[Simbelmynë]

Hey,

So I have tested some native OpenFOAM with a base model Mac Studio (32 GB RAM, 8 performance cores)

Since the old benchmark thread is really old I opted to go with just running the motorbike tutorial case. Since the case creates a rather small mesh I beefed it up by changing the blockMesh parameters to (200 80 80) instead of (20 8 8), this produced a mesh of about 4.4 million cells. Time is measured including meshing and simpleFoam ("time ./Allrun")

OF2206 M1 Max, 6 cores: 30 minutes

OFv10 5800X3D, 6 cores: 47 minutes
OFv10 5800X3D, 8 cores: 45 minutes
Only running the memory @3200 MT/s CL14 since I am lazy.

Could not manage to run properly on all eight performance cores since I am completely new to how macOS handles priority, running on 8 cores sank the performance to 74 minutes . Perhaps there is a bit more performance to be had. Also, the meshing part does not utilise the high memory bandwidth of the M1 Max, so a straight up simpleFoam run would probably be even more favourable.

The CPU temperature never increased beyond 37 oC during the simulation so the cooling solution and power draw is very impressive.

Nothing new perhaps, but it gives some insight how the CFD performance can be if using native arm compiled code.

[Simbelmynë]

Some more testing:

I changed to OF2206 on the Linux machine so that both run the same version of OF. I also did a fresh install of Ubuntu 22.04 for good measure.

Only testing the simpleFoam solver (so excluding all meshing etc.) I get 1320 and 1118 seconds for 6 and 8 cores respectively on the M1 Max.

The same case runs at 2355 seconds with 6 cores on the 5800X3D.

So the M1 Max has roughly twice the performance of the 5800X3D when it comes to the CFD solver. It is also fun to point out that the fans are really not audible during the simulation, just a minor hum if you pay attention.

** Edit **

So I found that there were some inconsistent behaviour in the results. This is likely connected to problems with memory initialisation, using MPI on macOS. https://github.com/open-mpi/ompi/issues/7393

The "solution" was to do the following export:

export TMPDIR=/tmp

I also converted the benchmark case to OpenFOAM.com from OpenFOAM.org, and ran it on 8 cores, which resulted in a solution time of 84 seconds. Energy per simulation (or iteration) seems like an interesting metric so I will try to measure that at the wall if I have time.