Distributed memory, GPU accelerated Fock construction for hybrid, Gaussian basis density functional theory.

Area of Science:

• Computational Chemistry
• Materials Science
• High-Performance Computing

Background:

• Modern supercomputers increasingly utilize graphics processing units (GPUs) for accelerated computations.
• Optimizing electronic structure methods for massively parallel GPU architectures is a growing research priority.
• Existing GPU acceleration for Gaussian basis atomic orbital methods often targets shared-memory systems, limiting massive parallelism.

Purpose of the Study:

• To develop and present distributed memory algorithms for evaluating Coulomb and exact exchange matrices in hybrid Kohn-Sham Density Functional Theory (DFT).
• To enable efficient electronic structure calculations on large-scale systems using Gaussian basis sets.
• To address the need for massively parallel GPU algorithms in quantum chemistry.

Main Methods:

• Implementation of distributed memory algorithms for direct density-fitted (DF-J-Engine) Coulomb matrix evaluation.
• Development of distributed memory algorithms for seminumerical (sn-K) exact exchange matrix evaluation.
• Utilized hybrid Kohn-Sham DFT with Gaussian basis sets.

Main Results:

• Demonstrated absolute performance of the developed distributed memory algorithms.
• Showcased strong scalability for systems ranging from hundreds to over a thousand atoms.
• Validated performance on up to 128 NVIDIA A100 GPUs on the Perlmutter supercomputer.

Conclusions:

• The presented distributed memory algorithms effectively exploit massively parallel GPU resources for electronic structure calculations.
• The methods show significant performance and scalability improvements for large atomic systems in hybrid DFT.
• This work advances the capability of large-scale quantum chemistry simulations on modern supercomputing architectures.