GPU Acceleration for Hermitian Eigensystems

As a recurrent problem in numerical analysis and computational science, eigenvector and eigenvalue determination usually employs high-performance linear algebra libraries. This paper explores the implementation of high-performance routines for the solution of multiple large Hermitian eigenvector and eigenvalue systems on a Graphics Processing Unit (GPU). We report a performance increase of up to two orders of magnitude over the original \textscEispack routines with a NVIDIA Tesla C2050 GPU, providing an effective order of magnitude increase in unit cell size or simulated resolution for Inelastic Neutron Scattering (INS) modelling from atomistic simulations