Summary
Running realistic simulations of materials requires complex models that predict atomic interactions, but these models can be hard to run efficiently on GPUs. The authors developed a specialized software tool that allows a broad class of machine-learned atomic models to run on both Nvidia and Intel GPUs using the popular LAMMPS simulation code. They demonstrate that their implementation scales to large simulations involving hundreds of thousands of atoms and runs at practical speeds across different hardware types. This approach makes it easier and faster to perform advanced molecular dynamics simulations with a unified software setup.
What this means in practice
- •For materials simulation engineers: Run large-scale isothermal-isobaric molecular dynamics simulations using machine-learned potentials efficiently on heterogeneous GPU hardware with a unified interface.
- •For high performance computing teams: Integrate a scalable, cross-vendor GPU molecular dynamics workload into multi-node clusters with optimized communications and single-source code for diverse GPUs.
Abstract
Foundation machine-learning interatomic potentials (MLIPs) offer one model for many chemistries, but their deep, globally coupled architectures do not fit the spatial decomposition that scales classical and short-ranged MLIPs. We present pair_style uma, a LAMMPS pair style that evaluates the Universal Model for Atoms (UMA) through LibTorch on NVIDIA and Intel GPUs from a single source tree. It reimplements FairChem's edge-sharded graph parallelism in C++ over exported TorchScript modules, with one MPI rank per device, NCCL or oneCCL collectives, and the whole energy and force path in double precision. The strain-gradient virial is carried through the exported modules and the cross-device reductions, so that fix npt runs on both vendors' hardware. We measure NPT capacity and cost on one 4xA100-40GB node, one 8xH200 node, and one 12-tile Intel node, and on up to 32 Intel nodes. Without activation checkpointing, one node holds NaCl cells of about 8,000 atoms (A100, 0.21 ns/day), 22,000 atoms (Intel, 0.10 ns/day), and 47,000 atoms (H200, 0.16 ns/day). Graph parallelism raises the capacity of a single device by 8.0x on H200 and 4.6x on A100. Across nodes, capacity grows to about 779,000 atoms on 32 Intel nodes, with a slower speed at 0.007 ns/day. On H200 and A100, a FLOP count of FairChem's graph-parallel predictor gives lower bounds of 31% and 51% of FP64 peak. From one omat checkpoint and with no refit, nine 100 ps NPT heating ramps (NaCl, equiatomic Pt-Ru, and equiatomic (HfNbTaTiZr)N, on all three node types) ran to completion.