High-Performance, High-Angular-Momentum J Engine on Graphics Processing Units.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Materials Science

Background:

• Evaluating electron repulsion integrals (ERIs) with high-angular-momentum Gaussian functions is computationally intensive for GPUs.
• Existing methods face register pressure and memory bottlenecks due to intermediate generation.
• Efficient computation of Coulomb matrices (J) is crucial for electronic structure calculations.

Purpose of the Study:

• To present a high-performance Coulomb-matrix (J) engine optimized for GPU execution.
• To address computational challenges in evaluating high-angular-momentum ERIs on GPUs.
• To enhance the efficiency of electronic structure calculations.

Main Methods:

• Developed a GPU-optimized McMurchie-Davidson recurrence algorithm.
• Implemented a tailored integral batching scheme to minimize intermediates and redundant computations.
• Partitioned high-angular-momentum ERI classes into sub-batches to shift kernels from memory-bound to compute-bound regimes.

Main Results:

• Achieved individual kernel speedups of up to 9×.
• Improved overall J-matrix formation performance by up to 64%.
• Demonstrated performance on polyglycine chains, water clusters, and boron nitride crystals using cc-pVQZ basis set on NVIDIA A100 GPU.

Conclusions:

• The proposed GPU-optimized engine significantly enhances computational throughput for high-angular-momentum ERIs.
• The approach effectively reduces time to solution for electronic structure calculations.
• The method is scalable and applicable to various chemical systems and basis sets.