Acceleration of a QM/MM-QMC simulation using GPU.

Area of Science:

• Computational chemistry
• High-performance computing
• Quantum mechanics

Background:

• Ab initio molecular Quantum Monte Carlo (QMC) calculations are computationally intensive.
• Accelerating these calculations is crucial for advancing scientific discovery.

Purpose of the Study:

• To accelerate ab initio molecular QMC calculations using Graphics Processing Units (GPUs).
• To evaluate the performance gains and accuracy of GPU-accelerated QMC.

Main Methods:

• Implemented a hybrid approach, replacing the bottleneck of QMC calculations with CUDA subroutines executed on GPUs.
• Compared performance of a (single core CPU + GPU) setup against a (single core CPU with double precision) setup.
• Assessed energy deviation in single-precision treatments on GPU.

Main Results:

• Achieved 23.6x and 11.0x speedups in single and double precision, respectively, using GPU acceleration.
• Single-precision treatment on GPU resulted in energy deviations within the required accuracy (∼10⁻⁵ hartree).
• Demonstrated linear performance scaling in a hybrid Message Passing Interface (MPI) cluster of GPU-accelerated nodes.

Conclusions:

• GPU acceleration is an effective strategy for speeding up ab initio molecular QMC calculations.
• The hybrid approach offers significant performance improvements without compromising essential accuracy.
• Scalable performance of GPU-accelerated nodes enables larger and more complex simulations.