KVM with GPU passthrough crashes nvlddmkm or CLOCK_WATCHDOG_TIMEOUT on Z690 AORUS PRO, i9 12900K and RTX 3050

[Burdeman]

I have a case open with Gigabyte but wanted to get the forum insight on an issue I’ve been dealing with for a while. My issue is that under normal hardware config and KVM Guest with GPU passthrough I get random crashes under any GPU load. Both gaming and encoding.

My setup:

OS: Archlinux 6.2.8-arch1-1
Motherboard: Z690 AORUS PRO
CPU: Intel Core i9 12900K
Memory: Corsair 2x16GB CMK32GX5M2A4800C40
GPU: GeForce RTX 3050 EAGLE OC 8G

VM: Windows 10 and Windows 11 (tried both) via KVM with 16 vCPU and GPU passthrough

Depending on the KVM config I get a hard crash memory dump with “CLOCK_WATCHDOG_TIMEOUT (101)” or “Display driver nvlddmkm stopped responding and has successfully recovered”.

The only configuration that is somewhat stable is if I move the GPU from PCIEX16 slot to PCIEX4_1 slot and all E-Cores disabled. No other hardware configuration gives me stability under GPU load. E.g. Having card in PCIEX16 and E-Cores disabled still crashes. I’ve tweaked KVM XML to death under normal configuration without success. I’ve also done driver cleans/updates to death. The issue just manifests itself with a different error.

Any thoughts why I have to run the hardware in this configuration? I’d rather be running the GPU in that PCIEX16 slot and be able to use all cores for other VM configurations. I’m open to any suggestions.

[jxdking]

Probably you’ve already done these.

1. In bios, set igpu as intial/primary gpu.
2. Add boot parameter “vfio-pci.ids=“ to mask both vga and audio devices for rtx 3050. Also enable iommu. For example, GRUB_CMDLINE_LINUX_DEFAULT="intel_iommu=on iommu=pt vfio-pci.ids=10de:100c,10de:0e1a"
3. When you passthrough, you have to pass through both vga and audio devices.

ps: if you are using i440fx in qemu, try to use q35 instead. Newer gpu does better with q35.

[Burdeman]

Thanks for your suggestions. I can confirm all those are already the case.