LAMMPS-KOKKOS: Performance Portable Molecular Dynamics Across Exascale Architectures

Since its inception in 1995, LAMMPS has grown to be a world class molecular dynamics code, with thousands of users, over one million lines of code, and multi-scale simulation capabilities. In this presentation, I will discuss how LAMMPS has adapted to the modern heterogeneous computing landscape by integrating the Kokkos performance portability library with a modular and user-friendly approach. I will also discuss the seamless ways in which LAMMPS supports advances in GPU-accelerated machine-learned models for more physically realistic simulations.

The focus of this presentation is a recent set of benchmarks across GPUs from different vendors and generations, combined with scaling studies on exascale machines. The benchmarks investigate the performance portability of simple pairwise, manybody reactive, and machine-learned interatomic potentials, and analyze performance trends, probing FLOPS throughput, memory bandwidths, cache capabilities, and thread-atomic operation performance. Finally, I will discuss scaling results for the different potentials, including full-machine benchmarks on the NNSA El Capitan, OLCF Frontier and ALCF Aurora exascale machines.