high performance computing on graphics processing units: hgpu.org

Previous research in the field of statistical testing of random number sequences using Graphics Processing Units (GPU) has shown that this approach yields a significant increase in performance for a subset of the statistical tests proposed by National Institute of Standards and Technology (NIST). The present paper aims at further improvements in the performance of statistical testing of random number sequences, by focusing on another technology dedicated to GPU computing, the Compute Unified Device Architecture (CUDA). CUDA extends the C programming language with functionality for massively parallel programming on GPUs. Due to the flexibility given by the CUDA memory and thread model plus the optimizations that take advantage of the Parallel Data Cache, we were able to further improve the performance of the statistical testing algorithms proposed by NIST. Experimental results show speedups of up to 219, depending on the test and the size of the input data, with an overall average speedup of 51.