A GPU-Accelerated Fast Multipole Method for GROMACS: Performance and Accuracy.

Area of Science:

• Computational chemistry
• Molecular dynamics simulations
• High-performance computing

Background:

• Calculating long-range electrostatic interactions is computationally intensive in molecular dynamics.
• Particle Mesh Ewald (PME) is the standard method but scales poorly in large parallel simulations.
• Exascale supercomputing demands methods with improved scaling properties.

Purpose of the Study:

• To implement and integrate a GPU-accelerated Fast Multipole Method (FMM) into GROMACS as an alternative to PME.
• To compare the accuracy and performance of FMM and PME for molecular dynamics simulations.
• To evaluate FMM's potential for exascale biomolecular simulations.

Main Methods:

• Implemented a performance-optimized GPU FMM and integrated it into the GROMACS package.
• Assessed the accuracy of FMM and PME using various input parameters.
• Compared the performance of GROMACS with FMM against PME on benchmark systems, including mixed-precision settings.

Main Results:

• FMM with a multipole order of 8 achieves accuracy comparable to standard PME.
• FMM does not increase energy drift in mixed-precision simulations for multipole orders of 8 or higher.
• FMM is slower than PME for dense, ≈50,000 atom systems but significantly outperforms PME for large, inhomogeneous systems like aerosols on a single node.

Conclusions:

• FMM is a viable and accurate alternative to PME for electrostatic interactions in molecular dynamics.
• FMM demonstrates superior scaling properties, making it suitable for future exascale simulations, particularly for specific system types.
• Further development of FMM in GROMACS can enhance the scalability of large-scale biomolecular simulations.