High-Performance Multi-GPU Analytic RI-MP2 Energy Gradients.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• High-Performance Computing

Background:

• Second-order Møller-Plesset perturbation theory (MP2) is crucial for accurate electronic structure calculations.
• Analytic energy gradients are essential for geometry optimization and molecular dynamics.
• Existing methods face computational bottlenecks, especially for large systems.

Purpose of the Study:

• To develop a high-performance algorithm for RI-MP2 analytic energy gradients.
• To leverage multiple graphical processing units (GPUs) for significant computational speedup.
• To enable accurate gradient calculations for larger molecular systems.

Main Methods:

• Implementation of a novel GPU-accelerated algorithm within the EXtreme Scale Electronic Structure System (EXESS) software.
• Utilization of GPUs for integral generation, tensor formation, Z-vector equation solution, and gradient accumulation.
• Integration with a molecular fragmentation framework to reduce computational scaling.

Main Results:

• Achieved over 80% of theoretical peak floating-point performance on nodes with 8 A100 GPUs.
• Demonstrated up to a 95-fold speedup compared to established CPU-based methods (Q-Chem, ORCA).
• Reduced RI-MP2 gradient calculation scaling from quintic to subquadratic using molecular fragmentation.

Conclusions:

• The developed GPU-accelerated RI-MP2 gradient algorithm offers substantial performance gains.
• The integration with molecular fragmentation provides significant computational savings with high accuracy.
• This approach enables efficient and accurate electronic structure calculations for larger and more complex molecules.