Engineering Supercomputing Platforms for Biomolecular Applications

TL;DR

This study benchmarks representative biomolecular workloads across a diverse set of HPC platforms to understand how hardware choices interact with software stacks in molecular dynamics, quantum chemistry, and cryo-EM data processing. It finds that GPUs are generally the most efficient for MD, but no single platform supports all methods; next-generation AI accelerators like GH200 offer substantial gains yet demand more software tuning and ecosystem maturity. The results highlight significant data storage and software-deployment challenges, and argue for DevOps practices, containerisation, and broader consortium-based support to improve uptime and accessibility. Ultimately, the work informs hardware procurement, software development, and community collaboration strategies to enable scalable, energy-aware biomolecular computing across heterogeneous HPC architectures.

Abstract

A range of computational biology software (GROMACS, AMBER, NAMD, LAMMPS, OpenMM, Psi4 and RELION) was benchmarked on a representative selection of HPC hardware, including AMD EPYC 7742 CPU nodes, NVIDIA V100 and AMD MI250X GPU nodes, and an NVIDIA GH200 testbed. The raw performance, power efficiency and data storage requirements of the software was evaluated for each HPC facility, along with qualitative factors such as the user experience and software environment. It was found that the diversity of methods used within computational biology means that there is no single HPC hardware that can optimally run every type of HPC job, and that diverse hardware is the only way to properly support all methods. New hardware, such as AMD GPUs and Nvidia AI chips, are mostly compatible with existing methods, but are also more labour-intensive to support. GPUs offer the most efficient way to run most computational biology tasks, though some tasks still require CPUs. A fast HPC node running molecular dynamics can produce around 10GB of data per day, however, most facilities and research institutions lack short-term and long-term means to store this data. Finally, as the HPC landscape has become more complex, deploying software and keeping HPC systems online has become more difficult. This situation could be improved through hiring/training in DevOps practices, expanding the consortium model to provide greater support to HPC system administrators, and implementing build frameworks/containerisation/virtualisation tools to allow users to configure their own software environment, rather than relying on centralised software installations.

Figures (55)

• Figure 1: Example rendering of a molecular dynamics system featuring protein and lipid (hEGFR Dimer of 1IVO and 1NQL described in table \ref{['tab:hecbms']}).
• Figure 2: Molecular dynamics performance on a single GPU of JADE2 (higher ns/day is better).
• Figure 3: Molecular dynamics performance on a single node of ARCHER2 (higher ns/day is better).
• Figure 4: MD Performance comparison between a single node of ARCHER2 and single GPU on JADE2 (higher ns/day is better).
• Figure 5: Comparison of MD energy use between a single node of ARCHER2 and a single node of JADE2 (Lower J/ns is more efficient).