I would like to accelerate a Windows VM without passing a whole GPU through (my dedicated A310 is used for docker containers and passing my CPUs iGPU through has not worked for me in the past as detailed here).
Is there any way to set the GPU type of the VM to virtio (similarly as described here) so the VM gets accelerated via a GPU for normal desktop use without passing one through?
Just as an update: I tried enabling my iGPU again and passing it through to the Windows VM.
I had hope since my last try was still under 24.04 and with a Linux VM.
I could actually isolate the GPU without problems and there also was no warning when I added it via “Edit” in the settings.
The VM started but did not recognize the GPU. When I searched for new devices in device manager it hung. After I waited a while and killed the VM (which took way longer than expected) the PCIe devices that were passed through were all shown as “Not available” and when I tried booting the VM again I only got an “Unknown PCI header type ‘127’” error message.
This also appeared after I removed all PCIe devices but the actual GPU from the VM.
There were all sorts of familiar error messages in syslog:
Jan 6 20:46:09 truenas1 kernel: vfio-pci 0000:0f:00.6: Unable to change power state from D3cold to D0, device inaccessible
Jan 6 20:46:09 truenas1 kernel: pcieport 0000:00:08.1: broken device, retraining non-functional downstream link at 2.5GT/s
Jan 6 20:46:09 truenas1 kernel: pcieport 0000:00:08.1: retraining failed
Jan 6 20:46:09 truenas1 kernel: pcieport 0000:00:08.1: broken device, retraining non-functional downstream link at 2.5GT/s
Jan 6 20:46:09 truenas1 kernel: pcieport 0000:00:08.1: retraining failed
Jan 6 20:46:09 truenas1 kernel: vfio-pci 0000:0f:00.0: not ready 1023ms after bus reset; waiting
Jan 6 20:46:09 truenas1 kernel: vfio-pci 0000:0f:00.0: not ready 2047ms after bus reset; waiting
I should also note that the machine is type q35 with TPM as it is Windows 11.
So all in all: still no luck with GPU acceleration even when trying to pass a whole GPU through.