An Efficient Multi-GPU Implementation for Linear-Response Time-Dependent Density Functional Theory

Abstract

Nowadays, Kohn-Sham density functional theory (DFT) calculation has drawn more and more attention in chemistry and material science simulations. However, due to the extreme large Hamiltonian matrix needed to be generated during the calculation, when the studied system increases, the cost of calculation becomes unbearable both in ground and excited state electronic structure simulations with large uniform basis. In this paper, we propose a high-performance multi-GPU approach for linear-response time-dependent density functional theory (LR-TDDFT) calculation to compute the excitation energies in molecules and solids with the plane wave basis set under the periodic boundary condition. We carefully design the parallel implementation, calculation steps and data distribution schemes in the naive CPU implementation to maintain good scalability when the studied system expands, then port the most time-consuming part to multi-GPU platform along with several effective optimization steps. The results show that with dual V100 GPUs, the proposed approach can achieve an average of 6.68x speedup compared with dual 12-core Xeon CPU with bulk silicon systems that comprises thousands of atoms (1,024 atoms).