SHAPE AUDIONAMIX: A benchmark of linear algebra libraries for HPC

The source separation technology developed by Audionamix requires a significant amount of computation. To increase the speed of the technology, several High Performance Computing libraries for linear algebra have been benchmarked with two applications: a simple matrix multiplication and a Non-Negative Matrix Factorization. The benchmarked libraries were Eigen, Armadillo, MKL and the GPU library Magma. The benchmark returned that for both applications the GPU-based solution performed better than the other solutions. But among CPU-based solutions, Eigen performed better on a complex task than MKL. Thus, making the Magma library as an Eigen back-end seems to have a good potential for HPC-based linear algebra, since this solution would benefit at the same time from the GPU-accelerated computation of Magma and from Eigen's optimized data transfer. 1