OpenCL-enabled Parallel Raytracing for Astrophysical Application on Multiple FPGAs with Optical Links

In an earlier study, we optimized the Authentic Radiative Transfer (ART) method to solve the space radiative transfer problems in early universe astrophysical simulations using an Intel Arria 10 Field Programmable Gate Array (FPGA). In this paper, we optimize this method for use on the latest FPGA, an Intel Stratix 10, and evaluate its performance by comparing the GPU implementation on multiple nodes. For the multi-FPGA computing and communication framework, we apply our original system, called as Communication Integrated Reconfigurable CompUting System (CIRCUS), to realize OpenCL based programming and utilize multiple optical links on an FPGA for parallel FPGA processing, and this study is the first implementation of a real application applied using CIRCUS. The FPGA implementation is 4.54-, 8.41-, and 10.64-times faster than that of a GPU on one, two, and four nodes, respectively, for multi-GPU cases using an InfiniBand HDR100 network. It also achieves 94.2% parallel efficiency running on four FPGAs. We believe this efficiency is brought about from the low-latency and high-efficiency pipelined communication of CIRCUS, which provide easy programming on multi-FPGAs using OpenCL for high-performance computing applications.