The Pulse
NVIDIA targets materials-calculation bottlenecks with Daviator
NVIDIA has introduced GPU kernels that batch calculations used to generate training data for machine learning interatomic potentials. It reports projected speed and cost gains against CPU execution, depending on hardware and configuration.

A new NVIDIA GPU-kernel library batches calculations that materials researchers use to generate training data for machine learning interatomic potentials. In a Nov. 18 technical post, NVIDIA said the library, called Daviator, achieved up to 22 times faster time-to-solution and up to 75% lower cost than CPU execution for a projected reference dataset workload, depending on hardware and configuration.
The company also described a mixed-precision layer intended to accelerate the calculations further on certain GPUs. The results concern the generation of reference data, rather than the training of an interatomic potential itself.
Small systems, many calculations
Training a next-generation machine learning interatomic potential requires tens to hundreds of millions of density functional theory, or DFT, calculations of energies, forces and stresses across different structures and chemistries. Daviator targets calculations run with VASP, a plane-wave DFT implementation.
In periodic systems, calculations sample a discrete set of points known as k-points. VASP's OpenACC GPU implementation processes those points sequentially, even though each solve is largely independent. For small systems with dense grids, the resulting dispatch overhead can dominate compute time.
NVIDIA says a single-atom iron system with 752 k-points ran eight times slower on more expensive GPU hardware than on a CPU under the existing approach. Large systems can provide enough work per k-point to fill a GPU, but smaller systems do not necessarily do so.
Batching the work
Daviator batches work across k-points and spins, including through a batched Davidson eigensolver. Data remains in GPU memory while VASP retains control of the self-consistent-field loop, mixing, symmetry and input and output.
The benchmarks compared Daviator with VASP 6.6.0 CPU and OpenACC baselines on NVIDIA H100, H200 and B200 GPUs. The company reports that the single iron atom ran more than twice as fast as the CPU baseline with Daviator. For a 64-atom bismuth-doped gallium arsenide structure, it says Daviator cut execution time by more than 85%.
Precision and access
The separate mixed-precision layer uses FP32 arithmetic and storage for selected Davidson solver components while retaining FP64 for numerically sensitive steps. NVIDIA says this provides roughly twice the end-to-end acceleration on GPUs with high FP32-to-FP64 throughput ratios, such as the RTX PRO 6000.
The precision analysis uses configurable floating-point types to screen individual kernels against a reference. Its workflow repeatedly reduces the mantissa and exponent bit counts until a comparison fails, helping identify where lower precision can be used and where higher precision is needed.
Daviator will be offered through the NVIDIA ALCHEMI VASP microservice, enabled by setting LDAVIATOR=True in the INCAR file. Existing VASP inputs work without modification.
For now, VASP with Daviator is limited to single-node execution. Multi-node support, coverage for systems with non-collinear spins and explicit gamma-only runs are expected at a later stage.
Sources
- Original source · Blog post NVIDIA Technical Blog developer.nvidia.com
Companies in this story
Related stories
- Mat3ra adds NVIDIA ALCHEMI to browser-based materials simulations
- AMD details GPU-initiated DMA transfers for multi-GPU workloads
- WEKA-backed scorecard challenges GPU use as an AI inference metric
- TurboPairFormer speeds OpenFold3 training up to 1.73× in tests
- AWS adds AI data preparation tools to SageMaker AI