Intel 125U chipset unusable with proxmox or bad hardware? Constant system hangs/freezes - FIX: Pre-allocate iGPU memory using lgfx Gsm2 in BIOS

R

richardh_nz

New Member
Jul 9, 2026
8
0
1
I purchased a new GMKtec K15 AI Mini PC Oculink with Intel Core Ultra 5 125U Processor for use with proxmox and it has been unstable from day one. Longest uptime is 3 days, shortest is a few minutes. System hangs with no messages in any log file and requires forced reboot. Works brilliantly when it works, but obviously totally unusable as a home server. Seems to hang faster when the GPU is used.

Has anyone successfully used proxmox on this or similar hardware?

I have put in the hours with this... here is a list of attempted fixes (likely forgetting a few).

  • Memtest86 overnight - no issues found with memory
  • Turning off power management for CPU cstates both in BIOS and kernel
    • GRUB_CMDLINE_LINUX_DEFAULT="quiet intel_idle.max_cstate=1 pcie_aspm=off i915.force_probe=!7d45 xe.force_probe=7d45"
  • Forcing GPU to use i915 drivers
    • disabling panel refresh with i915.enable_psr=0
  • Forcing GPU to use xe drivers
  • Checking that Intel GUC firmware is up to date
  • Forcing GPU interrupts to use one CPU core and disabling power mgmt on that core (https://bbs.archlinux.org/viewtopic.php?id=308313)
  • Trying both 6.17 and 7.0.14-6 kernels
    • Found a mention that 6.12 is stable with similar hardware but it wasn't easily accessible anymore
Have basically run down every rabbit hole that Google, Gemini and Claude have suggested and can't keep the system up for more than a few hours if I am using the GPU in a LXC container via passthrough and maybe a day or two if not.

So.. is this hardware just not really supported by linux at this point so it goes back to Amazon or is the hardware defective (so it goes back to Amazon) or do you have any further suggestions to troubleshoot/fix?

As a followup.. any suggestions on similar hardware that is known good using some very expensive DDR5 SODIMM memory

TIA
 
Last edited:
In Germany i see may owner of this sort mini pc, who think they are Proxmox VE Servers.
No they don't.

Most of the time a Intel E1000 Problem and Temperate cooling Problem of the output TDP the CPU, the RAM and NVMe.

Is this your device:
# https://de.gmktec.com/en/products/gmktec-nucbox-k15-intel®-core™-ultra-5-125u

If there are LAN Chip with: 2 × 2,5G LAN, Realtek 8125BG, then this my be fine.

Can your monitor with some DS18B20 on different places on your Mini PC the temperatures?
My be you can program a AVR Microcontroller with external storage?
 
I build some Promox VE and BS System, based on AMD AM4 Hardware with Ryzen 5000 CPU.
You can start with a Ryzen 5 5600(GT) 6c/12T and 32 GB DDR4 Ram.
 
news said:
In Germany i see may owner of this sort mini pc, who think they are Proxmox VE Servers.
No they don't.

Most of the time a Intel E1000 Problem and Temperate cooling Problem of the output TDP the CPU, the RAM and NVMe.

Is this your device:
# https://de.gmktec.com/en/products/gmktec-nucbox-k15-intel®-core™-ultra-5-125u

If there are LAN Chip with: 2 × 2,5G LAN, Realtek 8125BG, then this my be fine.

Can your monitor with some DS18B20 on different places on your Mini PC the temperatures?
My be you can program a AVR Microcontroller with external storage?
Click to expand...
Yes it is that system you linked to. I don't think it is an overtemp issue as the machine is very lightly loaded almost all of the time and the fans aren't particuarly ramped up until it hangs at which point the fans go 100%.
 
news said:
I build some Promox VE and BS System, based on AMD AM4 Hardware with Ryzen 5000 CPU.
You can start with a Ryzen 5 5600(GT) 6c/12T and 32 GB DDR4 Ram.
Click to expand...
Unfortunately I have very expensive DDR5 that I would want to re-use. Were you able to get the Ryzen GPUs available to LXC? That seemed troublesome as well.
 
richardh_nz said:
Unfortunately I have very expensive DDR5 that I would want to re-use. Were you able to get the Ryzen GPUs available to LXC? That seemed troublesome as well.
Click to expand...
No i don't need or habe i GPU on some systems, they run headless.

So what will you do now? get a DDR5 conversion kit for your SO-DIMM DDR5 Kit?

And by a Ryzen 7000 or 8000 CPU, example AMD Ryzen™ 7 8700G, and a good Mainboard.
A Silentware Bora Performance, TDP 215 Watt CPU Cooler, and a used case.

Please tell us, what will you do with your Proxmox VE?
Do you plan to use a Proxmox BS (Backup Server) on ZFS?
 
that k15 sounds flaky. that computer shouldnt have any issue with proxmox. its essentially a laptop without screen and battery.
my first suspects would be memory or drives.
run memtestx86 on that machine for a week (yes a whole week).
the drive could be a dud. if you have another drive, see if it works better with that one (after letting memtest run for a week).

i have the k12 with proxmox and dont have any noteworthy issues with it.
just in case you are looking for an alternative with ddr5 memory.
 
I will be receiving the same k15 (with 48gb ram) this week. and goal is to install proxmox for jellyfin, dockers containers, and vm's (like ubuntu & home assistant).
I'll keep you posted about my progress.

Jellyfin will run as a lxc container for easy(??) igpu passthrough.

Perhaps some of your problems are due to the hybrid architecture of the intel processor (P en E cores) ?

I read somewhere that it is best to leave blank the number of cores on lxc containers, when using those intel processors and instead change the CPU units (percentage). In doing this, the Linux kernel scheduler (specifically CFS) automatically assigns tasks to Performance cores (P-cores) or Efficient cores (E-cores) based on workload demand.
 
jackydec said:
I will be receiving the same k15 (with 48gb ram) this week. and goal is to install proxmox for jellyfin, dockers containers, and vm's (like ubuntu & home assistant).
I'll keep you posted about my progress.

Jellyfin will run as a lxc container for easy(??) igpu passthrough.

Perhaps some of your problems are due to the hybrid architecture of the intel processor (P en E cores) ?

I read somewhere that it is best to leave blank the number of cores on lxc containers, when using those intel processors and instead change the CPU units (percentage). In doing this, the Linux kernel scheduler (specifically CFS) automatically assigns tasks to Performance cores (P-cores) or Efficient cores (E-cores) based on workload demand.
Click to expand...
Will let you know if I find any fixes. So far no real confirmation either way. I am all out of ideas and am going to put Ubuntu 26.04 on it shortly to see if the problem is general or OS related.
 
So, the K15 (48gb ram 1tb ssd) arrived on tuesday. Didn't touch the bios, only to put in on 'Silent' mode to be energy efficient.
installed 'vanilla' Proxmox on it. (network via wired connection & fixed ip address)

installed Jellyfin as a priviledged lxc container using the community script ( https://community-scripts.org/scripts ) . (cores on 'unlimited' to avoid issues with P & E cores)
The script detected the iGPU and installed & configured the necessary drivers on the lxc container. I just had to install nfs-common package and adapt fstab to access the media on NAS (synology) and since that moment jellyfin was also up and running.
tested that iGPU transcoding was working with intel_gpu_top and that seems to be ok as well.

Will keep it 'on' for a couple of days and nights to see if the system remains stable.

Tried also to install Immich with the community script ( https://community-scripts.org/scripts/immich) but that failed during install, but this seems to be a known issue, not related to the device.

Thinking about clustering this install with my old 10th gen Nuc ( i7-10710U CPU). (perhaps for HA for Home Assistant?)
 
Last edited:
jackydec said:
So, the K15 (48gb ram 1tb ssd) arrived on tuesday. Didn't touch the bios, only to put in on 'Silent' mode to be energy efficient.
installed 'vanilla' Proxmox on it. (network via wired connection & fixed ip address)

installed Jellyfin as a priviledged lxc container using the community script ( https://community-scripts.org/scripts ) . (cores on 'unlimited' to avoid issues with P & E cores)
The script detected the iGPU and installed & configured the necessary drivers on the lxc container. I just had to install nfs-common package and adapt fstab to access the media on NAS (synology) and since that moment jellyfin was also up and running.
tested that iGPU transcoding was working with intel_gpu_top and that seems to be ok as well.

Will keep it 'on' for a couple of days and nights to see if the system remains stable.

Tried also to install Immich with the community script ( https://community-scripts.org/scripts/immich) but that failed during install, but this seems to be a known issue, not related to the device.

Thinking about clustering this install with my old 10th gen Nuc ( i7-10710U CPU). (perhaps for HA for Home Assistant?)
Click to expand...
Please report back in a few days how it is going. If you can get more than 3 days uptime it has to be broken hardware on my end.
 
reporting back :
richardh_nz said:
Please report back in a few days how it is going. If you can get more than 3 days uptime it has to be broken hardware on my end.
Click to expand...
reporting back : +- 1 week 24/7 up without any problems : no reboot of the proxmox server, no crashes of lxc or vm's.

running 5 lxc containers (jellyfin, notediscovery, immich, podman, plex,) and 2 VM's (win11 & ubuntu server) :for test purposes.
( some of the lxc have iGPU passthrough and working)

also i've put home assistant vm on it, and did a test migration of my old config, without any problems.

How did it go, installing ubuntu 26.04 on it instead of proxmox. Did you experience the same issues?

So, my guess : unfortunately indeed faulty hardware on your side.
 
Thanks for your report. Bad hardware is also my conclusion. Trying to install 26.04 or 24.04 I couldn't get through GUI installer without hanging. I think the lack of GPU usage by proxmox hid the bad hardware that would have been instantly obvious w/ a desktop install.

That one is going back to Amazon and new one is in box next to me now.
 
FIX: Pre-allocate fixed iGPU memory using lgfs Gsm2 setting in BIOS

After some additional troubleshooting involving borrowed memory that worked perfectly without hangs (Kingston KVR56S46BS8-16 PC5 5600 CL46) I finally got the system stable using my memory (Corsair CMSX32GX5M2A4800C40 Vengeance DDR5 SODIMM 32GB (2x16GB) DDR5 Up to 4800MHz C40) by finding the Igfx Gsm2 (Graphics Selected Memory / Dedicated Pre-allocated VRAM) setting in BIOS and pre-allocating fixed 4GB to the iGPU. All of the triggers that made the system hang no longer do.

My memory is slightly slower. Somehow the dynamic re-sizing of iGPU memory either didn't happen fast enough or smth triggering the system hangs.

I have wasted a huge number of hours over past 5 weeks on this. Hope it helps someone else finding this thread.

​

 
You must log in or register to reply here.
Top Bottom
Back