# What is ACEMD

ACEMD is a GPU-accelerated molecular-dynamics engine. Given a parameterised system and an input file, it can minimize the structure and propagate the integrator under your chosen thermostat, barostat, and restraints — covering minimization, equilibration, and production runs in the same engine. It does **not** build systems, prepare proteins, parameterise small molecules, or analyse trajectories; those tasks are handled upstream and downstream by [HTMD](https://software.acellera.com/htmd/) and [moleculekit](https://software.acellera.com/moleculekit/).

## What it does

- **Integrator on the GPU.** ACEMD is built on top of [OpenMM](http://openmm.org/), with custom tuning aimed at sustained single- and multi-GPU throughput.
- **Classical and machine-learned potentials.** Classical CHARMM, AMBER, and OpenMM XML force fields are first-class. Neural-network potentials run either as a hybrid (NNP/MM — molecule on NNP, rest on MM) or pure NNP — see [NNP and NNP/MM](nnp-and-nnp-mm.md).
- **Ensembles.** NVE, NVT (Langevin thermostat), and NPT (Monte Carlo barostat, isotropic / anisotropic / membrane).
- **Box shapes.** Orthorhombic and triclinic (truncated octahedra, rhombic dodecahedra).
- **Output.** XTC or DCD trajectories, optional velocity and force trajectories. CSV-formatted log with energies and timings.
- **Restart.** Periodic checkpoints; resumable on the same GPU model.

## Platform support

- **NVIDIA GPUs** via CUDA — the recommended target. Tested back to the *Maxwell* architecture; performance scales linearly with newer cards.
- **Other GPUs** via OpenCL.
- **CPU** for debugging — orders of magnitude slower than GPU, not for production.

ACEMD can be parallelised over multiple GPUs on a single host, though for typical system sizes splitting a single simulation across cards doesn't speed it up — see [Select GPU devices](../how-to/select-gpu-devices.md) for when multi-GPU actually helps.

## Integrations

- **[HTMD](https://software.acellera.com/htmd/)** — system building and analysis pipelines.
- **[PLUMED](https://www.plumed.org/)** — enhanced sampling (metadynamics, steered MD) and complex restraints. See [Run with PLUMED](../how-to/run-with-plumed.md).
- **[moleculekit](https://software.acellera.com/moleculekit/)** — reading, writing, and analysing trajectories.

## What it doesn't do

- No system preparation (protonation, mutation, solvation) — use [moleculekit](https://software.acellera.com/moleculekit/)'s {py:func}`~moleculekit.tools.preparation.systemPrepare` or HTMD.
- No parameterisation of non-canonical chemistry — small molecules, stapled peptides, non-canonical amino acids, modified residues. Use [HTMD](https://software.acellera.com/htmd/), which wraps OpenFF, GAFF/Antechamber, CGenFF, and related generators in a unified parameterisation pipeline, then bring the topology to ACEMD.
- No trajectory analysis — once the run finishes, hand `output.xtc` to moleculekit / HTMD / MDAnalysis / your tool of choice.

ACEMD is widely used in academic and industrial MD research; the original engine paper has been cited over 700 times.

## See also

- [Capabilities at a glance](../index.md) — the landing page summary.
- [Performance and benchmarks](performance-and-benchmarks.md)
- [Integrator and constraints](integrator-and-constraints.md)
- [NNP and NNP/MM](nnp-and-nnp-mm.md)