[Simbelmynë]

Originally Posted by moritzb

I am actually surprised by the massive difference compared to the 7800X3D, that solves in ~90s on 8c with 6000 memory. Is this just the enormous benefit from bigger cache or is there something wrong with my system?

Most cases are too large to fit in the L3 cache. However, a larger L3 cache decreases cache-misses and therefore has the possibility to effectively decrease the RAM latency of your system. CFD greatly benefits from low latency and high bandwidth memory. Not sure if you are comparing apples to apples though in your case as many results reported in this thread do not include latency.

At any rate, cache is king.

[xuegy]

Threadripper 9960X (24 cores), 4x DDR5 5600C28, tested OpenFOAM-v2506 with various compilers:

Ubuntu 24.04 LTS pre-compiled (gcc-13?)
cores MeshTime(s) RunTime(s)
-----------------------------------
1 551.46 367.46
2 383.03 211.9
4 225.46 140.84
8 151.32 92.86
12 108.17 65.21
16 89.61 49.59
24 72.77 38.26

Spack AOCC 5.0.0
cores MeshTime(s) RunTime(s)
-----------------------------------
1 514.21 320.56
2 354.52 194.31
4 223.53 134.3
8 131.54 89.95
12 109.11 64.82
16 88.74 48.48
24 74.71 37.85

Manually installed AOCC 5.0.0 with -zopt(-march=znver5 leads to SIGFPE)
cores MeshTime(s) RunTime(s)
-----------------------------------
1 530.89 335.27
2 376.58 194.06
4 220.48 137.53
8 131.96 90.71
12 111.99 65.99
16 89.38 49.24
24 71.35 38.81

gcc-15.2.0 with -march=znver5
1 491.02 340.73
2 344.24 197.2
4 232.5 135.62
8 135.01 91.41
12 120.38 64.84
16 92.59 49.13
24 70.05 38.07

Conclusion: If you don't have time, just use the pre-compiled one. At 24 cores everything is marginal.

[EmanuelMeisel]

Really nice work. The correlation you highlighted between memory controllers and speedup shows up clearly in the results, and it’s interesting how quickly the 6148 hits its scaling wall compared to the other Xeons. I’ll try your setup on my system and post the runtimes should help expand the dataset and maybe confirm whether this behavior generalizes across more recent CPUs. Thanks for maintaining the wiki page as well; having a central place for updated plots is extremely helpful.

[nikoscham]

System specifications
CPU: AMD Ryzen Threadripper PRO 7995WX (96 cores)
RAM: 512 GB, 4800 MHz
MB: Asus Pro WS WRX90E-SAGE SE
OS: OpenSUSE Leap 16
OpenFOAM-v2412
occDrivAerStatic benchmark with 65M cells (coarse mesh)

Results
cores time (h)
1 577.8
12 57.7
24 46.3
48 43.9
96 45

I was expecting better scaling, particularly given the number of available cores.

[pitter]

Great work putting all this together. The comparisons are really helpful, especially the notes on memory controllers and scaling limits. Thanks for maintaining the wiki page too-it’ll be a solid reference as more results come in.