Accelerating lattice QCD computations using general purpose graphic cards

点击翻译按钮获取中文摘要

We employ commodity graphic cards produced by Nvidia to accelerate the quark solver for lattice QCD simulations. With a single GPU card we obtain factors of 7-10 speed up for the quark solver compared with a single CPU case. The sample GPU solver program is freely available from the Web page of this project. We have developed the blocked iterative solver in which the independent linear equations are assigned to individual GPUs in a PC box. With this algorithm we can achieve an ideal improvement through the embarrassingly parallel usage of the multiple GPUs. Using the blocked GPU solver, we could estimate the running coupling constant of the 10-flavor SU(3) gauge theory very efficiently. In order to achieve large scale lattice QCD simulations with multiple PCs and GPUs, we have developed the additive Schwarz preconditioner to solve the domain decomposed quark linear equations. We did not observe the efficient improvement with the additive Schwarz preconditioner.