NVIDIA Now Allows GeForce GPU Pass-Through For Windows VMs On Linux

[mppix]

You don't actually need to have 2 GPUs, with some setups you can do it with one GPU. Problem is you can't simply share one GPU between 2 or more operating systems. There is GPU virtualization but it's basically limited to professional cards for AMD and Nvidia. As far I know only Intel supports this on consumer integrated graphics but these cards are not very good match for gaming. So if you want to do it with customer AMD or Nvidia card you need to unbind it from your host before VM start and bind again when VM stops. Some cards were (or even still are) problematic with such scenario (check AMD reset bug or patching Nvidia ROM). There is also main disadvantage compared to dual GPU setup - you can't easily use both operating systems at the same time. When VM starts your host loses control over your GPU and only VM will be able to display things. It's something like dual booting without rebooting.

Single GPU makes only sense if the host/hypervisor is headless today. For the rest we need to wait for GPUs with proper virtualization support such as SR-IOV

[oleid]

oleid , MadWatch , until now, needed to mask your VM such that the Nvidia driver in the Windows VM would not detect a VM. This creates all sorts of compromises.
This is pretty much why every GPU passthrough tutorial starts with 'if you have Nvidia, see also the other 128 steps.. and hope that certain stars are aligned'.

Ah, the joys of non-free drivers xD

[Serafean]

Read as:
nvidia removed arbitrary (maybe legal...) limitation from their windows drivers.

The more interesting SR-IOV is explicitly not supported.

[plantroon]

So now I can use what I've been using for the past couple of years without hacks. The hack being one line in libvirt xml. No big deal...

[oleid]

Just wondering : does anybody know how virtio-gpu is working on windows these days?

[Snaipersky]

So now I can use what I've been using for the past couple of years without hacks. The hack being one line in libvirt xml. No big deal...

The amount of work required can vary based on driver update vs install, and whether windows has tried to install drivers already. I've had some that required only one change, and others that required every paravirt feature masked.

[leipero]

Just wondering : does anybody know how virtio-gpu is working on windows these days?

It doesn't AFAIK there's no drivers for Windows. QXL works, but aside from basic acceleration, you can't expect much from it.

[rbalint]

It would be nice to have input lag benchmarks comparing native and guest Windows with GPU passthrough.

[bachchain]

So what are they replacing it with? Nvidia loves market segmentation, and I doubt they would just remove one of their artificial barriers without putting a new one up.

[Schmellow]

So what are they replacing it with? Nvidia loves market segmentation, and I doubt they would just remove one of their artificial barriers without putting a new one up.

Probably nothing. They just removed "error 43" which everyone worked around and they didn't enforce it much. The only question people had if nvidia starts enforcing it some day or not. Now we know, i guess.

Something on SR-IOV would be hella nice, but alas only Intel has something like this on their consumer gpu's (if you can call them that) with their GVT-G