Time to solution

Whole runs of the dam break to 6 s of flow: brae on 1 × NVIDIA H100 PCIe 80 GB and 1 CPU core, OpenFOAM v2412 on 26 cores of an Intel Xeon Platinum 8480+, both on the same rented machine.

Time to the same simulated time

398,720 cells

5.1× sooner

OpenFOAM, 26 cores

366 s

brae, 1 GPU

72 s

1,369,380 cells

7.3× sooner

OpenFOAM, 26 cores

2,226 s

brae, 1 GPU

304 s

4,894,216 cells

11.9× sooner

OpenFOAM, 26 cores

5 h 37 min

brae, 1 GPU

28 min

Wall clock, start-up and writes included; one run each. Each pair of bars is drawn against that mesh's OpenFOAM time, so compare within a pair, not across them. brae's AMG hierarchy is read from its cache (built once per mesh) and OpenFOAM's decomposePar is not counted.

The speed table

With $T$ the wall clock of a whole run to the same simulated time, the speed-up is

S=\frac{T_{\mathrm{OF}}}{T_{\mathrm{brae}}}

(1)

Mesh	Simulated	brae, 1 × H100	OpenFOAM, 26 cores of a Xeon Platinum 8480+	Speed-up
398,720 cells	6 s	72 s	366 s	5.1×
1,369,380 cells	6 s	304 s	2,226 s	7.3×
4,894,216 cells	6 s	1,700 s (28 min)	20,196 s (5 h 37 min)	11.9×

At 4,894,216 cells a time step takes 207 ms against 2,500 ms; brae took 8,202 steps and OpenFOAM 8,079. The opening 0.6 s at that size, run separately: 160 s against 1,762 s, 11.0×.

Cost of the same result

From the measured times, in the unit a reader can price for themselves: core-hours against GPU-hours. More cores finish sooner and use about the same core-hours. With $n=26$ cores and $c_{\mathrm{GPU}}$ the hourly price of the GPU instance, the cores one GPU stands in for and the core-hour price below which OpenFOAM is the cheaper run are

n_{\mathrm{eq}}=\frac{n\,T_{\mathrm{OF}}}{T_{\mathrm{brae}}},\qquad c^{*}=\frac{c_{\mathrm{GPU}}\,T_{\mathrm{brae}}}{n\,T_{\mathrm{OF}}}=\frac{c_{\mathrm{GPU}}}{n_{\mathrm{eq}}}

(2)

Mesh	OpenFOAM core-hours	brae GPU-hours	One H100 does the work of	OpenFOAM is cheaper below
398,720 cells	2.65	0.020	132 of these cores	2.5 cents a core-hour
1,369,380 cells	16.1	0.084	190 cores	1.7 cents a core-hour
4,894,216 cells	145.9	0.472	309 cores	1.1 cents a core-hour

The break-even is at 3.29 USD an hour for the H100 instance: Lambda Cloud, on demand, on 2026-10-11. At that rate the three brae runs cost 0.07, 0.28 and 1.55 USD. The same instance running OpenFOAM on its 26 cores costs 0.33, 2.03 and 18.46 USD, which is true but weak alone: nobody rents an H100 to run OpenFOAM on its CPU cores.

This is one provider's on-demand rate on one date. The cores are these cores, 26 of a Xeon Platinum 8480+; another CPU may be faster or slower per core. At 0.4 M cells the cost is close to a wash against cheap (spot or on-premises) cores.