexible SoC generation. OverGen can compete in performance with state-of-the-art HLS techniques, while requiring 10,000x less compile time and reconfiguration time. The essential idea in OverGen is to develop a hardware generation framework targeting a highly-customizable overlay design space, so that FPGA implementations can be automatically synthesized and specialized to applications of interest.
The work was led by co-primary authors Sihao Liu, a 4th year PhD, and Jian Weng, who is on the academic job market this year. Six other UCLA students contributed to the effort, including four PhDs: Dylan Kupsh, Atefeh Sohrabizadeh, Zhengrong Wang, Licheng Guo, and four other MS, undergraduate, and interns.