Multi-GPU MBE(3)-OSV-MP2 for Performant Large-Scale Ab Initio Calculations

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Materials Science

Background:

• Computational acceleration of orbital-invariant local correlation methods on GPUs is challenging due to algorithmic complexities.
• Runtime efficiency is limited by parallelism in orbital localization, wave function solution, and CUDA kernel adaptation.

Purpose of the Study:

• To present a multi-GPU implementation for large-scale third-order many-body expansion orbital-specific virtual MP2 (MBE(3)-OSV-MP2) energy calculations.
• To address GPU parallelization challenges for maximizing utilization and local MP2 computation parallelism.

Main Methods:

• Developed algorithms for Jacobi-Pipek-Mezey localization, randomized OSV (rOSV) generation, and direct MP2 integral regeneration.
• Adapted CUDA kernels for inherently local or sparse operations on GPUs.
• Implemented a multi-GPU approach for MBE(3)-OSV-MP2 energy calculations.

Main Results:

• Achieved effective empirical N^1.9 scaling for MBE(3)-OSV-MP2 energy computation up to (Gly)40/def2-TZVP.
• Demonstrated 84% parallel efficiency without localization using up to 24 GPUs.
• Obtained significant speedups: 40-fold over RI-MP2 and 10-fold over CPU-based MBE(3)-OSV-MP2 for various molecular systems.

Conclusions:

• The GPU-based implementation significantly accelerates local correlation calculations.
• Enables large-scale electronic structure computations on complex macromolecules, such as a 784-atom insulin peptide.
• Opens new avenues for fast GPU-based local correlation calculations in computational chemistry.