high performance computing on graphics processing units: hgpu.org

With the advent of parallel processing architectures and a steep increase in parallelism found among the recent applications, GPGPUs have gained attention with respect to their importance in the execution of these applications. In this document, we specifically analyze Sparse-Matrix Vector Multiplication(SPMV) across different architectures, libraries and matrix formats. The experimental platforms include but are not limited to GTX770, Tesla K40c, Tesla K20Xm while the different libraries we have used are CUSP, cuSPARSE, VexCL, ViennaCL and MKL. The purpose of this effort is to identify several trade-offs with respect to architectures and libraries while also accounting for density, size and number of non-zeros(NNZ) of the sparse matrices we worked with.