Abstract: As we approach the post-Moore era, hardware specialization (e.g., ASIC, FPGA) is becoming increasingly promising for sustaining performance growth and energy efficiency for specific tasks. Due to their inherently parallel nature in digital circuits, specialized hardware outperforms algorithms on general-purpose architectures, even with significantly slower clocks. Moreover, specialized hardware can be placed close to data sources, drastically reducing data movement, where general-purpose architectures cannot reach. However, there is no free lunch, as hardware development (in a traditional sense) can be very costly due to expert labor, license fees, fabrication cost, etc. To address some problems, open-source hardware development ecosystems are emerging as a recent trend; RISC-V inspires many, which could lower the hardware development cost and open new opportunities to bring our software expertise and HPC/AI experiences into the domain of hardware development.
In this talk, I will brief the traditional hardware development process, summarize emerging open-source hardware development ecosystems, and discuss possible new directions, opportunities, and challenges in hardware specialization. I will also showcase some of our ongoing hardware specialization projects at Argonne and give a live hardware construction and verification demo.