Understanding the Performance of Elementary NLA Kernels in FPGAs

Our work compares the use of multi-core hardware and FPGAs to address Numerical Linear Algebra (NLA) kernels. Specifically, we study the behavior of highly tuned kernels in a multi-core CPU processor and OpenCL implementations on FPGAs. For this purpose, we select two of the most significant kernels in the NLA field, the general matrix multiplication (GEMM) and the sparse matrix-vector multiplication (SPMV), which are the most characteristic kernels of dense and sparse NLA respectively. Finally, we perform the experimental evaluation of our implementations on a low-end FPGA platform, and state-of-the-art implementations for traditional CPUs, such as those included in the Intel MKL library, both in terms of runtime and energy consumption.