Porting Fragmentation Methods to Graphical Processing Units Using an OpenMP Application Programming Interface:...

Area of Science:

• Computational Chemistry
• High-Performance Computing
• Quantum Chemistry

Background:

• Efficient computation of electron repulsion integrals is crucial for quantum chemistry.
• Existing methods face challenges with large molecular systems.
• Graphical Processing Units (GPUs) offer potential for accelerating these computations.

Purpose of the Study:

• To develop and evaluate a framework for offloading four-index two-electron repulsion integrals to GPUs.
• To assess the performance of this framework within Restricted Hartree-Fock (RHF) and Effective Fragment Molecular Orbital (EFMO) methods.
• To demonstrate the scalability and efficiency of GPU acceleration for large-scale molecular calculations.

Main Methods:

• Implementation of a GPU-accelerated framework using OpenMP for integral calculations.
• Application of the framework to Fock build in RHF and EFMO methods.
• Benchmark calculations on water clusters and a solvated mesoporous silica nanoparticle system.

Main Results:

• Significant speedups (up to 52×) achieved for RHF calculations on water clusters compared to CPU-based methods.
• High parallel efficiency (up to 94%) observed for RHF on increasing system sizes.
• Excellent linear scalability (up to 4608 V100 GPUs) and 96% parallel efficiency demonstrated for EFMO calculations on a large nanoparticle system.

Conclusions:

• The developed GPU framework effectively accelerates key quantum chemistry computations.
• The approach demonstrates strong scalability and efficiency for both RHF and EFMO methods.
• This GPU acceleration is promising for tackling larger and more complex molecular systems in computational chemistry.