[SOLVED] Is there any fix for Proxmox crashing when using GPU passthru with AMD

W

wesnewb

New Member
Sep 13, 2025
6
0
1
Florida
I used to use Proxmox GPU passthru on my old computer with a Quadro M4000, and it worked without a hitch. I believe this was on 8.2

Now, after building a new system (All AMD, AM5, 64GB DDR5, 7800XT, etc.), I cannot for the life of me get passthru working fully with the AMD card. I spent a few days troubleshooting just getting it to power on and display the VM. I ended up having to download and use a rom file for the card to get it to work. I thought that was it, but when restarting a VM with passthru (I really only have a Windows and Linux one), It would always crash all of Proxmox. I checked to see if the card was in a pci group with multiple things, but it wasn't. I also was going to use vendor-reset, until I realized it doesn't support my card, or any new cards at that.

Is this a dead end for me or is there a possible fix? Thanks
 
I am using slot 1

Motherboard: MSI MAG B650 Tomahawk WiFi
CPU platform: AMD Ryzen 7 7700X (Raphael / Granite Ridge, AM5 socket)

I don't have Proxmox on the computer at the moment as i had to install Debian to get something done. I feel like I have checked the IOMMU groups before and they seemed fine, but i don't remember
 
Last edited:
You can run for d in /sys/kernel/iommu_groups/*/devices/*; do n=${d#*/iommu_groups/*}; n=${n%%/*}; printf 'IOMMU group %s ' "$n"; lspci -nns "${d##*/}"; done on any Linux. But remember that Debian uses a different kernel than Proxmox (which is based on the Ubuntu kernel).
Maybe update the motherboard BIOS if the x16 PCIe slot connected to the CPU causes problems on the Proxmox host with PCI(e) passthrough. Also make sure to early bind the GPU to vfio-pci to work around unbind and reset issues.

There are also many, many threads about PCI(e) passthrough on this forum with tips and work-arounds and even more on the internet as a whole.
 
leesteken said:
You can run for d in /sys/kernel/iommu_groups/*/devices/*; do n=${d#*/iommu_groups/*}; n=${n%%/*}; printf 'IOMMU group %s ' "$n"; lspci -nns "${d##*/}"; done on any Linux. But remember that Debian uses a different kernel than Proxmox (which is based on the Ubuntu kernel).
Maybe update the motherboard BIOS if the x16 PCIe slot connected to the CPU causes problems on the Proxmox host with PCI(e) passthrough. Also make sure to early bind the GPU to vfio-pci to work around unbind and reset issues.

There are also many, many threads about PCI(e) passthrough on this forum with tips and work-arounds and even more on the internet as a whole.
Click to expand...
Just got off work, my motherboard's bios is up to date. i need to reinstall and set up the vm for testng passthru
Will edit once I am done and will link some info about what I did
 
leesteken said:
You can run for d in /sys/kernel/iommu_groups/*/devices/*; do n=${d#*/iommu_groups/*}; n=${n%%/*}; printf 'IOMMU group %s ' "$n"; lspci -nns "${d##*/}"; done on any Linux. But remember that Debian uses a different kernel than Proxmox (which is based on the Ubuntu kernel).
Maybe update the motherboard BIOS if the x16 PCIe slot connected to the CPU causes problems on the Proxmox host with PCI(e) passthrough. Also make sure to early bind the GPU to vfio-pci to work around unbind and reset issues.

There are also many, many threads about PCI(e) passthrough on this forum with tips and work-arounds and even more on the internet as a whole.
Click to expand...
I finally had some time to put into this. I did the same install and setup as last time, which happens to include all of your suggestions. Still the same result. From what I'm seeing, my card is effected by the radeon reset glitch. It appears to be very specific, as one series might have it on some of the cards, but not all of them. I've seen vendor-reset as a fix but it has yet to be updated for the 7000 series. Outside of that, I've seen temporary work arounds from scripts inside the guest that run upon booting and shutting down, but no whole system-wide true fix

What are your thoughts about this?
 
I've been bitten by reset (and other passthrough) issues in the past and got into the habit of checking/web-searching before buying. It's also really hard to find USB PCIe add-in cards that use chips that reset properly for example (and have not silently switched to other cheaper chips). I hope you find a work-around for your GPU.

EDIT: AMD 6950XT works great for me with a Windows and Linux VMs. Other AMD GPU work fine with vendor-reset. I've had lots of issues with NVidia in the past.
 
Last edited:
leesteken said:
I've been bitten by reset (and other passthrough) issues in the past and got into the habit of checking/web-searching before buying. It's also really hard to find USB PCIe add-in cards that use chips that reset properly for example (and have not silently switched to other cheaper chips). I hope you find a work-around for your GPU.
Click to expand...
No real workaround. Most likely I'm going to use the Quadro, while i sell the 7800XT and get something like a 4070 TS or maybe 5060 TI If i wanna go lower end. Definitely sad that even with 3-4 years to fix this problem, AMD has taken no action. Should have looked into this before buying the card. And i thought this would be (relatively) easy!

Edit: I have bought a 4070 TI Super, and am selling the 7800XT. Of course when i set up gpu passthrough with my Quadro while I'm waiting for the card to ship, absolutely everything works fine, including rebooting VMs.

Overall, do NOT use AMD cards for passthrough in Proxmox, until they get their sh*t together and fix it. Go NVIDIA.
Anyways.... Who wants a 7800XT??? https://www.ebay.com/itm/336181527379
 
Last edited:
You must log in or register to reply here.
Top Bottom