MINT32: A Minimum-Image INT32 Coordinate Representation for Fast and Accurate Molecular Dynamics on GPUs.

Area of Science:

• Computational Chemistry
• Molecular Dynamics Simulations
• GPU Computing

Background:

• GPU-accelerated molecular dynamics (MD) simulations often face a trade-off between computational performance and numerical precision.
• Existing mixed-precision methods (e.g., AMBER SPFP, OpenMM) use single precision (FP32) for coordinates, leading to quantization errors that introduce artificial heat and degrade simulation stability.

Purpose of the Study:

• To introduce and evaluate MINT32, a novel coordinate representation for GPU-based MD simulations.
• To overcome the limitations of FP32 coordinates and achieve higher numerical precision and stability in MD simulations.

Main Methods:

• Developed MINT32, mapping simulation box coordinates onto a 32-bit integer grid with high spatial resolution (~0.01 fm).
• Implemented MINT32 in a modified AMBER simulation package as a testbed for evaluating coordinate representations.
• Conducted benchmark simulations on systems ranging from 12K to 91K atoms, including PME water and protein systems.

Main Results:

• MINT32 reduced energy drift in microcanonical (NVE) simulations by 5-10× compared to conventional methods, achieving double-precision stability.
• Coordinate precision was identified as the dominant factor for simulation stability, outperforming conventional mixed-precision models.
• MINT32 integer arithmetic showed negligible overhead (0-5%) on consumer GPUs, with potential for performance gains in optimized implementations.

Conclusions:

• MINT32 offers a path to highly stable and accurate GPU-accelerated MD simulations without sacrificing performance.
• The findings provide a foundation for next-generation MD software, emphasizing the critical role of coordinate representation.
• MINT32 enables tighter SHAKE tolerances (10⁻⁷ Å) for production-length simulations.