Tweet
Intel ARC DGPU dramatically lower reported "idle" / desktop power observed.
I hadn't tested the ARC (A770LE DGPU) power consumption between the end of October 2023 and the past few days.
In that elapsed time something may have changed upstream since before recently when I tried to achieve lower
power than ~38-42W in desktop use I failed despite all known / attempted changes in ASPM / kernel / desktop / monitor configurations.
In the past couple of days I tried again, this time running kernels 6.7 and 6.8 on tumbleweed with a single HDMI display.
Now I'm seeing:
~0W GPU-reported power consumption when no monitor is attached or the attached monitor is in "screen blank/off" mode.
~5W GPU power when the single attached (HDMI) monitor is in 1080p30 mode.
~14-17W GPU power when the single attached monitor is in 4k-30Hz mode.
~22W GPU power in 4k-30Hz mode running glxgears.
Using sensors-detect and looking through sysfs I haven't noticed (at a cursory glance) any new controls / sensors associated with the ARC.
So sadly I seem to be lacking "windows intel control center" style exposure of GPU controls / sensors.
If the reported power consumption is accurate and also generally usable / stable, though, it represents a vast improvement over the "at least ~40W" power
consumption I've measured in 2023 so now it may be using ~17W vs 42W in basic desktop use and even "near 0W" vs 42W when the monitor is
detached / off / in screen blank mode -- there's still a lot of room for verification / validation / stability-checking / tuning after my initial experiments
today.
On a more negative note in the past couple of months (perhaps not coincidentally) I have noticed a few GPU-related "crashes" resulting in the loss of
GPU output video until the system was rebooted (and also some syslog errata).
Today, I've seen more syslog complaints from i915 and also the GPU associated intel audio device driver relating to power management typically
correlated wit KVM switching the monitor away from the system, seeing the system's GPU reported power go down to ~0W, and then leaving
the monitor disconnected for a prolonged time and eventually switching the monitor back to the still awake system and trying to get the desktop video
output to come back (use keyboard / move mouse -- monitor may have been seen as detached / asleep but the system itself was always "up").
Sometimes the desktop video comes back quickly, sometimes it has remained absent for some time until reappearing, there may be variable
severity of impact depending on circumstances.
If the gpu utility can be maintained in low power "in use" modes, and can automatically power down greatly concurrent with screen blanking / screen saver / low demand usage modes, et. al. and resume higher performance capability seamlessly upon demand (GPU load, compute, graphics, video, ...) then I'll class it as a major
UX improvement for ARC DGPU & LINUX.
The one setting I've noticed was associated with the savings initiating was setting one of the power saving modes available per.:
cat /sys/module/pcie_aspm/parameters/policy
default performance powersave [powersupersave]
echo powersave > /sys/module/pcie_aspm/parameters/policy
Interestingly, though, when changing the setting "back" to "default" or "performance" I didn't notice any immediate change in the
reported (low) power consumption, which surprised me.
The only other possibly / probably relevant thing in my case is that my MB / BIOS (AMD/X570/ASUS) has no PCIE-ASPM mode control UI options at all,
so some months ago I started adding "pcie_aspm=force" to the kernel command line arguments in case it helped anything, but
it wasn't observed to be useful as of ~october 2023, maybe it helps me now.
0.7405669073935952 W; 0.371582 J; 0.5017534490000001 s;
11.04990341431887 W; 5.535827 J; 0.5009841979999999 s;
I hadn't tested the ARC (A770LE DGPU) power consumption between the end of October 2023 and the past few days.
In that elapsed time something may have changed upstream since before recently when I tried to achieve lower
power than ~38-42W in desktop use I failed despite all known / attempted changes in ASPM / kernel / desktop / monitor configurations.
In the past couple of days I tried again, this time running kernels 6.7 and 6.8 on tumbleweed with a single HDMI display.
Now I'm seeing:
~0W GPU-reported power consumption when no monitor is attached or the attached monitor is in "screen blank/off" mode.
~5W GPU power when the single attached (HDMI) monitor is in 1080p30 mode.
~14-17W GPU power when the single attached monitor is in 4k-30Hz mode.
~22W GPU power in 4k-30Hz mode running glxgears.
Using sensors-detect and looking through sysfs I haven't noticed (at a cursory glance) any new controls / sensors associated with the ARC.
So sadly I seem to be lacking "windows intel control center" style exposure of GPU controls / sensors.
If the reported power consumption is accurate and also generally usable / stable, though, it represents a vast improvement over the "at least ~40W" power
consumption I've measured in 2023 so now it may be using ~17W vs 42W in basic desktop use and even "near 0W" vs 42W when the monitor is
detached / off / in screen blank mode -- there's still a lot of room for verification / validation / stability-checking / tuning after my initial experiments
today.
On a more negative note in the past couple of months (perhaps not coincidentally) I have noticed a few GPU-related "crashes" resulting in the loss of
GPU output video until the system was rebooted (and also some syslog errata).
Today, I've seen more syslog complaints from i915 and also the GPU associated intel audio device driver relating to power management typically
correlated wit KVM switching the monitor away from the system, seeing the system's GPU reported power go down to ~0W, and then leaving
the monitor disconnected for a prolonged time and eventually switching the monitor back to the still awake system and trying to get the desktop video
output to come back (use keyboard / move mouse -- monitor may have been seen as detached / asleep but the system itself was always "up").
Sometimes the desktop video comes back quickly, sometimes it has remained absent for some time until reappearing, there may be variable
severity of impact depending on circumstances.
If the gpu utility can be maintained in low power "in use" modes, and can automatically power down greatly concurrent with screen blanking / screen saver / low demand usage modes, et. al. and resume higher performance capability seamlessly upon demand (GPU load, compute, graphics, video, ...) then I'll class it as a major
UX improvement for ARC DGPU & LINUX.
The one setting I've noticed was associated with the savings initiating was setting one of the power saving modes available per.:
cat /sys/module/pcie_aspm/parameters/policy
default performance powersave [powersupersave]
echo powersave > /sys/module/pcie_aspm/parameters/policy
Interestingly, though, when changing the setting "back" to "default" or "performance" I didn't notice any immediate change in the
reported (low) power consumption, which surprised me.
The only other possibly / probably relevant thing in my case is that my MB / BIOS (AMD/X570/ASUS) has no PCIE-ASPM mode control UI options at all,
so some months ago I started adding "pcie_aspm=force" to the kernel command line arguments in case it helped anything, but
it wasn't observed to be useful as of ~october 2023, maybe it helps me now.
0.7405669073935952 W; 0.371582 J; 0.5017534490000001 s;
11.04990341431887 W; 5.535827 J; 0.5009841979999999 s;