AI · Nov 9, 2026
OpenAI’s Control Room puts a fail-closed kill switch inside the inference engineNvidia's Daviator aims to make GPUs the default for VASP, even on tiny crystals
· Updated · Signal Desk newsroom
Analysis of reporting by NVIDIA Developer. How we work

Batched k-point kernels and a carefully mapped mixed-precision solver target two weaknesses of GPU DFT. Nvidia's projections promise large savings for building AI training data, though the figures are its own and the tool sits behind a microservice.
Key takeaways
- Daviator replaces VASP's sequential per-k-point GPU work with batched kernels, including a batched Davidson eigensolver, to keep GPUs busy on small systems.
- Nvidia cites a single-atom iron case with 752 k-points that previously ran 8x slower on GPU than CPU; with Daviator it runs over 2x faster than the CPU baseline.
- A mixed-precision mode runs much of the solver in FP32 and switches to FP64 near convergence; Nvidia reports about 2x end-to-end gains on RTX PRO 6000 but marginal gains on H100 and H200.
- For a projected OMat24-scale dataset, Nvidia estimates up to 22x faster time-to-solution and 60% to 75% lower cost than CPU, about $5 million to $6 million.
- Daviator is single-node for now, requires a VASP license and will ship through the NVIDIA ALCHEMI VASP microservice.
Why DFT throughput suddenly matters to AI
Machine learning interatomic potentials, the models that let chemists simulate materials far faster than quantum mechanics alone, need enormous training sets. Nvidia notes that next-generation models want tens to hundreds of millions of density functional theory calculations of energies, forces and stresses. Plane-wave DFT, and VASP in particular, is the trusted way to generate that ground truth. So the speed and cost of VASP has become a bottleneck for AI in materials science, a theme that ran through our coverage of tensor-core speedups for interatomic potentials.
The awkward fact is that the small unit cells common in these datasets are exactly where GPUs have struggled. VASP's OpenACC port processes k-points one at a time. For small cells that need hundreds or thousands of k-points, the GPU spends its time launching tiny jobs. Nvidia's example is stark: a single iron atom with 752 k-points ran 8x slower on GPU than on CPU, forcing researchers to route jobs by size.
Batching the problem away
Daviator is Nvidia's library of GPU kernels that batches k-points and spins. At its core is a batched Davidson eigensolver built on batched FFTs, matrix multiplies and custom kernels, which cuts the number of kernel launches and lets the GPU approach full occupancy while data stays on the device. VASP still owns the self-consistent loop, mixing, symmetry and I/O; Daviator plugs in at defined boundaries and is switched on with a single input flag.
In Nvidia's benchmarks against VASP 6.6.0 on H100, H200 and B200, the single iron atom runs more than 2x faster than CPU, a LiFePO4 calculation with the r2SCAN functional sees a 22x Davidson speedup, and execution time for a 64-atom bismuth-doped gallium arsenide system drops by more than 85%. Nvidia's claim that Daviator beats prior results at all system sizes would, if it holds, end the size-based routing problem. Current limits matter: single-node only, with multi-node, non-collinear spin and gamma-only runs promised later.
Precision as an engineering discipline
The more broadly useful part of the post is about precision. Data center GPUs have held FP64 throughput flat across Hopper, Blackwell and Rubin while FP32 keeps rising, and workstation cards like the RTX PRO 6000 have a 64-to-1 FP32-to-FP64 ratio. Codes that insist on FP64 everywhere are leaving performance unused. Nvidia's approach is to measure, kernel by kernel, the minimum mantissa and exponent bits each step needs, and then run FFTs, the Hamiltonian and orbital updates in FP32 while keeping the eigenproblem, sensitive matrix products and energies in FP64.
That honesty about failure modes is welcome. The post warns that soft failures such as energy drift or wrong convergence can look valid at a glance, which is exactly the risk scientists should worry about. The payoff depends on hardware: about 2x end-to-end on RTX PRO 6000, but only marginal gains on H100 and H200, where FP64 tensor cores match FP32. Researchers should validate results against FP64 on their own systems before trusting mixed-precision data at scale.
The cost claim, read carefully
Nvidia projects that generating a dataset the size of OMat24, over 100 million VASP data points, would be up to 22x faster than CPU and 60% to 75% cheaper, saving roughly $5 million to $6 million, with the lowest cost about a quarter of CPU on RTX PRO 6000 with FP32. These are extrapolations from 13 benchmarked inputs, priced against an average of on-demand rates from AWS, Azure, CoreWeave and Google, and using a more precise projector setting than OMat24 itself. They are plausible and useful for planning, but they are a vendor projection, not a measured run. Access through Nvidia's ALCHEMI microservice also means another layer between scientists and their code.
Who feels it
- Computational materials scientists
- Small-cell, dense k-point jobs may no longer need to stay on CPU.
- MLIP builders
- Cheaper ground-truth data could enlarge training sets for materials models.
- HPC centers
- Workstation-class GPUs become competitive for some DFT workloads.
- VASP licensees
- Benefits arrive through an Nvidia microservice, adding a dependency.
What to watch
- Release timing for Daviator multi-node, non-collinear and gamma-only support.
- Independent benchmarks validating accuracy of mixed-precision results.
- Whether large dataset projects adopt Daviator for their next generation.
- Similar batching or mixed-precision work in open-source DFT codes.
FAQ
What is Daviator?
An Nvidia library of GPU kernels that batches k-points and spins in plane-wave DFT, including a batched Davidson eigensolver, to improve GPU utilization in VASP.
Does mixed precision help on all GPUs?
No. Nvidia reports about 2x end-to-end gains on RTX PRO 6000, but only marginal gains on H100 and H200, where FP64 tensor cores match FP32 throughput.
How do researchers get Daviator?
Through the NVIDIA ALCHEMI VASP microservice, with a VASP user license, by setting the LDAVIATOR flag in the INCAR file.
How much could Daviator save?
Nvidia projects 60% to 75% lower cost than CPU, about $5 million to $6 million for an OMat24-sized dataset. That is a vendor estimate.
More on AI
AI · Nov 11, 2026
OpenAI’s $2 million for children’s AI safety research matters for one promise: publishing bad newsAI · Oct 28, 2026
RTX Spark laptops get open speech models and a dial for background agentsAI · Nov 17, 2026
OpenAI says its models now shave area off production chips, including AMD’s next flagshipsAI · Nov 5, 2026
NVIDIA’s four open Cosmos 3 Nano models let robot and driving teams simulate worlds that react