The issue isnt EXACTLY the baseline virtual cpu model (although this comes to play too) but rather the presence/absence of specific feature flags and/or hw vulnerability mitigations.
The X86-vX models essentially are presets for flags, and attempt to mimic a specific "age" of underlying hardware. In context of guests, Linux guests are much more forgiving then Windows, as Windows iterations have more and more hardware dependencies. For example, x86-v2-aes works fine for windows 2016 but terrible for Windows 11/2025 as they make heavy use of modern cpu features (which is why they dont support hardware prior to 8-9th gen intel.)
In summary- the virtual cpu model you choose depends on:
- The guest OS
- the host cpu- virtual cpu cannot report capabilities not present
- the Cluster makeup (it applicable)
- Security stance (for homelab you might not care about process breakout potential)
In the case of Epyc processors there is another potential issue. PVE doesnt have good control (or at least didnt) of CCD NUMA placement. What that means is that your observations on L1/L2/L3 cache and DRAM performance can be skewed and may even be different between reboots; search the forums for specific discussions on fixing core pinning to mitigate.
