Power and performance tradeoff of a floating-point intensive Kernel on OpenCL FPGA platform

As power is recognized as one of first-order constraints in high-performance computing, understanding how power and performance are affected using a high-level synthesis tool for a floating-point intensive kernel is important. FPGAs offer a promising solution for high-performance and energyefficient computing applications. This paper presents the impact of the optimizations of a floating-point intensive kernel from a geographical information system upon the performance and power using an FPGA. Using an OpenCL-based high-level synthesis tool for FPGAs, we evaluate the resource usage, performance and power consumption of the kernel implementations on an Arria10-based FPGA platform. We compare the performance and energy efficiency of the kernel implementations on an Arria 10 GX1150 FPGA, an Intel's Xeon Phi Knights Landing CPU, and an NVIDIA's Tesla K80 GPU. Our experiment shows that the performance per watt of the kernel implementation on the FPGA is 1.79X better than the CPU and 1.56X better than the GPU. The execution time on the FPGA is approximately 2.9X slower.