high performance computing on graphics processing units: hgpu.org

In order to build the radio telescopes needed for the experiments planned for the years to come, it will be necessary to design computers capable of performing thousands more floating point operations per second than the actual most powerful computers of today, and do it in a very power efficient way. In this work we focus on the parallelization of a specific operation (part of the pipeline of most modern radio telescopes): the beam forming. We aim at discovering if this operation can be accelerated using Graphics Processing Units (GPUs). To do so we analyze a reference beam former, the one that ASTRON uses for the LOFAR radio telescope, discuss different parallelization strategies, and then implement and test the algorithm on a NVIDIA GTX 480 video card. Furthermore, we want to compare the performance of our algorithm using two different implementation frameworks: CUDA and OpenCL.