What this does not show
Where the measurements stop, where brae is not ahead, and what is not yet understood. Read these beside the figures on the other pages.
Single runs, no repeats
One OpenFOAM run and one brae run per setting. Timings of the small tutorials moved by up to 2× between two identical sweeps.
No speed claim below about 100,000 cells
Against 26 cores brae is ahead on 19 of 40 tutorials per time step (median 0.98×): ahead on the large meshes and behind on the small ones. The dam break is the speed claim.
The default path is not the 1e-08 path
By default brae replaces the case's pressure solver with its own AMG-preconditioned PCG and sweeps some fields in another order. At each tutorial's own loose solver tolerance that path differs from serial OpenFOAM by a median 1.3e-02 in velocity after 30 steps (alpha 5.0e-04); OpenFOAM differs from itself the same way between serial and 26 ranks. The 1e-08 figure is brae running the case's own solvers.
Thirteen tutorials sit above 1e-06
Thirteen of the 42 are above 1e-06 in velocity even with the case's own solvers (worst 1.1e-03); ten are moving-mesh cases. The cause is not established.
One known defect, not yet fixed
On the motorBike tutorial brae's end fields depend slightly on how often it writes (up to 2.7e-07 relative, on omega).
The impact peak is not converged
The impact pressure peak is not converged in the time step in either code, and most sensor quantities oscillate with the mesh rather than converge (16 of 64 monotone for brae, 20 for OpenFOAM).
After the impact the splashes differ
The two runs' splashes no longer coincide cell for cell: the overlap of the water regions is 0.9999 before the impact, 0.86 at its lowest and about 0.95 at the end. Gauges, volume, energy and momentum stay together.
OpenFOAM's reservoir gauge H4
In its parallel runs OpenFOAM's gauge H4 starts at 0.60 m (0.4 M cells) and 0.5675 m (4.89 M) instead of 0.55 m. That is its gauge, not the flow.
The same-solver check
The same-solver check through the impact was made at 0.4 M cells; at 4.89 M cells it covers the first eight time steps.
The converged OpenFOAM runs
The converged OpenFOAM runs at 1.37 M and 4.89 M cells use OpenFOAM's PCG with a GAMG preconditioner; at 0.4 M cells OpenFOAM's two solvers were both converged and agree to 0.001 Pa.
One sensor, one run per setting
Only sensor P1's first impact was examined this way, with one run per setting.
A second machine
These checks ran on a second machine with OpenFOAM v2412 at another patch level than the timed runs.
One GPU
brae runs on one GPU here. Multi-GPU and parallel runs are not part of this.