An Efficient RI-MP2 Algorithm for Distributed Many-GPU Architectures.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• High-Performance Computing

Background:

• Second-order Møller-Plesset perturbation theory (MP2) is crucial for accurate molecular energy calculations beyond the Hartree-Fock approximation.
• Current RI-MP2 methods face computational cost and scalability challenges on modern supercomputing architectures.
• Efficient algorithms are needed to apply RI-MP2 to larger, more complex molecular systems.

Purpose of the Study:

• To present the first distributed-memory many-GPU algorithm for RI-MP2 calculations.
• To optimize RI-MP2 computations for hundreds of GPU accelerators.
• To enable efficient and scalable quantum chemistry simulations on modern hardware.

Main Methods:

• Developed a novel distributed memory algorithm for forming RI-MP2 intermediate tensors with minimal communication.
• Implemented a distributed memory algorithm for the energy reduction step, sustaining high performance on GPU clusters.
• Utilized hundreds of GPU accelerators for all computational steps.

Main Results:

• Achieved near-peak performance on GPU-based supercomputers.
• Outperformed state-of-the-art quantum chemistry software by over 3.5 times in speed.
• Reduced computational power consumption by 8-fold.
• Demonstrated 11.8 PFLOP/s performance on the Perlmutter supercomputer for a large water cluster.

Conclusions:

• The novel many-GPU RI-MP2 algorithm offers significant time-to-solution and power consumption benefits.
• This work paves the way for applying advanced quantum chemistry methods to large molecules on GPU-accelerated systems.
• Leveraging modern GPU computing environments is crucial for advancing computational chemistry.