Distributed Multi-GPU Ab Initio Density Matrix Renormalization Group Algorithm with Applications to the P-Cluster of...

Area of Science:

• Quantum Chemistry
• Computational Chemistry
• Materials Science

Background:

• Polynuclear transition-metal compounds, like iron-sulfur clusters in nitrogenase, present significant computational challenges due to degenerate d/f orbitals.
• Existing quantum chemistry methods struggle with the complexity of these systems.

Purpose of the Study:

• To develop a novel computational approach for accurately simulating complex transition-metal compounds.
• To overcome the limitations of current quantum chemistry methods for systems with degenerate d/f orbitals.

Main Methods:

• Introduction of the first distributed multi-graphics processing unit (GPU) *ab initio* density matrix renormalization group (DMRG) algorithm.
• Implementation on modern high-performance computing (HPC) infrastructures.
• Parallelization of operator-wave function multiplication using operator parallelism and batched GPU contractions.

Main Results:

• Achieved an unprecedented bond dimension (D = 14,000) for an active space model of the P-cluster (114 electrons in 73 active orbitals).
• Utilized 48 NVIDIA A100 80 GB SXM GPUs for the calculation.
• Exceeded previous DMRG bond dimensions for this system by nearly threefold compared to CPU-only calculations.

Conclusions:

• The new distributed multi-GPU DMRG algorithm significantly enhances the capability of quantum chemistry methods for complex systems.
• This advancement allows for more accurate and larger-scale simulations of challenging polynuclear transition-metal compounds.
• The developed algorithm is suitable for modern HPC environments, paving the way for future discoveries.