My apologies if this topic would fit better in the Virtualization & Cloud section; it seemed like a better fit here.

I have an odd question that frankly I am not sure where best to ask, so I thought I'd give it a shot here.

For my work, being able to run multiple OSs in KVM VMs with dedicated discreet GPUs would be very convenient. I've done some tinkering with a second video card, QEMU, and macOS (don't worry, this is not a macOS-specific question!).

I have gotten to where I can pass the second GPU through to the VM, and can boot the machine with acceleration to a display connected physically to the second GPU.

My question: is it possible to get the accelerated framebuffer back into the host, e.g. inside the QEMU window in the host?

I'm thinking that the answer may be 'no' for a few reasons, but I hope I am wrong.

Any thoughts appreciated.

Isn't that dependent on the physical GPU/hardware?
Note: I came accross this vGPU functions while looking up sr-iov (different cat)
https://wiki.archlinux.org/title/Intel_GVT-g

That's very interesting, thank you. I just so happen to be using a Radeon Pro but it's hard to find information about which models are supported -- some sources say all "Pro" models are. I will definitely be looking into this more.

I think generally the framebuffer, once rendered in the GPU, is not passed back into the host -- it just gets polled over the display cable by the monitor. There are functions in OpenGL, for example, that can copy the buffer back into main memory, but in my experience they are slow.

So that's why I'm wondering, without the use of something like SR-IOV (how does this actually work?) what needs to happen to get that framebuffer back into the host and displayed in the QEMU window.

Note to anybody else reading this -- I am not sure about any of this so take it with a grain of salt.

Tbh, virtualisation isn't my specialty. We have a specialised team at my workplace just doing that at enterprise scale.

Hope this article helps a bit:
https://www.techtarget.com/searchvir...ktop-GPU-cards

Not sure I understand the question correctly, but, wouldn't that break the principle of the Virtual Machine?

It is however possible for the virtual machine to SHARE things with the host, like for example a filesystem. But in practice with QEMU, that is most easily done with plan9 network file system.

Having done alot of experimentation with GPU passthrough (and failed), albeit not with a second GPU, but sharing the same GPU, I did notice that the principle of the GPU in many cases also works in the same way as sharing. Meaning the way you actually get the screen on a non-GPU-passthrough is basically also over a kind of virtual network video stream. Several of the methods of using a display on a KVM/QEMU VM works in that way, and if I don't remember incorrect Looking Glass being one of them.

So, in theory, it should (maybe) be possible to stream back /dev/fb0 to the host in such a way, or something similar.

Maybe ramfb is a key possibility in all this. A tiny piece of technical detail about it here:
https://bugzilla.redhat.com/show_bug.cgi?id=1679680