cuPotential benchmarks on Leonardo
We’ve shared scripts for running NVIDIA’s cuPotential with OpenMM on the Leonardo supercomputer. In addition to AMBER, we benchmarked CHARMM36 and CHARMM36m for DHFR on a single A100 GPU. In OpenMM’s default CHARMM36 setup, cuPotential speeds up electrostatics, but the separate Lennard-Jones calculation limits the overall gain to 2–11%. We moved the Lennard-Jones terms that follow standard mixing rules into the force that cuPotential accelerates and kept CHARMM36m’s pair-specific corrections in OpenMM. With this change, cuPotential ran 2.6–3.1× as fast as standard OpenMM using the same reorganized force terms. Setup instructions, benchmark results, and run logs