Quantum Mechanics/Molecular Mechanics Simulations on NVIDIA and AMD Graphics Processing Units.

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• High-Performance Computing

Background:

• Quantum mechanics/molecular mechanics (QM/MM) methods are crucial for simulating complex molecular systems.
• Efficient implementation on graphics processing units (GPUs) is key to accelerating these simulations.
• Optimization for diverse GPU architectures, including AMD, expands accessibility and computational power.

Purpose of the Study:

• To port and optimize the GPU-accelerated QUICK and AMBER QM/MM implementation on AMD GPUs.
• To enhance the performance of the entire Fock matrix build and force calculation within the QUICK code.
• To benchmark the performance on both NVIDIA and AMD GPUs and compare with previous versions.

Main Methods:

• Porting and optimization of QUICK and AMBER QM/MM code for AMD GPUs.
• Inclusion of Fock matrix build and force calculation optimizations (one-electron integrals, two-electron repulsion integrals, exchange-correlation quadrature, linear algebra).
• Performance benchmarking on NVIDIA V100 and AMD MI100 GPUs for standalone HF/DFT and QM/MM molecular dynamics.

Main Results:

• Achieved similar performance for standalone HF/DFT and QM/MM simulations on NVIDIA and AMD GPUs.
• Observed significant speedups for QM/MM molecular dynamics simulations compared to QUICK/AMBER version 21.
• Demonstrated general performance improvements in the QUICK GPU code.

Conclusions:

• The optimized QM/MM implementation on AMD GPUs provides comparable performance to NVIDIA hardware.
• Significant speedups enhance the feasibility of complex simulations using open-source QM/MM software.
• This advancement broadens the scope of scientific inquiry addressable with modern computational resources.