Evaluating Radial Basis Function Kernel on OpenCL FPGA Platform

Field-programmable gate arrays (FPGAS) are becoming a promising heterogeneous computing component for scientific computing when floating-point optimized architectures are added to the current FPGAS. The emerging high-level synthesis (HLS) tools provide a streamlined design flow to facilitate the use of FPGAS for researchers who have little FPGA development experience. In this paper, we choose the kernel, Radial Basis Function, in a support vector machine as a case study to evaluate the potential of implementing machine learning kernels on FPGAS, and the capabilities of an HLS tool to convert a kernel written in high-level language to an FPGA implementation. We explain the HLS flow and the RBF kernel. We evaluate the kernel in an OpenCL-to-FPGA HLS flow, and describe the optimizations of the kernel. Our optimizations using kernel vectorization and loop unrolling improve the kernel performance by a factor of 15.8 compared to a baseline kernel on the Nallatech 385A FPGA card that features an Intel Arria 10 GX 1150 FPGA. In terms of energy efficiency, the performance per watt on the FPGA platform is 2.8X higher than that on an Intel Xeon 16-core CPU, and 1.7X higher than that on an NVIDIA Tesla K80 GPU. On the other hand, the performance per watt on an Intel Xeon Phi Knights Landing CPU and an NVIDIA Tesla P100 GPU are 5.3X and 1.7X higher than that on the FPGA, respectively.