Eigen-G: GPU-based eigenvalue solver for real-symmetric dense matrices

Abstract

This paper reports the performance of Eigen-G, which is a GPU-based eigenvalue solver for real-symmetric matrices. We confirmed that Eigen-G outperforms state-of-the-art GPU-based eigensolvers such as magma-dsyevd and magma-dsyevd-2stage implemented in the MAGMA version 1.4.0. Applying the best-tuned CUDA BLAS libraries and the GPU-CPU hybrid DGEMM yields an even better performance improvement. We observe an approximately 2.3 times speedup over magma-dsyevd on a Tesla K20c. © 2014 Springer-Verlag.