Mixed-Precision AMG method for Many Core Accelerators

There is a large gap between single- and double-precision performances on GPUs. Single-precision arithmetic is more than twice as fast as double-precision arithmetic on GPUs. However, single-precision arithmetic cannot achieve sufficient accuracy. By using a combination of single- and double-precision arithmetics, accuracy is maintained and calculations are performed quickly. We introduce a mixed precision method to the algebraic multigrid (AMG) method and evaluate the performance on various GPUs. The results indicate that, for anisotropic problems, the performance of the mixed-precision AMG is approximately 1.9 times greater than the double-precision AMG at a size of 1003.