Hey, have there been any advances in Navi 22 Host VFIO with Windows guest in the last few years? (Or was it always fine and I’m misinformed?(!))
My setup:
CPU: Ryzen 7 7800X3D
Mobo: Gigabyte X670 Aorus Elite AX Firmware F31
GPU: Radeon RX 6750 XT
RAM: 64GB
OS: Manjaro
It was a long time ago so the details are hazy on what exactly didn’t work, but I remember initially trying to use the CPU’s iGPU for the Host and the RX 6750 for passthrough, then pivoting to trying single GPU with the RX 6750.
I guess my question is, before I start plugging away at this, am I wasting my time? I vaguely recall Navi 22 incompatibility being the blocker, but I accept I may be completely wrong here.
1 Like
I just happened to see this on my Level1Techs summary and thought I’d chime in. I ran this EXACT same GPU for 3 years using Linux as the host and Windows 10 IoT LTSC as the guest and had absolutely no issues with it at all, I was even able to get ReBar to work (and be detected in the AMD driver) by compiling and installing the latest version of QEMU (9.2 at the time).
I am using an Intel CPU (i5-12600K) with an iGPU for the Linux host, I have also blacklisted the AMD drivers so the card isn’t initialized by the Linux Kernel at all. I can start, shutdown and reboot the VM as many times as I want (I usually just leave it running all the time).
The only reason I’m not running this setup anymore is that I was fortunate enough to grab an RX 9070XT at close to MSRP before the prices started going up. It’s also working with zero issues so far using the same setup, I was able to basically just install the new GPU and upgrade the AMD drivers in the guest and everything just worked.
1 Like
Hey thanks for your reply. I will try again!
I can’t do the iGPU I guess as it would involve needing to use the same AMDGPU drivers that the dGPU needs, so blacklisting would be a problem.
If I can get single GPU working reliably I’d be delighted.
Any particular reason for using the IoT version of Win10?
“Any particular reason for using the IoT version of Win 10?”
No Copilot, less telemetry, less crap, no Windows 11 
Good luck, let me know if I can be of any help troubleshooting. I’ve never tried doing single-GPU, it looks like it’s kind of a PITA and prone to issues but I understand for some they don’t have any other choice.
The most challenging VFIO I ever did was getting my Clevo laptop with an RTX 2070 Super to do VFIO passthrough. nVidia really REALLY doesn’t want people doing VFIO with their consumer GPUs. In the end, what got me out of Code 43 driver hell was spoofing the PCI IDs of the GPU and then using a fake ACPI battery so that the VM Guest would treat the system as a laptop.
1 Like
Rather than blacklist the entire amdgpu module, you can load the vfio_pci driver early and make it bind to the dGPU by its vendor+device ID. Then, amdgpu will only be able to grab the iGPU.
See PCI passthrough via OVMF - ArchWiki (→ probably mkinitcpio in your case)
iGPU passthrough tends to be fiddly. I haven’t tested this, but these look like good instructions: AMD 7000 series/Raphael/RDNA2 iGPU passthrough : VFIO