GPU Acceleration for Density Functional Theory with Slater‐Type Orbitals

Abstract

This chapter describes the graphics processing unit (GPU) acceleration of density functional theory (DFT) calculations with Slater-type orbital (STOs) as developed for the Amsterdam Density Functional (ADF) program. This implementation is focused on accelerating the numerical integration step in ADF, which consumes the majority of CPU time in the existing implementation. Most implementations of DFT developed for molecular systems use atom-centered basis functions, typically with Gaussian-type orbitals (GTOs), as these facilitate the evaluation of Coulomb interactions. The chapter identifies the hardware and software requirements for accelerating the ADF package with GPUs. It then describes the CUDA code, developed for the numerical integration and analyzes its performance, followed by a description of the hybrid computational scheme used to accelerate ADF. Finally, the chapter compares the performance of the GPU-accelerated ADF code with the reference central processing unit (CPU) code for real simulations.