proxmox_i915_sriov.md

The following works for pve 8 with 6.5 kernel.

The driver is still experimental, use with your own discretion.

• prepare env

git clone https://github.com/strongtz/i915-sriov-dkms
apt install dkms build-essential proxmox--headers-6.5 

• modify dkms.conf

PACKAGE_NAME="i915-sriov-dkms"
PACKAGE_VERSION="6.1"

• build dkms deb package

# add module
dkms add /root/i915-sriov-dkms

# check module status
dkms status

# build
dkms build i915-sriov-dkms/6.1

# install
dkms install i915-sriov-dkms/6.1

# if messed up
dkms uninstall i915-sriov-dkms/6.1
dkms remove i915-sriov-dkms/6.1 --all

• update grub command line, and add intel_iommu=on i915.enable_guc=3 i915.max_vfs=7 to GRUB_CMDLINE_LINUX_DEFAULT, and run update-grub

• reboot and check dmesg to see if it is working. If some i915 firmware file is missing, and there are errors like this

i915 0000:00:02.0: GuC firmware i915/tgl_guc_70.bin: fetch failed with error -2
i915 0000:00:02.0: [drm] GuC firmware(s) can be downloaded from https://git.kernel.org/pub/scm/linux/kernel/git/firmware/linux-firmware.git/tree/i915
i915 0000:00:02.0: [drm] GuC firmware i915/tgl_guc_70.bin version 0.0.0
i915 0000:00:02.0: GuC initialization failed -2

so download the missing firmware file and place it in /usr/lib/firmware/i915, here it is tgl_guc_70.bin for gen 11 CPU. Firmware can be downloaded from here

If you see the following, it is successful.

[  +0.000121] i915 0000:00:02.0: 7 VFs could be associated with this PF

• now it is possible to specify the number of VFs needed with the following.

echo 2 > /sys/devices/pci0000:00/0000:00:02.0/sriov_numvfs

Thank you for the quick response. I feel like I didn't do anything different and now it is working. I do have a 12th gen i5-12500 (UHD77)0)

This is the guide that got me there in case someone comes across this comment section.
https://www.michaelstinkerings.org/gpu-virtualization-with-intel-12th-gen-igpu-uhd-730/

VM, no. Because I am using an unfortunate gen 11 CPU, which has some serious issue that breaks vt-d. No matter if is a physical CPU or VF, it will not work when passed through to a VM. In Windows, usually an error code 43 will be reported. If you are using gen 12 intel, i.e. alder lake, you should have no issue.

For now, I am passing through one VF to a lxc container with bind mount, and jellyfin hardware accelerated encoding works just fine.

Is that confirmed? Can I give up trying to get the GPU of my Intel i5 1135G7 used by a VM in any way?

Is that confirmed? Can I give up trying to get the GPU of my Intel i5 1135G7 used by a VM in any way?

It is confirmed (not officially from intel) there is some issue with i915 driver on Windows with gen 11 CPU. Considering gen11 is no more a shiny new generation, I would not be surprised that Intel does not hurry to have it fixed.

Also I hear people have some success with latest Debian, though I could not recall if it was bookworm or sid. Anyway it could work with a Linux VM provided that it comes with up-to-date driver.

It is confirmed (not officially from intel) there is some issue with i915 driver on Windows with gen 11 CPU. Considering gen11 is no more a shiny new generation, I would not be surprised that Intel does not hurry to have it fixed.

Also I hear people have some success with latest Debian, though I could not recall if it was bookworm or sid. Anyway it could work with a Linux VM provided that it comes with up-to-date driver.

Damn. I really need to give some sort of GPU to a Win 11 VM. I failed with a Ryzen based mini PC and when I needed a second mini PC I went with an Intel based one in order to swap them around. One of the few that fit my criteria was a 12th gen but had an overheating problem and the replacement I found is (unfortunately) an I5 1135G7 Intel® 11th Gen CoreTM Processor ( Tiger Lake UP3) so it looks like I failed again.
I was hopeful when I read about the new code regarding i915 in the 6.8 kernel but after reading your reply I realize it's the Windows driver part which fails :-(

Thanks for your guide, at least I have managed to give GPU access to a few docker containers although that was not my main objective :-)