How FPGA can contribute to HPC ?
Taisuke Boku · 2022
Summary form only given, as follows. The complete presentation was not made available for publication as part of the conference proceedings. FPGA is one of the promised technologies to achieve high flexibility and performance to apply various applications, not just for small embedded system but also for large scale HPC (high performance computing) these days. The advanced large scale FPGAs have been growing their capability on the capacity of circuit as well as memory bandwidth, both are quite important to support HPC applications. Moreover, the high performance direct optical interconnection on FPGAs is quite important for parallel FPGA system construction to scale out the system size. Currently, the main player of accelerated HPC is GPU (graphics processing unit), however the FPGA solution becomes a new hero to support HPC system and application as well as GPU today. We have been developing an original technology to combine both FPGA and GPU for large scale accelerated PC clusters by compensating with each other for multi-physical simulation on advanced scientific calculation. Cygnus supercomputer is running at Center for Computational Sciences, University of Tsukuba, is a large scale proof-of-concept system for our CHARM (cooperating heterogeneous acceleration by reconfigurable multidevices) solution, and we have developed supporting subsystems such as FPGA-FPGA interconnection system and API for HLS, GPU-FPGA DMA module, multi-language compilation system, and several real applications to apply CHARM methodology to speed up the performance drastically from GPU-only solution. In this talk, I will summarize the current accelerated HPC solutions mainly depending on GPUs, and introduce our CHARM concept and its implementation as well as practical applications.