A comparative investigation of device-specific mechanisms for exploiting HPC accelerators

A variety of computational accelerators have been greatly improved in recent years. Intel's MIC (Many Integrated Core) and both GPU architectures, NVIDIA's Kepler and AMD's Graphics Core Next, all represent real innovations in the field of HPC. Based on the single unified programing interface OpenCL, this paper reports a careful study of a well thought-out selection of such devices. A micro-benchmark suite is designed and implemented to investigate the capability of each accelerator to exploit parallelism in OpenCL. Our results expose the relationship between several programing aspects and their possible impact on performance. Instruction level parallelism, intra-kernel vector parallelism, multiple-issue, work-group size, instruction scheduling and a variety of other aspects are explored, highlighting their interaction that must be carefully considered when developing applications for heterogeneous architectures. Evidence-based findings related to microarchitectural features as well as performance characteristics are cross-checked with reference to the compiled code being executed. In conclusion, a case study involving a real application is presented as a part of the verification process of statements.