Harnessing the Power of Multi-GPU Acceleration into the Quantum Interaction Computational Kernel Program.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• High-Performance Computing

Background:

• Accurate electronic structure calculations are crucial for understanding molecular properties.
• Scaling quantum chemistry methods to large systems requires efficient parallelization strategies.
• The QUantum Interaction Computational Kernel (QUICK) program provides a platform for quantum chemistry simulations.

Purpose of the Study:

• To develop and implement a multi-GPU accelerated ab initio Hartree-Fock/density functional theory (HF/DFT) method within the QUICK program.
• To optimize load balancing algorithms for electron repulsion integrals and exchange-correlation quadrature across multiple GPUs.
• To evaluate the parallel efficiency and scalability of the new implementation on various GPU architectures.

Main Methods:

• Integration of multi-GPU capabilities into the QUICK program for HF/DFT calculations.
• Development of load balancing algorithms for electron repulsion integrals and exchange-correlation quadrature.
• Benchmarking the implementation on NVIDIA V100-SXM2, A100, P100, and K80 GPUs using medium to large molecular systems.

Main Results:

• Excellent load balancing and high parallel efficiency achieved across multiple GPU nodes.
• Parallel efficiencies exceeding 82% for Kohn-Sham matrix formation and 90% for nuclear gradients on four V100 GPUs.
• Demonstrated parallel efficiencies greater than 68% on NVIDIA A100, P100, and K80 platforms.

Conclusions:

• The new multi-GPU implementation significantly enhances the performance of ab initio HF/DFT calculations within QUICK.
• The efficient parallelization strategy enables large-scale electronic structure calculations on modern GPU hardware.
• This work paves the way for more extensive computational studies of complex molecular systems.