Enabling Multireference Calculations on Multimetallic Systems with Graphic Processing Units.

Area of Science:

• Computational Chemistry
• Materials Science
• High-Performance Computing

Background:

• Efficient modeling of multimetallic systems is crucial for predicting chemical properties and designing new materials.
• Current computational methods can be time-consuming for complex systems.

Purpose of the Study:

• To implement and evaluate a multi-GPU accelerated localized active-space self-consistent field (LASSCF) method.
• To demonstrate the performance gains and scalability of the parallelized LASSCF calculations.

Main Methods:

• Utilized multiple graphics processing units (GPUs) for parallel computation.
• Employed density fitting to reduce memory requirements.
• Performed LASSCF calculations on multimetallic catalyst systems.

Main Results:

• Achieved 5-10x speedup in total runtime for LASSCF calculations.
• Demonstrated efficient utilization of multi-GPU compute nodes.
• Observed comparable performance on NVIDIA A100 and Intel Max Series GPUs, indicating performance portability.

Conclusions:

• The multi-GPU implementation significantly accelerates LASSCF calculations for multimetallic systems.
• This approach facilitates faster material property prediction and design.
• The method shows promise for exascale computing environments.