A single-GPU implementation of first-principles molecular dynamics

Abstract

We present a single-Graphics Processing Unit (GPU) implementation of First-Principles Molecular Dynamics (FPMD) using plane wave basis functions and pseudopotentials, for the NVIDIA CUDA platform. We discuss various design choices made to exploit the high memory bandwidth available on a GPU while minimizing host–device data transfers. Applications to FPMD simulations of superionic NH3, liquid water, defects in SiC and MgO, and various systems including up to 512 atoms and 4096 electrons are demonstrated on an NVIDIA A100 GPU and an NVIDIA Grace-Hopper GH200 platform. Performance benchmarks show a significant speedup compared to other GPU-enhanced implementations, enabling the efficient use of computational resources. Applications to ensemble FPMD simulations of free energy barriers of a solvated molecule and residual stress in liquid water are also presented.