high performance computing on graphics processing units: hgpu.org

Graphic Processing Units (GPUs) are widely employed in many applications in which high computing capabilities are required and parallelism can be fruitfully exploited. A higher amount of parallel threads bring to the GPU a higher throughput, but may also increase the code neutron-induced error rate. The GPUs sensitivity depends not only on the code throughput, but also on the chosen threads distribution. We experimentally evaluate how the neutroninduced output error rate of some benchmark codes varies when their throughput is increased. Experiments found that increasing the block size minimizes the application neutron-induced output error rate.