Evaluating floating-point intensive applications on OpenCL FPGA platforms: A case study on the SimpleMOC kernel

FPGAs are becoming a promising choice as a heterogeneous computing component for scientific computing when floating-point optimized architectures are added to current FPGAs. The high-level synthesis tools such as Intel FPGA SDK for OpenCL provide a streamlined design flow to facilitate the use of FPGAs for researchers. In this paper, we choose a nuclear reactor simulation application, the SimpleMOC kernel, as a case study to evaluate the potential and effectiveness of using an FPGA for floating-point intensive applications. We describe the OpenCL implementations of the kernel optimized with low-latency floating-point operators, on-chip memory accesses, loop transformations, kernel vectorization, and compute-unit duplication on an Intel Arria10-based FPGA platform, and evaluate their performance and resource utilizations. Compared to the baseline OpenCL implementation of the kernel, our optimizations improve the kernel performance by a factor of 102. We also evaluate the kernel application on an Intel Xeon 16-core CPU and an Nvidia Tesla K80 GPU. The GPU is approximately 2X faster than the CPU and 7.5X faster than the FPGA. The power consumption on the FPGA is 4.5X and 6.4X lower than that on the GPU and CPU, respectively. The performance per watt on the FPGA is 1.74X higher than that on the CPU, and 1.65X lower than that on the GPU.