NVIDIA Launches ALCHEMI Toolkit for GPU-Accelerated Chemistry Simulations

Luisa Crawford Apr 14, 2026 16:51

NVIDIA releases ALCHEMI Toolkit enabling researchers to build custom atomistic simulation workflows with up to 33x speedups for batched molecular dynamics on GPUs.

NVIDIA has released its ALCHEMI Toolkit, a PyTorch-native framework that lets researchers build custom GPU-accelerated atomistic simulation workflows for chemistry and materials science applications. The toolkit addresses a persistent bottleneck in computational chemistry: while machine learning interatomic potentials (MLIPs) run efficiently on GPUs, the surrounding simulation infrastructure has remained stuck on legacy CPU-centric code.

The release expands on NVIDIA's ALCHEMI ecosystem, first announced at Supercomputing 2024. Early adopters are already reporting significant performance gains. Orbital, which develops AI foundation models for sustainable materials discovery, achieved approximately 1.7x acceleration for large systems and 33x speedups for batched smaller systems using the toolkit's GPU-accelerated graph construction.

What the Toolkit Actually Does

ALCHEMI Toolkit sits between domain-specific GPU kernels and deep learning models, managing data flow and enabling composable simulation workflows. The initial release includes geometry relaxation, molecular dynamics capabilities, and pipeline infrastructure for combining multiple simulation stages.

Key technical features include batched dynamics kernels, FIRE and FIRE2 optimizers, and integrators for various thermodynamic ensembles including Langevin thermostats and constant-pressure molecular dynamics. The toolkit supports neighbor list construction, DFT-D3 dispersion corrections, and long-range electrostatic interactions through its Toolkit-Ops layer.

System requirements include Python 3.11 or higher, PyTorch 2.8+, CUDA Toolkit 12+, and an NVIDIA GPU with compute capability 7.0 or greater (RTX 20-series and newer).

Industry Partners Already Integrating

Three major platforms have announced integrations. Orbital is incorporating ALCHEMI Toolkit into their OrbMolv2 model, using PME electrostatics for periodic Coulomb interactions and the MTK integrator for batched molecular dynamics.

Materials Graph Library (MatGL), an open-source framework for graph-based MLIPs, is integrating with the toolkit's TensorNet model to accelerate materials simulations while reducing memory consumption.

Matlantis, which combines universal MLIPs with cloud computing for industrial materials discovery, reports speedups of up to 10x using ALCHEMI Toolkit-Ops components including Warp-optimized neighbor list construction. The company is exploring how the composable dynamics could enable parallel relaxation of millions of molecular configurations.

Technical Architecture

The toolkit offers two scaling approaches. FusedStage wraps end-to-end workflows in torch.compile with shared CUDA stream contexts for single-GPU deployment. For multi-GPU scenarios, a distributed pipeline approach splits stages across devices—the documentation example shows eight GPUs running geometry optimization while another eight handle Langevin dynamics.

Data management eliminates the traditional "memory tax" of CPU-GPU transfers. The AtomicData and Batch objects keep simulation data GPU-resident throughout execution, with native support for ASE and Pymatgen interfaces plus Zarr-based storage for efficient batch writing.

The framework supports custom model integration through standardized wrappers, with built-in support for MACE, TensorNet, and AIMNet2 architectures.

The toolkit is available now on GitHub under the NVIDIA/nvalchemi-toolkit repository. JAX support is scheduled for the v0.2.0 release.

• nvidia
• alchemi
• computational chemistry
• machine learning
• gpu computing