Numerical Investigations of Variable Preconditioned GCR with Mixed Precision on GPU

抄録

The Variable Preconditioned GCR (VPGCR) with mixed precision on Graphics Processing Unit (GPU) using CUDA is numerically investigated. The convergence theorem of VPGCR is guaranteed that the residual equation can be solved in the range of single precision, which means that VPGCR is applicable method to elicit the high performance of GPU. The results of computations show that VPGCR with mixed precision demonstrated significant achievement than that with double precision operation. Especially, the hybrid VPGCR on GPU is 10.10 times faster than that of CPU, and the standard VPGCR on GPU is 10.36 times faster than that of CPU.