NVIDIA has released Daviator, a library of GPU kernels for plane-wave density functional theory (DFT) calculations in VASP, designed to eliminate small-system performance bottlenecks and enable mixed-precision execution. The company reports that for projected reference dataset workloads, the combined capabilities deliver up to 22x faster time-to-solution and up to 75% lower cost compared to CPU execution.
The release targets the data generation pipeline for machine learning interatomic potentials (MLIPs), which requires tens to hundreds of millions of DFT calculations. VASP remains the dominant implementation for generating this ground truth data due to its accuracy and validation across materials classes. However, as MLIP datasets have grown, the pressure to generate DFT data faster has exposed hardware inefficiencies. Small unit cells with dense k-point grids leave significant GPU performance unused, and the stagnation of FP64 throughput relative to FP32 on modern GPUs has widened the gap between HPC code capabilities and available hardware.
Batching k-points to saturate GPUs
In periodic systems, electronic wavefunctions are parameterized by crystal momentum, sampled by a discrete set of k-points. For small unit cells, hundreds to thousands of irreducible k-points are often required for convergence. Each k-point solve is largely independent, yet VASP's OpenACC GPU port processes them sequentially. For large systems, the work per k-point can saturate a GPU, but for small systems, the hundreds of small GPU dispatches per step cause kernel launch overhead to dominate compute time. Previously, a single-atom iron system with 752 k-points ran 8x slower on GPU hardware than on CPU, forcing researchers to route jobs manually based on system size.
Daviator replaces this sequential processing with batched GPU operations. At the core is a batched Davidson eigensolver that handles k-points and spin in parallel, replacing sequential loops with batched FFTs, GEMMs, and custom kernels. This drastically reduces the number of GPU kernel launches, allowing streaming multiprocessor occupancy to move from a fraction of capacity to near saturation. Data remains resident in GPU memory throughout the calculation. VASP retains ownership of the self-consistent field loop, mixing, symmetry, and I/O; Daviator integrates at well-defined call boundaries over data already on device.
Benchmarks on NVIDIA H100, H200, and B200 GPUs show significant improvements across system sizes. The single iron atom achieves over 2x speedup relative to the CPU baseline. Gains increase for higher-accuracy functionals: the Davidson solver on a LiFePO4 structure with the r2SCAN functional sees a 22x speedup, enabling meta-GGA runs at production scale. For mid-size systems, such as a 64-atom Bi-doped GaAs structure, VASP with Daviator cuts execution time by more than 85%. Even at 348 atoms with a single k-point, Daviator delivers a significant improvement over OpenACC and CPU baselines, eliminating the need to route jobs between CPU and GPU based on system size.
Mixed precision for inference-class hardware
Daviator initially relied on FP64 arithmetic and storage. NVIDIA has now added a mixed-precision execution layer to leverage the high FP32 throughput available on modern hardware. This is particularly relevant for inference-class and workstation GPUs like the RTX PRO 6000, which offers a 64:1 FP32-to-FP64 throughput ratio. On data center GPUs like the H100 and H200, FP64 Tensor Core throughput has held flat while FP32 throughput has increased, though Hopper and Blackwell architectures maintain a 1:1 ratio between FP64 and FP32 Tensor Cores.
Implementing mixed precision in a Davidson solver is non-trivial. Hard failures, such as a solver rejecting an ill-conditioned input, are obvious, but soft failures like energy drift, incorrect convergence, or diverging molecular dynamics trajectories can appear valid on casual inspection. NVIDIA's approach involved a precision requirements analysis workflow. The team instrumented the code with a generic floating-point type that allows configurable mantissa and exponent bits. By running screening scripts that decrease bit counts until results diverge from a reference, they mapped the precision requirements for each kernel.
Illustrative image: Magnascan / Pixabay
This analysis revealed that large parts of the Davidson algorithm could run in FP32. The FFTs, the application of the Hamiltonian, and the orbital updates run in FP32, while the subspace eigenproblem, the tall-skinny GEMMs behind the subspace matrices, and the convergence measures remain in FP64. The tall-skinny GEMMs in the Rayleigh-Ritz projection proved to be the remaining FP64 bottleneck; pure FP32 accumulation is insufficient because the output feeds directly into a numerically sensitive eigenvalue solver. The implementation uses a split-k algorithm that breaks dot products into FP32 chunks with a final FP64 accumulation step to resolve this.
Performance gains from mixed precision depend heavily on the hardware ratio. On H100 and H200, where FP64 Tensor Cores match FP32 throughput, mixed precision delivers only marginal end-to-end gains. The benefit concentrates on hardware with a high FP32:FP64 ratio. On the RTX PRO 6000, the FP32 Davidson solver runs 3x to 7x faster, delivering roughly 2x end-to-end acceleration. On GaAsBi-64, FP32 enables a 2x speed-up at an EDIFF of 10^-5 eV, and the combined FP32-to-FP64 run is 1.5x faster than pure FP64 at an EDIFF of 10^-8 eV.
Convergence is controlled by the EDIFF parameter, which specifies the threshold for total energy change between electronic iterations. Mixed precision converges to an EDIFF of 10^-4 to 10^-6 eV depending on the input, while continuing with FP64 allows the calculation to converge below the FP32 noise floor. Users enabling mixed precision must configure switching thresholds carefully. The calculation starts in mixed precision with FP32 storage and switches to FP64 based on either an energy difference threshold or a step count.
| INCAR Parameter | Value | Description |
|---|---|---|
| LDAVIATOR_MIXED_PRECISION | True/False | Enable/Disable mixed precision |
| DAVIATOR_EDIFF_FP32 | Real number > 0 | EDIFF threshold for FP32 |
| DAVIATOR_SWITCH_TO_FP64_AFTER_STEP | Integer N >= 0 | Switch to FP64 after step N |
The DAVIATOR_EDIFF_FP32 parameter sets the energy threshold, typically between 10^-4 and 10^-5 eV. The DAVIATOR_SWITCH_TO_FP64_AFTER_STEP parameter controls the switching logic, allowing users to switch after a fixed number of steps, switch dynamically after convergence, or run entirely in FP32.
Availability and constraints
Daviator and its mixed-precision option are available through the NVIDIA ALCHEMI VASP microservice. A VASP user license is required. Users enable the library by setting LDAVIATOR=True in the INCAR file. Existing VASP inputs work without modification; Daviator intercepts the Davidson solver internally when ALGO = Normal.
Mixed precision is enabled by setting LDAVIATOR_MIXED_PRECISION=True. Users can configure the precision behavior using the parameters listed above. For example, a run that uses FP32 until EDIFF reaches 5x10^-5 eV and then switches to FP64 would set DAVIATOR_SWITCH_TO_FP64_AFTER_STEP=-2 and DAVIATOR_EDIFF_FP32=5e-05.
There are several constraints to note. Daviator currently supports single-node execution only. Multi-node support, coverage of systems with non-collinear spins, and explicit gamma-only runs are expected to be released at a later stage. Some GPU kernels are specialized for the batch sizes they encounter and may trigger compilation on first use. Users should set CUDA_CACHE_PATH to fast storage, such as /tmp, and ensure CUDA_CACHE_DISABLE=0 to avoid latency during initial runs.
Sources
PNEUMETRON EDITORIAL
Pneumetron's pipeline pairs AI-assisted drafting with human editorial review before publishing — our goal is to make staying informed easier, not to replace real reporting.




