Runtime Behavior Comparison of Modern Accelerators and Coprocessors

Abstract

Recently, a variety of accelerator architectures became available in the field of high performance computing. Intel's MIC (Many Integrated Core) and both GPU architectures, NVIDIA's Kepler and AMD's Graphics Core Next, all represent the latest innovation in the field of general purpose computing accelerators. This paper explores several important characteristics of these architectures and investigates the impact of certain design factors on the achieved performance using the uCLbench micro-benchmarks, the NPB (NAS Parallel Benchmark) suite and diverse real-world applications from the field of physics. Based on the single unified programming interface OpenCL, we observe the run-time behavior of each test program on several test platforms. Major architectural discrepancies are studied and a higher level examination is discussed in details.