high performance computing on graphics processing units: hgpu.org

Graphics processing units (GPUs) greatly improved their performance over the last ten years. The first graphics cards have been developed in the late 90’s and were targeted for the mass market. These first cards were special purpose hardware, designed to accelerate graphic processing required in computer games. As the interest in computer games continued, GPU developers such as NVIDIA and AMD/ATI continuously improved the performance and level of parallelism of their GPUs. To this end, it became desirable to exploit this special purpose hardware for general purpose computing. Therefore, parallel programming models such as CUDA and OpenCL are developed in order to utilize the hundreds of cores of modern GPUs which yield a peak performance of up to 1331 GFLOP/s on a single GPU. These programming models greatly decreased the effort. necessary to program these architectures. At their beginning in 2007 (CUDA) and 2008 (OpenCL), they suffered from drawbacks such as low double precision performance, few debugging capabilities and the lack of atomic operations [36]. Today, the development of these models and the underlying architectures are by no means completed. Nevertheless, the fact that more and more heterogeneous clusters within the TOP500 [3] consist of a mixture of CPUs and GPUs emphasizes the increasing importance of GPUs in the field of scientific computing. Hence, it is crucial to provide new benchmark suites which enable the evaluation of heterogeneous systems in order to evaluate and compare the performance of these systems.