Technology1 min read

Tackling GPU Bottlenecks in Plane-Wave DFT with K-Point Batching and Mixed Precision

By · Published by Everything Blog

In short

The standard GPU implementation of the plane-wave software VASP processes k-points sequentially and relies on FP64 precision, causing single-atom iron calculations to run eight times slower on a GPU than on a CPU. NVIDIA addressed this issue with Daviator, which groups k-points and spins into batched kernels centered on a batched Davidson eigensolver using resident GPU memory. Combined with mixed precision, this approach yields up to a 22-fold improvement in time-to-solution and cuts costs by up to 75 percent compared to CPU-only execution. Users can utilize Daviator without altering standard VASP input setups by enabling a single configuration parameter in the NVIDIA ALCHEMI VASP microserv…

Key points

  • Batching k-points and spins eliminates launch overheads that traditionally limit GPU perf…: Batching k-points and spins eliminates launch overheads that traditionally limit GPU performance when simulating small atomic systems.
  • Utilizing mixed precision takes advantage of modern workstation and inference-class GPUs…: Utilizing mixed precision takes advantage of modern workstation and inference-class GPUs that prioritize high FP32-to-FP64 throughput ratios.
  • Researchers can deploy these performance improvements seamlessly through the NVIDIA ALCHE…: Researchers can deploy these performance improvements seamlessly through the NVIDIA ALCHEMI VASP microservice with standard input setups.

Sources checked by our analysis pipeline: 1

Training a next-generation machine learning interatomic potential (MLIP) requires ground truth: tens to hundreds of millions of density functional theory (DFT) calculations of energies, forces, and stresses across diverse structures and chemistries.

Plane-wave DFT, with VASP as its dominant implementation, is the gold standard for generating that data. It is accurate, well-validated across materials classes, and ubiquitous at the national labs and research institutions running these calculations at scale.

As MLIP training datasets have grown, so has the pressure to generate DFT data faster and cheaper. But the small unit cells and dense k-point grids common in MLIP data generation can leave significant GPU performance unused. VASP’s OpenACC implementation has delivered meaningful speedups, especially for large-system calculations. However, for smaller systems, sequential k-point processing limits GPU utilization. At the same time, AI workloads are reshaping GPU hardware development toward lower-precision data types, creating a growing gap between what modern NVIDIA GPUs can deliver and what FP64-only HPC codes are optimized for.

This post introduces two new NVIDIA capabilities that address these bottlenecks. The first, Daviator, resolves the small-system GPU performance…

Technology