NVIDIA releases cuPotential molecular-dynamics primitives: its OpenMM plugin tops 3x end-to-end speedup on selected benchmarks
Synopsis
NVIDIA released cuPotential, a molecular-dynamics library that exposes modular primitives through a C API and ships an OpenMM plugin replacing the nonbonded force, Langevin integrator and Monte Carlo barostat. Across six systems from 24,000 to 400,000 atoms under NPT, newer RTX PRO 6000 Blackwell GPUs reached over 3x end-to-end speedup in selected benchmarks, while other tested GPUs often gained around 2x or more; these numbers come from standard, non-alchemical systems.
Interpretation
In molecular dynamics, long-range interactions under periodic boundary conditions, namely electrostatics and dispersion, remain the main computational and engineering bottleneck; cuPotential splits this work into modular primitives that developers can call selectively for acceleration. MD codes were among the first ported to GPUs, but as GPUs evolve, kernel optimization and compatibility across heterogeneous hardware have become harder than ever for many MD engine developers. With the OpenMM plugin, systems from 24,000 to 400,000 atoms (DHFR, TYK2, BACE, ApoA1, 100,000-atom water and cellulose) were tested under NPT; newer RTX PRO 6000 Blackwell GPUs exceeded 3x end-to-end speedup in selected benchmarks, and B200, H100, A100, L40S, A10, L4 and T4 often gained around 2x or more.
The library has two parts, a core methods library and an OpenMM plugin that supplies replacements for the nonbonded force, the Langevin integrator and the Monte Carlo barostat, aiming at drop-in replacement with minimal code changes. Integration requires replacing the system's OpenMM nonbonded force before creating the simulation context and switching to the cuPotential integrator, with its barostat optional for NPT. The sample systems use the AMBER14/TIP3P force field, a 0.9 nm nonbonded cutoff, a 5×10⁻⁴ electrostatic accuracy target and mixed precision, with a 4 fs timestep paired with constrained hydrogen bonds, rigid water and hydrogen mass repartitioning to 1.5 amu; throughput runs use 500 warm-up steps followed by 3,000 timed steps.
New C2 continuous splitting and spreading functions reduce the real-space cutoff or reciprocal-space grid size with no loss in accuracy, measured as L2 force RMSE at OpenMM default settings, and substantially lower energy drift. The PME variant of Ewald splitting has been the de facto method for 30 years, while newer techniques such as u-series, midtown splines and prolates have not been widely adopted because of their technical demands. On a 10 ns, 10,000-molecule rigid water system, cuPotential with a 9 Å cutoff reduced drift several hundred fold versus PME default settings, with converged PME using a 20 Å cutoff, 2 fs timesteps and 300 K, with drift normalized over 60,000 degrees of freedom.
Perspective
The demonstrated scope is solvated, approximately uniform-density systems in rectangular periodic boxes, typical of biomolecular simulations; the plugin accepts PME source forces and does not support LJPME, triclinic boxes or OpenMM particle or exception parameter offsets, so the plugin README should be consulted before adapting an existing simulation. Performance figures were measured on DHFR, TYK2, BACE, ApoA1, 100,000-atom water and cellulose systems from 24,000 to 400,000 atoms, and on standard, non-alchemical systems.
The dispersion accuracy gain is described as several thousand fold in dispersion energy versus conventional methods while keeping comparable force errors at significantly reduced computational cost; the text notes that reducing the cutoff to 6 Å still poses enormous practical challenges because of NBFix and combining rules. The u-series was not optimized for force error but for ensemble error, and its parameters were chosen to match OpenMM's force-error target.
