Intel Arc A580 idle power regression with kernel 7.0.14-20/22-pve

seylon79

New Member
Oct 8, 2026
10
1
3
Hi,

I've noticed what appears to be a kernel regression affecting the idle power consumption of an Intel Arc A580.

Hardware:

  • Proxmox VE 9.2
  • AMD Ryzen 5 9600X
  • Intel Arc A580
  • Arc is using the i915 driver
I originally noticed that the Arc was drawing significantly more power than expected at idle. I tested this by booting different installed Proxmox kernels.

Results:

KernelArc A580 idle power
7.0.14-16-pve~6 W
7.0.14-20-pve~40–50 W
7.0.14-22-pve~40–50 W
The difference is reproducible. I can boot 7.0.14-16-pve and the Arc returns to approximately 6 W idle. Booting 7.0.14-20-pve or 7.0.14-22-pve causes the idle consumption to increase substantially.

I also disabled GPU monitoring in Beszel (SKIP_GPU=true) during the testing, so the higher power draw isn't being caused by Beszel polling the GPU.

I verified the power difference using the i915 energy counter rather than relying only on monitoring software. On 7.0.14-22-pve, the counter increased by approximately 3.0 kJ in about one minute, corresponding to roughly 50 W.

Interestingly, I also noticed different behavior from the AMD iGPU between the kernels, although I haven't investigated that separately.

For now I've pinned the host back to 7.0.14-16-pve, where the Arc returns to its normal ~6 W idle consumption.

Is this a known regression in the newer 7.0 kernels, particularly around i915/DRM power management?

Thanks!
 
mhmm could you also check the -17, -18 and -19 kernels? this would shrink the possible regression window by a lot
from a good and a bad boot the following output would also be practical:

Code:
lspci -PP -vvv
 
Last edited:
Thanks, I did some further testing to narrow this down.

I tested 7.0.14-17-pve first. The high idle power consumption was immediately present after boot, so I didn't test -18 or -19 since the regression is already present in -17.

I also tested the latest available kernel, 7.0.14-23-pve, to see whether the issue had been fixed there. Unfortunately the regression is still present.

My current results are:

KernelArc A580 idle power
7.0.14-16-pve~6 W
7.0.14-17-pve~44 W
7.0.14-20-pve~40–50 W
7.0.14-22-pve~40–50 W
7.0.14-23-pve~42 W
I also verified the -17 result using the i915 energy counter. Over approximately one minute it increased from 7,582,748,413 µJ to 10,209,383,605 µJ, corresponding to approximately 44 W.

After returning to 7.0.14-16-pve, the Arc immediately returned to its normal ~6 W idle consumption.

I've attached the requested lspci -PP -vvv output and diagnostics from the 7.0.14-17-pve boot.

So at this point the regression appears to be introduced somewhere between 7.0.14-16-pve and 7.0.14-17-pve, and it is still present in the current 7.0.14-23-pve.

Thanks!
 

Attachments

can you also attach a lspci output from a bad kernel?

EDIT: also please run lspci as root, otherwise some information is not visible there (such as the capabilities, etc.)
 
Last edited:
My apologies. i thought i had run it as root. I've now captured the lspci -PP -vvv output as root from both kernels so there's a direct comparison between the known-good and known-bad versions.

I've attached both files:

  • lspci-PP-vvv-root-16.txt from 7.0.14-16-pve (known good, ~6 W idle)
  • lspci-PP-vvv-root-17.txt from 7.0.14-17-pve (first known bad, ~44 W idle)
Hopefully the comparison provides some useful information.
 

Attachments

mhmm the only thing that stands out is that ASPM L1 is enabled on the working system while not on the not working one.

between those kernels there are only a few ASPM related changes, and only one is generic:
https://lore.kernel.org/all/20260428040104.78524-1-carlos.bilbao@kernel.org/

not sure if it's the reason for the behavior though.

you might want to report the behavior upstream to the kernel though if you can reproduce with an upstream kernel..
 
Thanks for taking a look.

I'm fairly new to homelabbing, and I only noticed this because the server's fans suddenly started getting noticeably louder than they used to. That's what led me to investigate the GPU power draw in the first place.

I'm not familiar with kernel development, so I'm not entirely sure what you mean by reproducing this with an upstream kernel or how I'd go about testing that. Is that something I should look into, or is there anything else you'd suggest I do at this point?
 
it would entail installing an upstream kernel, best on a test machine and if it's still the upstream behavior open a bug on https://bugzilla.kernel.org/
i don't have a a580 to test here at the moment so I can't do the test myself sadly
 
Thanks for all the help and for taking the time to look into this. I think testing upstream kernels and opening a kernel bug report is getting a bit beyond my current skill set, so I'll stick with the known-good kernel for now and keep testing newer Proxmox kernels as they become available. Hopefully the issue gets resolved in a future update. Thanks again!
 
I can confirm a very similar ASPM regression on a different hardware platform, using an Intel Arc A380 with VFIO passthrough.

Hardware and software​

  • Motherboard: Supermicro H12SSL-C
  • CPU: AMD EPYC 7313
  • GPU: ASRock Intel Arc A380 (DG2, PCI ID 8086:56a5)
  • GPU configuration: PCIe passthrough via vfio-pci to a Debian 13 VM
  • Proxmox VE: 9.2.21
  • QEMU: pve-qemu-kvm 11.0.3-4
  • Guest kernel at the time of the regression: 6.12.107+deb13-amd64

When the problem started​

On September 21, 2026, I upgraded the Proxmox host kernel from 7.0.14-16-pve to 7.0.14-17-pve.

Following that update and reboot, my system's idle power consumption increased by approximately 15–20 W and remained elevated.

The Debian guest was running the same kernel (6.12.107) before and after the host update, so there was no simultaneous guest kernel change.

Kernel comparison​

Today (October 26), I booted the host back into 7.0.14-16-pve without changing the BIOS, hardware, GPU passthrough configuration or PCIe settings.

Both the TrueNAS VM and the Debian VM with the Arc A380 were running during the measurements. All eight HDDs were in standby.

7.0.14-25-pve7.0.14-16-pve
System idle power~140–145 W~125 W
AMD PCIe root port (80:01.1)ASPM DisabledASPM L1 Enabled
Intel PCIe bridge (81:00.0)ASPM DisabledASPM L1 Enabled
Power consumption was measured externally using a smart plug. The measurement also includes some network equipment, which remained unchanged between tests.

With the older kernel, the idle power consumption returned to its previous level immediately after booting and spinning down the HDDs.

PCIe ASPM configuration​

The ASPM policy with the working kernel is:

Code:
default performance [powersave] powersupersave
Relevant kernel parameters:

Code:
amd_iommu=on iommu=pt
There is no pcie_aspm=off or pcie_aspm=force override.

With kernel 7.0.14-16-pve, both affected PCIe links report:

Code:
LnkCtl: ASPM L1 Enabled
With kernel 7.0.14-25-pve, both report:

Code:
LnkCtl: ASPM Disabled

Conclusion​

The combination of the power measurements and the PCIe configuration differences strongly suggests an ASPM regression introduced between 7.0.14-16-pve and 7.0.14-17-pve.

This is particularly interesting because the affected Intel Arc GPU is passed through to a VM using VFIO, rather than being driven by i915 directly on the Proxmox host.

I have attached the complete lspci -PP -vvv output collected as root under the working kernel 7.0.14-16-pve, together with the PCIe topology, kernel parameters, ASPM policy and installed Proxmox package versions.

The full PCIe dump from the affected kernel was not captured, but the relevant ASPM states were verified directly.

I'm currently using 7.0.14-16-pve as a temporary workaround.
 

Attachments

Update (30 October 2026):

Tested kernel 7.0.14-26-pve. Unfortunately, the Intel Arc A580 power draw issue still persists, with idle power consumption returning to approximately 40W.

Kernel 7.0.14-16-pve remains pinned, as it is the last known working version without the issue.
 
I can confirm this as well. I have an A380 and an A580.

Kernel: Linux 7.0.14-29-pve

A380: 248w 12 hour average
A580: 292w 12 hour average

44w delta
 
Would be interesting to see what you see with the 7.3 series just announced:
 
I have this script that I use to enable ASPM on ConnectX-3 network cards.... I wonder if it'd work for you given your post in the kernel thread.

I call it as:
Code:
enable-aspm.sh 23:00.0 3

23:00.0 is my NIC from `lspci`, and 3 is to enable L1 and L0 entries.
 

Attachments

Update: Intel Arc A580 idle power regression traced to PCIe ASPM L1​

I've made some progress investigating the high idle-power consumption of my Intel Arc A580 on newer Proxmox kernels.

I believe I've found a reproducible workaround, and the results strongly suggest that the issue is related to ASPM configuration on the upstream PCIe link.

Hardware​

  • Motherboard: ASRock B850 Pro-A
  • CPU: AMD Ryzen 5 9600X
  • GPU: Intel Arc A580 (8086:56a2)
  • Driver: i915
  • Proxmox VE 9.2
Relevant PCIe topology:

AMD Root Port 00:01.1
└── Intel PCIe Switch Upstream Port 01:00.0
├── Downstream Port 02:01.0
│ └── Arc A580 Graphics 03:00.0
└── Downstream Port 02:04.0
└── Arc A580 Audio 04:00.0
The upstream link is PCIe Gen4 x16.

Observed behaviour​

The known-good kernel 7.0.14-16-pve gives approximately 5–6 W GPU idle power.

On newer kernels, idle consumption increases to approximately 40–45 W.

Some measurements:

KernelGPU idle power
7.0.14-16-pve~6 W
7.0.14-17-pve~43.8 W
7.0.14-23-pve~42 W
7.0.14-30-pve~42.7 W
7.3.5-1-pve~44.4 W

PCIe ASPM investigation​

On the affected kernels, I noticed that standard ASPM L1 was disabled on both:

  • AMD root port 00:01.1
  • Intel switch upstream port 01:00.0
The downstream links to the GPU graphics and audio functions already had L1 enabled.

Both upstream devices advertise support for standard ASPM L1. The AMD root port does not advertise L1.1/L1.2 support, so I have only tested standard L1.

Workaround​

Manually enabling L1 on both ends of the upstream link:

setpci -s 01:00.0 CAP_EXP+10.w=0002:0003
setpci -s 00:01.1 CAP_EXP+10.w=0002:0003
Immediately reduces GPU idle power from approximately 42–43 W to around 6 W.

I also performed a reverse test on 7.0.14-30-pve.

Disabling L1 again:

setpci -s 00:01.1 CAP_EXP+10.w=0000:0003
setpci -s 01:00.0 CAP_EXP+10.w=0000:0003
caused idle power to return to approximately 40–45 W.

Re-enabling L1 restored the low idle power.

This was repeatable.

Workload testing​

With L1 enabled on 7.0.14-30-pve, I tested:

  • Immich thumbnail processing
  • Immich machine-learning processing
  • Jellyfin hardware transcoding
All worked normally.

During Jellyfin transcoding, GPU power increased to approximately 60.5 W and returned to approximately 6 W after playback stopped.

The upstream link continued reporting ASPM L1 enabled during transcoding. No obvious i915, PCIe AER or DPC errors appeared in the kernel logs.

Current workaround​

I have implemented a small systemd oneshot service called arc-aspm-fix.service which executes the two masked setpci writes automatically at boot.

I subsequently tested this on 7.3.5-1-pve.

The service executes successfully after reboot, both upstream links report ASPM L1 enabled, and GPU idle power remains around 5–6 W.

Questions​

  1. Is there a known regression or change in PCIe ASPM handling that could explain why L1 is disabled on these upstream links?
  2. Is there a recommended way to configure this correctly through the kernel rather than using setpci?
  3. Could this be investigated as a kernel regression affecting this particular AMD root port / Intel PCIe switch combination?
The workaround is currently working, but I'd prefer a proper kernel-level fix rather than maintaining a local systemd service indefinitely.

Hopefully these findings are useful to others experiencing similar Intel Arc idle-power problems.