Performance of Floating-point Intensive Kernels on Low-power Processor-A Case Study with Geodesic Distance Kernel

A processor, with a GPU and a CPU integrated on the same chip, is a promising low-power system for floating-point intensive applications. While an integrated GPU is not designed to outperform a discrete GPU due to its power, area, and thermal constraints, there is a need to better understand the performance of a floating-point intensive kernel using an integrated GPU. Toward this end, we choose a representative floating-point intensive kernel as a case study. We port the kernel with a vendor-neutral framework, analyze the compiler optimizations of the kernel at the assembly code, evaluate the relationship between floating-point operations per second and arithmetic intensity, and compare the performance and power of the kernel implementations on the CPU and GPU. Our key findings are: 1) Compared to an un-optimized kernel, the floating-point optimizations improve the performance of the single-and double-precision floating-point kernels executing on an Intel^® GEN8 Iris Pro GPU by 15.4X and 5.4X, respectively; the optimizations also improve the performance of the two kernels by 5.6X and 3.4X on an Intel^® Xeon^® E3 CPU, respectively. 2) Achieving peak floating-point operations per second on the GPU requires much higher arithmetic intensity than that on the CPU. 3) Running the floating-point intensive kernel on the processor consumes 48 Watts, which is very close to the thermal power draw of the processor. The floating-point optimization can reduce the average GPU power from 35.7 W to 22.7 W for the double-precision kernel, and from 33.1 W to 8.8 W for the single-precision kernel.