If you don’t attaching it to any qube, it will be sitting in dom0, exposing it’s outputs (which are actually inputs as well) directly to the ‘environment’. Sure, to exploit such need physical access…
Hardware supported PCI Passtrough (IOMMU and VT-d) is the best isolation technique known today. Means there are no better way at the moment.
And if we forget gaming, then this is the reason for the sys-gui-gpu in the first place: to further reduce the attack surface, by attaching the GPU to a dedicated qube.
Compered to this, just attaching the GPU to a single (gaming/LLM/whatever) qube is much more simple operation with much less complexity and less attack surface as a result.
The real issue if you want to ‘share’ that GPU. And that’s what Qubes OS does not supports - for a very good reason.
well, the ‘–option permissive=True’ is actually weakening the hardware supported isolation, and it might be allowing DMA attacks - or more realistically: system crashes.
(all depends on your CPU, motherboard and it’s BIOS/UEFI)
The ‘–option no-strict-reset=True’ is not as problematic - as long as you have only one single qube where you GPU is attached.
So if your gaming qube is already hacked, and/or running malicious code, then it can be an attack surface compared to the case where you are not using your GPU at all.
Every system security can by tightened up to the point when it’s becoming unusable.