Turbulent pipe flow across two facilities: Polaris vs. Aurora¶
System(s): Polaris & Aurora · Code: NekRS · 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 turbulent pipe-flow simulation (Reynolds number 5200) on both Polaris and Aurora and compare performance."
What happened¶
Trinity ran an identical spectral-element CFD simulation of turbulent flow through a pipe on both Polaris (NVIDIA GPUs) and Aurora (Intel GPUs), a genuine cross-facility comparison. The Polaris run worked on the first try. The Aurora run took five iterations to get right — the compute backend name had to be set correctly for Intel's compiler stack, and a stale build cache from a previous run had to be cleared before Aurora would pick up the right MPI configuration.
Results¶
- Polaris: 26.73 simulation steps/second.
- Aurora: 23.3 steps/second — about 87% of Polaris's throughput on equivalent hardware generations, a useful cross-facility performance data point.
- Both runs produced numerically stable turbulent flow advancement, confirming the same physics setup works correctly on both vendors' hardware.