"Implementation of a nonsymmetric linear solver on GPU"
Chadov S.N.

An implementation of a nonsymmetric sparse linear solver on GPU is considered. The solver uses a version of the BiCG-STAB algorithm. This algorithm is shortly described. Several sparse matrix storage formats are given with consideration of the NVIDIA GPGPU hardware features. The performance of the implementation is analyzed compared to the performance of a similar algorithm run on contemporary CPUs. The impact of several factors on the performance is discussed with some suggestions on further development.

Keywords: parallel computing, sparse linear solver, GPGPU, CUDA, BiCG-STAB

Chadov S.N.   e-mail: sergei.chadov@gmail.com