Dual EPYC CPU nodes vs Single CPU performance for virtual machines

[harmonyp]

I'm searching for some newer upgraded hardware and am stuck on if I should buy in regards to the performance from 1 CPU Socket vs Dual CPU sockets.

When looking at CPU benchmarks I have noticed that Dual CPUs always have a lower score than having two separate nodes.

For example if you look at cpubenchmark.net results for AMD EPYC 7702 you get 71646 but for Dual EPYC 7702 you only get 91,878 - I would expect closer to 143,292

This is because the CPUs need to communicate with each other if you combine their total CPU power but how does this work with virtual machines? If I created 20 virtual machines each with 12 CPU threads each (using 240/256) would I expect total CPU computation power to be closer to the combined 143,292 result?

 

[LnxBil]

I suspect this is due to different NUMA nodes. In PVE, you can stick VMs to NUMA nodes so that the VM always have local (with respect to the CPU) ram in order to get the best performance. Dual socket system will have more cores and more RAM available, but if that is not the important for you, then two nodes can be better with respect to availablity, but you will have the problem with shared storage and quorum.