SUMMARY
After updating to kernel 7.0.14-28-pve, Jellyfin running in an unprivileged LXC with Intel iGPU passthrough started to lag when playing a movie, with much higher CPU usage than before the update. Booting 7.0.14-11-pve (same hardware, same config, nothing else changed) makes the problem disappear completely. I pinned 7.0.14-11-pve as a workaround.
ENVIRONMENT
- Proxmox VE 9.2.25
- Host: HP EliteBook laptop, Intel Core i5-7200U (Kaby Lake), HD Graphics 620 (PCI ID 8086:5916), 16 GB RAM, NVMe SSD (MAXIO MAP1202), Intel I219-V NIC
- Bad kernel: 7.0.14-28-pve #2 SMP PREEMPT_DYNAMIC PMX 7.0.14-28 (2026-11-02T11:54Z)
- Good kernel: 7.0.14-11-pve #1 SMP PREEMPT_DYNAMIC PMX 7.0.14-11 (2026-08-06T23:36Z)
- Other kernels installed (not tested): 7.0.14-6, 7.0.14-8, 6.17.13-19, 6.17.13-21, 6.17.4-1, 6.14.x
- CPU governor: performance, ~2.6 GHz, 36-44 C at idle (no throttling seen at idle)
SETUP
- LXC 113 running Jellyfin (installed with the community-scripts helper), jellyfin-ffmpeg 7.1.4
- iGPU shared via /dev/dri/renderD128 (also card1 present in the container), user jellyfin is in the video (44) and render (993) groups
- The same iGPU is also used by Frigate in another LXC
- Jellyfin transcoding configured as Intel QuickSync (QSV), hardware decoding enabled for H264, HEVC, MPEG2, VC1, VP8
- Media on an SMB share (NAS) mounted on the host and bind-mounted into the LXC
SYMPTOMS (kernel 7.0.14-28-pve)
- Playback lags/stutters when a movie starts
- CPU usage of the Jellyfin LXC is clearly higher than before the update (top inside the LXC showed the jellyfin process at about 93% CPU during a transcode/remux, with high I/O wait: about 24% wa in the LXC, about 19% on the host)
- Same behavior on two different clients (Xbox Series S and Samsung Galaxy S24), so it is not a client issue
- Same behavior with a 4K HEVC Dolby Vision file (transcoded to 1080p H264) and with a 1080p Blu-ray ISO (video stream copy, audio DTS-HD transcoded to MP3), and changing the audio track did not help
WHAT I CHECKED (all looked normal on 7.0.14-28-pve)
- lspci shows the iGPU normally
- dmesg: i915 loads without errors, DMC firmware kbl_dmc_ver1_04.bin loaded, no GPU hang/reset messages. The only unusual line is the informational CIFS message about the default SMB dialect change
- vainfo with the iHD driver (Intel iHD driver 24.1.0 on the host, VA-API 1.23 in jellyfin-ffmpeg) lists H264, HEVC Main/Main10, VP9, MPEG2, VC1, JPEG profiles
- jellyfin-ffmpeg log for the 4K HEVC Dolby Vision file: VAAPI decode + tonemap_vaapi + h264_qsv encode, running at about 100-105 fps (about 4.3x realtime), so the hardware pipeline did initialize and work in that test
- NAS throughput from the host: dd on a 3 GB read from the SMB mount gave about 95 MB/s, so storage/network are fine
- Permissions/groups inside the LXC are correct
WORKAROUND / RESULT
proxmox-boot-tool kernel pin 7.0.14-11-pve, then reboot. On 7.0.14-11-pve the lag is gone with exactly the same Jellyfin settings, LXC config and media. I did not test -6, -8 or the 6.17 kernels, and I have not bisected between -11 and -28, so I only know the regression is somewhere between these two builds.
CAVEATS
- I could not capture a side-by-side CPU/GPU comparison (for example intel_gpu_top) between the two kernels. The high CPU usage was observed in top on the bad kernel and was clearly higher than my earlier experience, but I do not have exact numbers from the good kernel.
- Not sure whether the card0 to card1 index in /dev/dri changed between kernels, it is just what I see now.
QUESTIONS
- Has anyone seen similar behavior (higher CPU usage or lag with QSV/VAAPI transcoding in LXC on Kaby Lake or older Intel iGPUs) with 7.0.14-28-pve?
- Is there a known i915, media driver or scheduler change between 7.0.14-11 and 7.0.14-28 that could explain this?
- I am happy to provide more output (intel_gpu_top, perf, journalctl, LXC config /etc/pve/lxc/113.conf, pveversion -v) or test other kernels if it helps.
Thanks!
After updating to kernel 7.0.14-28-pve, Jellyfin running in an unprivileged LXC with Intel iGPU passthrough started to lag when playing a movie, with much higher CPU usage than before the update. Booting 7.0.14-11-pve (same hardware, same config, nothing else changed) makes the problem disappear completely. I pinned 7.0.14-11-pve as a workaround.
ENVIRONMENT
- Proxmox VE 9.2.25
- Host: HP EliteBook laptop, Intel Core i5-7200U (Kaby Lake), HD Graphics 620 (PCI ID 8086:5916), 16 GB RAM, NVMe SSD (MAXIO MAP1202), Intel I219-V NIC
- Bad kernel: 7.0.14-28-pve #2 SMP PREEMPT_DYNAMIC PMX 7.0.14-28 (2026-11-02T11:54Z)
- Good kernel: 7.0.14-11-pve #1 SMP PREEMPT_DYNAMIC PMX 7.0.14-11 (2026-08-06T23:36Z)
- Other kernels installed (not tested): 7.0.14-6, 7.0.14-8, 6.17.13-19, 6.17.13-21, 6.17.4-1, 6.14.x
- CPU governor: performance, ~2.6 GHz, 36-44 C at idle (no throttling seen at idle)
SETUP
- LXC 113 running Jellyfin (installed with the community-scripts helper), jellyfin-ffmpeg 7.1.4
- iGPU shared via /dev/dri/renderD128 (also card1 present in the container), user jellyfin is in the video (44) and render (993) groups
- The same iGPU is also used by Frigate in another LXC
- Jellyfin transcoding configured as Intel QuickSync (QSV), hardware decoding enabled for H264, HEVC, MPEG2, VC1, VP8
- Media on an SMB share (NAS) mounted on the host and bind-mounted into the LXC
SYMPTOMS (kernel 7.0.14-28-pve)
- Playback lags/stutters when a movie starts
- CPU usage of the Jellyfin LXC is clearly higher than before the update (top inside the LXC showed the jellyfin process at about 93% CPU during a transcode/remux, with high I/O wait: about 24% wa in the LXC, about 19% on the host)
- Same behavior on two different clients (Xbox Series S and Samsung Galaxy S24), so it is not a client issue
- Same behavior with a 4K HEVC Dolby Vision file (transcoded to 1080p H264) and with a 1080p Blu-ray ISO (video stream copy, audio DTS-HD transcoded to MP3), and changing the audio track did not help
WHAT I CHECKED (all looked normal on 7.0.14-28-pve)
- lspci shows the iGPU normally
- dmesg: i915 loads without errors, DMC firmware kbl_dmc_ver1_04.bin loaded, no GPU hang/reset messages. The only unusual line is the informational CIFS message about the default SMB dialect change
- vainfo with the iHD driver (Intel iHD driver 24.1.0 on the host, VA-API 1.23 in jellyfin-ffmpeg) lists H264, HEVC Main/Main10, VP9, MPEG2, VC1, JPEG profiles
- jellyfin-ffmpeg log for the 4K HEVC Dolby Vision file: VAAPI decode + tonemap_vaapi + h264_qsv encode, running at about 100-105 fps (about 4.3x realtime), so the hardware pipeline did initialize and work in that test
- NAS throughput from the host: dd on a 3 GB read from the SMB mount gave about 95 MB/s, so storage/network are fine
- Permissions/groups inside the LXC are correct
WORKAROUND / RESULT
proxmox-boot-tool kernel pin 7.0.14-11-pve, then reboot. On 7.0.14-11-pve the lag is gone with exactly the same Jellyfin settings, LXC config and media. I did not test -6, -8 or the 6.17 kernels, and I have not bisected between -11 and -28, so I only know the regression is somewhere between these two builds.
CAVEATS
- I could not capture a side-by-side CPU/GPU comparison (for example intel_gpu_top) between the two kernels. The high CPU usage was observed in top on the bad kernel and was clearly higher than my earlier experience, but I do not have exact numbers from the good kernel.
- Not sure whether the card0 to card1 index in /dev/dri changed between kernels, it is just what I see now.
QUESTIONS
- Has anyone seen similar behavior (higher CPU usage or lag with QSV/VAAPI transcoding in LXC on Kaby Lake or older Intel iGPUs) with 7.0.14-28-pve?
- Is there a known i915, media driver or scheduler change between 7.0.14-11 and 7.0.14-28 that could explain this?
- I am happy to provide more output (intel_gpu_top, perf, journalctl, LXC config /etc/pve/lxc/113.conf, pveversion -v) or test other kernels if it helps.
Thanks!