Improvements to GPU support in LAMMPS

LAMMPS currently supports GPU acceleration through the KOKKOS and GPU packages. While the KOKOS package has been under active development, the GPU package is mostly under maintenance mode with occasional updates. To many LAMMPS users, it is useful to know which package would give optimal performance for a given problem. In this talk, I will present my recent contribution to both packages by adding new features and improving the GPU speedup. On the ALCF Polaris machine, the performance with either packages on NVIDIA A100 GPUs is 3X-16X faster than the MPI-only runs on a 32-core node, depending on the nature of simulated systems: simple liquids, solvated proteins and dipolar particles. I will conclude by discussing the considerations for choosing the suitable package for optimal performance.