Cray's Approach to Heterogeneous Computing.

Heterogeneous architectures, combining traditional multicore CPUs with energy-efficient accelerators, can provide a viable approach to exascale computing. Such hybrid architectures are, however, difficult to program, requiring specific low-level rewriting of application codes. Here we describe Cray's approach to heterogeneous computing and the first Cray hybrid supercomputing system, the Cray XK6, with its unified x86/GPU programming environment. We also describe how the proposed standard for OpenMP accelerator directives offers a high-level programming model, and develop a GPU version of the Himeno benchmark on an early Cray XK6 prototype.