NVIDIA open-sources cuPotential, speeding molecular dynamics up to ~4x on GPUs

AllThingsApx · x · 2026-11-17

NVIDIA released cuPotential, a CUDA-X library for efficiently computing molecular interactions from standard force fields in molecular dynamics simulations, usable via an OpenMM Python plugin or an experimental C API.

Official benchmarks (integrated with OpenMM 8.4.0, NPT mixed precision, vs stock OpenMM throughput) show roughly 3–4x speedups across systems from 24k to 409k atoms (DHFR, TYK2, BACE, water, ApoA1, cellulose): RTX PRO 6000 Blackwell up to 3.95x, L40S 3.99x, H100 PCIe 3.34x, B200 3.24x, T4 2.64x. The code is open-sourced on GitHub with docs, samples, and benchmark details.

Related event: NVIDIA Open-Sources cuPotential, Delivering 3-4x Real-World Molecular Dynamics Speedups(10 posts)→

Original post →

More from Infra

Infra channel →