RESEARCH · RESEARCH · #3838
NVIDIA's Daviator adds k‑point batching and mixed precision to VASP, boosting small-system DFT on GPUs
Daviator, an NVIDIA library of GPU kernels, replaces VASP’s sequential per–k-point processing with k-point/spin-batched GPU operations (including a batched Davidson eigensolver), eliminating small-system GPU inefficiencies. Benchmarks on a projected OMat24 workload report up to 22x faster time-to-solution and up to 75% lower cost vs CPU; an optional mixed-precision layer uses FP32 for selected Davidson components while keeping critical steps in FP64, providing roughly 2x additional speedup on GPUs with high FP32:FP64 ratios (e.g., RTX PRO 6000). Daviator is exposed via the NVIDIA ALCHEMI VASP microservice (enable with LDAVIATOR=True and LDAVIATOR_MIXED_PRECISION=1 in INCAR) and works with existing VASP inputs.
KEY POINTS
- Daviator, an NVIDIA library of GPU kernels, replaces VASP’s sequential per–k-point processing with k-point/spin-batched GPU operations (including a batched Davidson eigensolver), eliminating small-system GPU inefficiencies.
- Benchmarks on a projected OMat24 workload report up to 22x faster time-to-solution and up to 75% lower cost vs CPU; an optional mixed-precision layer uses FP32 for selected Davidson components while keeping critical steps in FP64, providing roughly 2x additional speedup on GPUs with high FP32:FP64 ratios (e.g., RTX PRO 6000).
- Daviator is exposed via the NVIDIA ALCHEMI VASP microservice (enable with LDAVIATOR=True and LDAVIATOR_MIXED_PRECISION=1 in INCAR) and works with existing VASP inputs.
WHY IT MATTERS
This removes a major GPU utilization bottleneck in plane‑wave DFT for small unit cells and cuts time/cost for large DFT datasets used as MLIP training ground truth while exploiting modern GPUs' FP32 strengths.