LAMMPS on one of Swing’s A100 GPUs is about ten times faster than LAMMPS on a Bebop Broadwell Node.
- Speedups increase as the size of the model increases. For models containing more than 50,000 atoms you can expect speedups of more than an order of magnitude.
- Parallel scaling on multiple GPUs require models contains millions of atoms.
- Loading the module lammps/15Jun2023 will setup your environment to use this build of LAMMPS
- An example script for running LAMMPS on Swing’s GPUs can be found at /soft/lammps/15Jun2023/swing/example.
- To see which packages have been accelerated for the GPUs with KOKKOS in this LAMMPS build read the file /soft/lammps/15Jun2023/swing/example/lammps.help.output.