NVIDIA Introduces Daviator to Address GPU Bottlenecks in Plane-Wave DFT
Home
Latest reporting
News
NVIDIA’s Daviator and mixed precision execution layer aim to optimize GPU performance in plane-wave DFT, achieving up to 22x speedups and 75% cost reductions.
NVIDIA has unveiled two new capabilities designed to tackle GPU performance bottlenecks in plane-wave density functional theory (DFT), a computational method widely used in materials science and chemistry. The innovations, named Daviator and a mixed precision execution layer, aim to significantly enhance the efficiency of DFT calculations, particularly for small systems with dense k-point grids. These advancements promise to accelerate machine learning interatomic potential (MLIP) training, which relies on tens to hundreds of millions of DFT calculations.
Addressing GPU Utilization Challenges
Plane-wave DFT calculations, often implemented using the Vienna Ab initio Simulation Package (VASP), are computationally intensive and require high accuracy. However, small unit cells and dense k-point grids commonly used in MLIP data generation can leave GPU resources underutilized. NVIDIA’s OpenACC implementation for VASP has previously delivered speedups for large systems, but sequential k-point processing has limited GPU performance for smaller systems. Additionally, the dominance of double-precision (FP64) operations in traditional DFT codes has created a mismatch with modern NVIDIA GPUs, which are increasingly optimized for lower-precision data types such as FP32.
Daviator: Batching K-Points for GPU Saturation
Daviator introduces a library of GPU kernels that replaces sequential k-point processing with batched operations. Central to this innovation is a batched Davidson eigensolver, which reduces the number of GPU kernel launches and maximizes streaming multiprocessor (SM) occupancy. Data remains resident in GPU memory throughout the process, minimizing latency. Daviator integrates seamlessly with VASP, requiring no modifications to existing input files. Users can activate it by setting the “LDAVIATOR=True” parameter in the INCAR file.
Initial benchmarks demonstrate significant performance improvements. For example, a single-atom iron system achieved over 2x speedups compared to CPU execution, while higher-accuracy functionals like meta-GGA showed up to 22x faster time-to-solution. For mid-size systems, such as a 64-atom Bi-doped GaAs structure, execution times were reduced by more than 85%.
Mixed Precision Execution Layer
The second capability introduced by NVIDIA is a mixed precision execution layer, which leverages the higher FP32 throughput available on modern GPUs. By selectively using FP32 for most computational steps while retaining FP64 for critical accuracy-sensitive operations, this approach bridges the gap between traditional HPC codes and AI-optimized hardware. This is particularly beneficial for inference-class GPUs like the RTX PRO 6000, which feature a 64:1 FP32-to-FP64 throughput ratio.
Implications and Future Developments
These advancements not only reduce computational costs but also enable researchers to perform higher-accuracy DFT calculations at scale. Daviator is currently limited to single-node execution, with multi-node support and additional features, such as compatibility with non-collinear spin systems, expected in future releases. Overall, NVIDIA’s innovations represent a significant step forward in optimizing GPU performance for computational chemistry and materials science applications.
