A discussion on calculating eigenvalues of real symmetric tridiagonal matrices on a GPU

While GPUs are attracting attention as an accelerator in wide-ranged application areas, compatibility between the architecture and selected algorithm is important to effectively bring out their potential performance. This paper focuses on eigenvalue calculation from a given real symmetric tridiagonal matrix and compares GPU implementations for the QR method and the bisection method. Implementation for a total of four different GPU architectures are shown and compared to reveal the affinity between algorithms and architectures.