GROMACS water box: GPU (Polaris) vs. CPU-only (Frontier)¶
System(s): Polaris & Frontier · Code: GROMACS · Outcome: Success
The ask¶
(This campaign predates verbatim prompt logging — the request below is reconstructed from the campaign's documented design, not a direct quote.)
"Run the same GROMACS water-box benchmark on Polaris (GPU) and Frontier (CPU-only) and compare performance."
What happened¶
Trinity ran a standard water-box molecular dynamics benchmark on both systems. On Polaris, it worked around a filesystem-access restriction on compute nodes and a thread-count conflict with the job scheduler, plus needed an energy-minimization step before the main simulation to avoid constraint-solver errors. On Frontier, it had to bypass a scheduler-integration issue by submitting the job through the login node directly, and had to explicitly reset a thread-count environment variable that the system's compiler modules reset to an unwanted default. This CPU-only Frontier baseline is what a follow-on engineering effort later used as the comparison point when adding GPU support to GROMACS on Frontier.
Results¶
- Polaris (GPU, 7,404 atoms): 178.7 nanoseconds/day.
- Frontier (CPU-only, 4,536 atoms): 93.2 nanoseconds/day.
- Both runs completed cleanly and established real cross-facility performance baselines for this workload.