HierCut: Enabling 16-bit Format Mixed Precision for Molecular Dynamics through Hierarchical Cutoff

Abstract

Mixed-precision methods offer the potential to achieve better performance while maintaining accuracy comparable to that of high-precision formats. However, the adoption of mixed precision—particularly with 16-bit formats—in scientific computing remains limited due to precision truncation.

To address this challenge, we propose HierCut, a mixedprecision strategy that leverages cutoff schemes in molecular dynamics (KOKKOS package of LAMMPS) by assigning different precision levels to particles across distinct cutoff layers. We further introduce techniques to improve the performance and accuracy of simulations using 16-bit numerical formats, and develop error analysis methods to guide the configuration of mixed-precision ratios. Our scheme achieves accuracy comparable to high-precision results, delivering a speedup of up to 3.75x over the original FP64 implementation of LAMMPS and up to 1.40x over the optimized FP32 kernels, based on NVIDIA A100 GPUs. Our implementation can be effectively scaled to boost large-scale multi-GPU simulations.