high performance computing on graphics processing units: hgpu.org

As hardware accelerators and especially GPUs become more and more popular to accelerate the compute intensive parts of an algorithm, standard high performance computing packages are starting to benefit from this trend. We present the first GPU acceleration of the ANSYS direct sparse solver. We explain how such a multifrontal solver may be accelerated using an optimized dense matrix factorization, and show that with the current generation of hardware, speed-ups of up to 4x can be obtained by utilizing mixed precision, and 2.9x in double precision. We discuss the impact of the model being studied on the overall performance, as well as the influence of using single precision factorizations on the accuracy of the solution.