Hi all,
after a routine apt full-upgrade, the Raphael iGPU on a testing PVE node stopped clocking down at idle. It now sits permanently at the highest DPM state (2200 MHz instead of 600 MHz), which costs roughly 7–9 W
Idle behaviour of the iGPU before and after the upgrade:
The decisive detail is gpu_busy_percent: with 3.18-6 the GPU reports 5–26 % utilisation while the fence counters show exactly one GFX job since boot - there is no work at all. The SMU apparently honours that bogus utilisation reading and keeps the clock pinned at maximum. With 3.18-5 the same idle system correctly reports 0 % and drops to 600 MHz.
Note that the reported utilisation does not correlate with actual power draw (from gpu_metrics): at "5 %" the GPU drew anywhere between 4.8 W and 17.4 W, at "26 %" as little as 5.6 W.
Kernel log confirms the firmware change:
before: Found VCN firmware ... Revision: 15 / DMUB firmware version=0x05003000
after: Found VCN firmware ... Revision: 21 / DMUB firmware version=0x05003300
Does anyone else have similar symptoms?
Happy to provide further data if useful.
after a routine apt full-upgrade, the Raphael iGPU on a testing PVE node stopped clocking down at idle. It now sits permanently at the highest DPM state (2200 MHz instead of 600 MHz), which costs roughly 7–9 W
Idle behaviour of the iGPU before and after the upgrade:
| pve-firmware 3.18-5 | pve-firmware 3.18-6 | |
|---|---|---|
| pp_dpm_sclk (idle) | 600 MHz | 2200 MHz (pinned) |
| vddgfx | 0.750–0.770 V | 0.969 V |
| gpu_busy_percent | 0 % | 5–26 % |
| CPU package power (PPT) | 25–27 W | 30–36 W |
The decisive detail is gpu_busy_percent: with 3.18-6 the GPU reports 5–26 % utilisation while the fence counters show exactly one GFX job since boot - there is no work at all. The SMU apparently honours that bogus utilisation reading and keeps the clock pinned at maximum. With 3.18-5 the same idle system correctly reports 0 % and drops to 600 MHz.
Note that the reported utilisation does not correlate with actual power draw (from gpu_metrics): at "5 %" the GPU drew anywhere between 4.8 W and 17.4 W, at "26 %" as little as 5.6 W.
Kernel log confirms the firmware change:
before: Found VCN firmware ... Revision: 15 / DMUB firmware version=0x05003000
after: Found VCN firmware ... Revision: 21 / DMUB firmware version=0x05003300
Does anyone else have similar symptoms?
Happy to provide further data if useful.