high performance computing on graphics processing units: hgpu.org

Cholesky decomposition has been widely utilized for positive symmetric matrix factorization in solving least square problems. Various parallel accelerators including GPUs and FPGAs have been explored to improve performance. In this paper, Cholesky decomposition is implemented on both FPGAs and GPUs by designing a dedicated architecture for FPGAs and exploiting massively parallel computation for GPUs. Performance of the Cholesky decomposition on GPUs, CPUs, FPGAs, and hybrid systems are compared in both single and double precision. Results show that the FPGA implementation has the highest efficiency with respect to clock cycles compared with our pure GPU implementation, a hybrid system with MAGMA, and a CPU with LAPACK. The GPU implementation is better than other implementations using MAGMA and LAPACK library for small matrices, and the hybrid system with MAGMA is the best for larger matrices.