AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Accelerate Molecular Dynamics Simulations with Accurate, Flexible Primitives Using NVIDIA cuPotential

Collected Nov 16, 2026

NVIDIA has published cuPotential, a focused library aimed at molecular dynamics (MD) that exposes modular primitives through a C API, along with an initial plugin for OpenMM. The announcement appeared on the NVIDIA Developer Blog and the code is available in the NVIDIA/cuPotential GitHub repository, which hosts examples, benchmarks and integration code. The work targets the costliest part of classical force-field simulations: evaluating long-range interactions under periodic boundary conditions, chiefly electrostatics and dispersion.

The design philosophy is composability rather than a monolithic engine. cuPotential offers direct-space and reciprocal-space utilities such as cutoff-based interaction evaluators and charge-spreading functions, and developers can pull individual layers into their own MD stack. The OpenMM plugin is presented as a reference integration, with the stated goal of drop-in replacement and minimal changes to existing user code.

In the OpenMM plugin, three components are replaced: the nonbonded force, the Langevin integrator and the Monte Carlo barostat. In practice a user constructs a system as usual, then adds an nv.NonbondedForce wrapping the existing OpenMM force and removes the original; the cuPotential integrator handles center-of-mass motion removal, so the separate OpenMM CMMotionRemover is dropped. NPT simulations add the cuPotential MonteCarloBarostat with a chosen attempt frequency. The sample code uses AMBER14/TIP3P, a 0.9 nm nonbonded cutoff, a 5e-4 electrostatic accuracy target, mixed precision and a 4 fs timestep with constrained hydrogens, rigid water and hydrogen mass repartitioning to 1.5 amu.

Building requires Linux x86-64, a supported NVIDIA GPU, a driver compatible with CUDA 13.1, and Conda or Micromamba. The documented route creates an environment with Python 3.12, OpenMM 8.4.0 and the CUDA 13.1 toolkit, installs pybind11 and nvidia-mathdx, edits a build helper so it points at the active environment, then runs the plugin build script and builds and installs a wheel. A bundled sample exercises DHFR, BACE and ApoA1 under both NVT and NPT, ending in PASS.

Benchmarks cover six systems from roughly 24,000 to 400,000 atoms — DHFR, TYK2, BACE, ApoA1, a 100,000-atom water box and cellulose — across GPUs including T4, L4, A10, L40S, A100, H100, B200 and the RTX PRO 6000 Blackwell in workstation and server editions. Under NPT conditions, the newest Blackwell cards reach over 3x end-to-end speedup on selected benchmarks, while the older GPUs typically show around 2x or more. The figures were measured on standard, non-alchemical systems, and because the cuPotential arm uses its own force, integrator and barostat, the speedup describes that integrated configuration rather than an isolated kernel.

Accuracy methodology separates error from throughput. Electrostatics is compared against a double-precision reference using twice the cutoff for the short-range part, twice the mesh resolution per dimension for the long-range part and a fixed Ewald splitting parameter; truncated Lennard-Jones uses a same-cutoff double-precision reference. Throughput runs use 500 warm-up steps followed by 3,000 timed steps.

The scientific contribution is a new set of C2 continuous splitting and spreading functions developed with D. E. Shaw Research and the Flatiron Institute's Center for Computational Mathematics. Particle Mesh Ewald has dominated the field for about 30 years, and alternatives such as the u-series, midtown splines and prolates have seen limited adoption given their technical demands. NVIDIA says its functions reach target force accuracy at lower grid resolution and reduced cutoffs compared with Ewald and u-series. A caveat is offered in the post: the u-series was tuned for ensemble error rather than force error, and parameters were chosen to match a desired force error for a fair comparison against OpenMM.

Dispersion gets similar treatment. Lattice dispersion interactions can be handled with the same splitting and spreading machinery, which matters for highly nonpolar and membrane systems where conventional homogeneous analytical corrections can produce severe artifacts. NVIDIA reports several-thousand-fold improvement in dispersion energy accuracy while keeping comparable force errors at lower computational cost. The post also notes that pushing cutoffs down to 6 Å remains difficult in practice because of NBFix parameters and combining rules.

Energy drift over 10 ns was measured on a 10,000-molecule rigid water system at 300 K with 2 fs steps, 9 Å cutoffs for cuPotential and standard OpenMM, and a 20 Å cutoff for converged PME reference, normalized over 60,000 degrees of freedom. cuPotential reduced drift by several hundred fold relative to default PME settings, which NVIDIA says exceeds typical biomolecular requirements. Drift matters because slow energy leakage can bias long trajectories; a force library that is not the dominant drift source lets users keep production settings.

The current plugin has a deliberately narrow scope: solvated, approximately uniform-density systems in rectangular periodic boxes, accepting PME source forces. LJPME, triclinic boxes and OpenMM particle or exception parameter offsets are not supported, so teams with those setups will need to check the README before adapting existing simulations.

Why it matters: MD engine developers and computational chemistry groups now have a set of reusable GPU primitives they can embed rather than rewriting long-range kernels for each new architecture, and OpenMM users can test the speedups with a small amount of substitution code. The main trade-offs are the plugin's restricted box and force-type support and the need for CUDA 13.1-era tooling, which suggests early adopters should validate against their own accuracy references before switching production pipelines.

Read at NVIDIA Developer Blog

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Molecular dynamics (MD) is an indispensable tool across physics, chemistry, biology, and materials science. Owing to its arithmetic structure, particularly the...