Faster Self-Consistent Field (SCF) Calculations on GPU Clusters.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• High-Performance Computing

Background:

• The self-consistent field (SCF) procedure is a cornerstone of quantum chemistry for determining molecular electronic structure.
• Efficiently parallelizing SCF calculations, particularly on modern hardware like GPUs, is crucial for tackling larger and more complex chemical systems.
• Existing parallel implementations often face limitations in scalability and performance on multi-GPU architectures.

Purpose of the Study:

• To develop and present a novel, high-performance implementation of the SCF procedure optimized for multi-GPU execution.
• To evaluate the computational efficiency and scalability of the new GPU-accelerated SCF algorithm.
• To compare the performance against established CPU-based and GPU-based quantum chemistry codes.

Main Methods:

• The study presents a new SCF algorithm that offloads key computational stages to GPUs.
• Major computational steps include one-electron integrals, electron repulsion integrals (ERI) calculation and digestion, and Fock matrix diagonalization.
• The implementation incorporates SCF acceleration techniques such as DIIS (Direct Inversion in the Iterative Subspace).

Main Results:

• The novel GPU implementation demonstrates remarkable speedups compared to the state-of-the-art parallel GAMESS CPU code.
• Significant performance gains were observed in both single and multi-GPU executions across various molecules and basis sets.
• The code outperforms existing multi-GPU implementations, achieving speedups of 1.2x-3.3x over Terachem using eight V100 GPUs, and up to 28x over QUICK on a single GPU.

Conclusions:

• The presented multi-GPU SCF implementation offers a substantial advancement in computational efficiency for quantum chemistry.
• The algorithm exhibits excellent strong scaling properties up to 8 GPUs and maintains high parallel efficiency up to 18 GPUs.
• This work paves the way for faster and more accessible high-performance quantum chemical calculations on GPU hardware.