Discussion of GPU-coded Quantum Mechanics Software Performance¶
Akira Masago (JAMSTEC)
Abstract¶
In materials physics, quantum mechanics and statistical mechanics are the theoretical foundations, and various macroscopic properties of materials are studied from the microscopic viewpoint such as atoms and molecules. On this basis, many simulation software packages have been developed for case studies to handle specific materials. These are generally called first-principles quantum mechanics software. Many people believe that GPU computers hardly achieve good performance due to the algorithms. Nevertheless, famous first-principles quantum mechanics software packages that are not GPU-coded are becoming a minority with the recent growth of GPU computers. In this presentation, I will compare the performance on GPU with CPU computers in first-principles quantum mechanics software. Herein, AkaiKKR, which is a first-principles quantum mechanics software package developed by our research group, is used. This package is especially effective in semiconductor materials containing impurity atoms because it can reduce the computational cost through an approximation method called CPA. However, recently it is expected to be applied to large atomic systems such as interface structures, which is difficult to reduce computational costs. We hope that GPU computers can be of great help in such systems.