Technical Papers Archive

Unleashing Shared Memory Potential for NEP Machine-Learning Potential in Molecular Dynamics on GPUs


Authors: Zeyu Song and Zikai Zhou (Tsinghua University); Lin Gan (Tsinghua University; Hetao Institute of Mathematics and Interdisciplinary Sciences, Shenzhen); Xiaohui Duan (Shandong University); and Guangwen Yang (Tsinghua University)

Abstract: Machine-learning potentials (MLPs) achieve near first-principles accuracy in molecular dynamics while retaining high efficiency. The neuroevolution potential (NEP) uses compact networks to deliver substantially higher GPU throughput than other MLPs. However, existing NEP implementations suffer from high per-thread memory demands that limit effective use of low-latency shared memory.

We present a shared-memory-centric optimization framework for NEP on GPUs. By introducing subwarp partitioning and two-dimensional neighbor reordering, it enhances inter-thread data sharing and reduces memory requirements, allowing most accesses to be served from shared memory. We further incorporate shared-memory-based matricized computation, buffering strategies, along with a series of GPU performance optimizations, and quantify shared memory usage to facilitate analysis.

Our framework achieves up to 18.09x speedup over the baseline GPU implementation and an average of 2.60x speedup over GPUMD across five representative material systems on NVIDIA H100 GPUs.



Presentation: file

Back to Technical Papers Archive Listing