Parallel Implementation of the Density Matrix Renormalization Group Method Achieving a Quarter petaFLOPS Performance...

Area of Science:

• Quantum Chemistry
• Computational Physics
• High-Performance Computing

Background:

• Accurate electronic structure calculations are crucial for understanding enzyme mechanisms.
• The Density Matrix Renormalization Group (DMRG) is a powerful method for quantum chemistry.
• Scaling DMRG to large systems on modern hardware remains a challenge.

Purpose of the Study:

• To report performance results of a spin-adapted DMRG implementation on NVIDIA DGX-H100 architectures.
• To evaluate the efficiency of tensor network algorithms on hybrid CPU-multi-GPU systems.
• To assess the feasibility of tackling challenging quantum chemistry problems with advanced hardware.

Main Methods:

• A single-node hybrid CPU-multi-GPU implementation of the spin-adapted DMRG method.
• Performance evaluation on NVIDIA DGX-H100 architectures.
• Calculations performed for active sites of FeMoco and cytochrome P450 enzymes.

Main Results:

• Achieved 246 teraFLOPS of sustained performance.
• Demonstrated a 2.5x performance improvement over DGX-A100 architectures.
• Showcased an 80x acceleration compared to a 128-core CPU implementation.

Conclusions:

• Tensor network algorithms can efficiently utilize high-performance multi-GPU hardware.
• The combination of tensor networks and GPU accelerators enables solving complex quantum chemistry problems.
• This work paves the way for advancements in computational chemistry and related fields.