Dual AMD EPYC Milan - 3 vs 4 Infinity Fabric connections? Impact?

[DaTruAndi]

Between the Gigabyte board MZ72-HB0

MZ72-HB0 (rev. 3.0/4.0) | Server Motherboard - GIGABYTE U.S.A.

and the ASRock Rack Board ROME2D16-2T
https://www.asrockrack.com/general/productdetail.asp?Model=ROME2D16-2T
I stumbled over that the Rome2D16-2T only seems to have 3 Infinity Fabric connections and the Gigabyte Board seems to have 4.
Is that accurate?

Now the question would be what impact that will have on the performance. I plan to do quite some virtualization, but a lot of Random AI things, at the beginning with one GPU but may add more, and I am equipping it with 1TB Ram.
I also plan to use it as a development workstation to do code compilation, Unity3D game development and so on.

I have somewhat of a preference towards the ASRock Rack board as it seems to be a bit more “workstation like”, but now seeing that the Infinity Fabric of one board may be inferior towards the other, I have second and third thoughts.

Has anybody real experience with the impact of having less Infinity Fabric connections? Not just hypothetical.
Which workloads may suffer?

And, besides of experience, if anybody has seen data looking into this question.

Thanks a lot,

[wendell]

You can do an nps0 mode and that makes both sockets one big numa node. Other than that use case it doesn’t practically matter too much

VMware 6.5 suffered a bit but now is pogy aware and that works fine now

Does what you do need a lot of cross thread comms? Where it might be split across sockets?

[DaTruAndi]

I would use it mostly as workstation with a good mix of code compilation, using unity3d (shader compilation is a thing), Linux through WSL, and there random Python AI projects. Starting out with a single GPU (RTX 4090) but may add more. 7

VMs too but with Windows as the host most likely. At least at the beginning.
I have to figure out what works better in terms of performance. as said windows and WSL plus VMs (hyper-v or free solutions) or running Linux as the main and virtualizing Windows (depending on how well I can pass through the GPU).
In other words completely unpredictable workloads.

Also plan to do video editing.

I don’t think I will have much control over processor affinity. Eg It’s very well possible that the RAM for one process in a heavily multithreaded scenario will be tied to the other processor.
I don’t expect it often to have a VM that only uses RAM of one processor and locks in a certain number of cores. I expect more scenarios where one process needs as much compute as it can get.

The big question for me is currently if I would end up with a real performance hit in some workloads when committing to the Asrock rack board with three vs four on the Gigabyte.

Btw
I followed a certain tuber’s advice and stacked up on some Optane drives for some workloads. Will probably be weeks until I decide about a configuration for those for best performance