Trillion-Atom Exascale Performance Portability of FLARE for Catalysis

Machine learning interatomic potentials (MLIPs) have become a prevalent approach to bridging the gap between slow-but-accurate ab initio calculations and fast-but-inaccurate empirical potentials for molecular dynamics. Among MLIPs, there is a Pareto front of models with different tradeoffs between accuracy and speed. The FLARE interatomic potential has previously pushed the boundary of scalability and performance on NVIDIA GPUs, reaching 500 billion atoms at record speed on 27336 GPUs [1], while maintaining sufficient accuracy to study complex, reactive systems.<br/><br/>In this work, we demonstrate recent performance results with FLARE and its Kokkos implementation in LAMMPS on OLCF Frontier, the current fastest supercomputer in the world [2]. Using the tens of thousands of AMD MI250X GPUs available on Frontier, we breach the trillion-atom barrier for molecular dynamics simulations on GPUs and achieve performance parity with NVIDIA A100 GPUs with no AMD-specific changes to the code. We also discuss recent optimizations of FLARE that more than double its performance across different platforms, as well as its new functionality for performing Bayesian active learning directly in LAMMPS. Finally, we highlight how the extreme performance of FLARE enables a detailed study of heterogeneous catalytic processes through large-scale simulations.<br/><br/>[1] arXiv:2204.12573<br/>[2] https://www.top500.org/