A Mixed-Precision Approach to a Preconditioned Eigensolver for Efficient Density Functional Calculations on...

Area of Science:

• Computational Chemistry
• Materials Science
• High-Performance Computing

Background:

• Graphics Processing Units (GPUs) have advanced computational chemistry.
• AI-focused GPUs are inefficient with double-precision (FP64) quantum chemistry operations.
• Limited GPU memory requires algorithmic adaptations for Density Functional Theory (DFT).

Purpose of the Study:

• To develop a mixed precision strategy for iterative matrix diagonalization in real-space DFT.
• To maintain numerical accuracy while improving computational efficiency on GPUs.
• To enable larger-scale electronic structure simulations and broaden GPU applicability.

Main Methods:

• Implemented a systematic mixed precision strategy for matrix diagonalization.
• Utilized single-precision (FP32) and Brain Floating Point (BF16) arithmetic.
• Developed and validated a mixed precision eigensolver across diverse material systems.

Main Results:

• Achieved up to 10x speedup in diagonalization compared to FP64.
• Extended feasible system size by approximately 50%.
• Demonstrated comparable performance of AI-focused GPUs to HPC-focused GPUs.

Conclusions:

• Mixed precision strategy preserves numerical accuracy in DFT calculations.
• AI-focused GPUs can be effectively utilized for large-scale electronic structure simulations.
• The developed method enhances accessibility to advanced computational chemistry tools.