NVIDIA open-sources Daviator GPU kernel library to accelerate VASP DFT calculations
NVIDIA's Daviator library replaces sequential k-point processing in VASP with batched GPU operations, achieving up to 22x speedup for small-system plane-wave DFT workloads that previously ran slower on GPU than CPU.
Summary
NVIDIA has released Daviator, an open-source library of GPU kernels designed to eliminate performance bottlenecks in plane-wave density functional theory (DFT) calculations using VASP, the dominant implementation in materials science. The library addresses a critical inefficiency where small unit cells with dense k-point grids caused sequential GPU kernel launches to run up to 8x slower than equivalent CPU execution. Daviator introduces batched operations including a batched Davidson eigensolver that replaces sequential loops with batched FFTs, GEMMs, and custom kernels, keeping data resident in GPU memory throughout computation.
The release also includes a configurable mixed precision execution layer that uses FP32 arithmetic for selected solver components while maintaining FP64 for numerically sensitive operations, delivering roughly 2x additional acceleration on GPUs with high FP32:FP64 throughput ratios such as the RTX PRO 6000. Both capabilities are available through the NVIDIA ALCHEMI VASP microservice via INCAR file flags without requiring modifications to existing VASP inputs, though the underlying kernel library itself is open-source.
Daviator architecture and mixed precision implementation
- Batched Davidson eigensolver: Replaces VASP's sequential per-k-point processing with batched GPU operations, drastically reducing kernel launch overhead and raising SM occupancy from fractional capacity to near saturation for small systems with hundreds to thousands of k-points.
- Memory residency: Data remains in GPU memory throughout the SCF loop; VASP retains control of the outer SCF iteration, mixing, symmetry, and I/O while Daviator integrates at well-defined call boundaries over data already on device.
- Mixed precision layer: Uses configurable floating-point types to identify minimum required precision per kernel, implementing FP32 arithmetic and storage for selected Davidson components with FP64 retained for convergence-critical operations.
- Hardware-targeted acceleration: The mixed precision approach specifically benefits inference-class and workstation GPUs with high FP32:FP64 ratios (64:1 on RTX PRO 6000), addressing the growing divergence between AI-optimized GPU hardware and traditional FP64-only HPC codes.
- Integration mechanism: Activated through the NVIDIA ALCHEMI VASP microservice via
LDAVIATOR=TrueandLDAVIATOR_MIXED_PRECISION=1environment flags in the INCAR file, with existing VASP inputs (POSCAR, INCAR, KPOINTS) requiring no modification.
Why it matters
Daviator directly addresses a pain point that has forced computational materials scientists to manually route small DFT jobs to CPU rather than GPU, improving hardware utilization for MLIP training datasets that require tens to hundreds of millions of calculations. The open-source release of the kernel library, combined with NVIDIA's use of standard interfaces and configurable precision through CCCL generic floating-point types, provides a template for adapting other legacy FP64 HPC codes to modern AI-optimized GPU architectures without sacrificing numerical reliability where it matters.
Original reference: developer.nvidia.com →