AdnanHodzic / auto-cpufreq

Automatic CPU speed & power optimizer for Linux
https://foolcontrol.org/?p=4603
GNU Lesser General Public License v3.0
5.31k stars 259 forks source link

Nvidia GPU not using all it's power when using auto-cpufreq #714

Open MiMillieuh opened 1 month ago

MiMillieuh commented 1 month ago

Fill out information requested in this template, without doing so issue will be ignored & closed!

Have you tried?


System information:

Add/paste output of:

mimillie@tataru-os:~/auto-cpufreq$ sudo auto-cpufreq --debug

Using settings defined in /etc/auto-cpufreq.conf file
could not get value from conservation mode

-------------------------------- Battery Info ---------------------------------

battery count = 1
ERROR: failed to read BAT0 thresholds: [Errno 2] No such file or directory: '/sys/class/power_supply/BAT0/charge_start_threshold'

-------------------------------------------------------------------------------

Linux distro: Fedora Linux 40 Silverblue
Linux kernel: 6.8.10-300.fc40.x86_64
Processor: 13th Gen Intel(R) Core(TM) i9-13905H
Cores: 20
Architecture: x86_64
Driver: intel_pstate

------------------------------ Current CPU stats ------------------------------

CPU max frequency: 2600 MHz
CPU min frequency: 400 MHz

Core    Usage   Temperature Frequency
CPU0      0.0%        51 °C      1221 MHz
CPU1      0.0%        51 °C       400 MHz
CPU2      1.0%        50 °C       400 MHz
CPU3      0.0%        50 °C       400 MHz
CPU4      7.1%        47 °C       971 MHz
CPU5      0.0%        47 °C       400 MHz
CPU6     11.0%        48 °C      1635 MHz
CPU7      0.0%        48 °C       400 MHz
CPU8      3.9%        47 °C       399 MHz
CPU9      0.0%        47 °C       400 MHz
CPU10      5.0%        47 °C      1531 MHz
CPU11      2.0%        47 °C       400 MHz
CPU12      2.0%        51 °C       400 MHz
CPU13      3.0%        51 °C       795 MHz
CPU14      0.0%        51 °C       919 MHz
CPU15      3.9%        51 °C       400 MHz
CPU16      1.0%        49 °C       400 MHz
CPU17      1.0%        49 °C      1154 MHz
CPU18      4.0%        49 °C       900 MHz
CPU19      0.0%        49 °C      1637 MHz

auto-cpufreq version: 2.3.0 (git: 3f6d7a3)

Python: 3.12.3
psutil package: 5.9.8
platform package: 1.0.8
click package: 8.1.7
distro package: 1.9.0

Computer type: Notebook
Battery is: charging

auto-cpufreq system resource consumption:
cpu usage: 0.0 %
memory use: 0.05 %

Total CPU usage: 2.4 %
Total system load: 0.76
Average temp. of all cores: 49.00 °C 

Currently using: powersave governor
Currently turbo boost is: off

-------------------------------------------------------------------------------

Hello, So since I'm trying to use auto-cpufreq on all my laptops, I had the same issue which is super annoying on newer laptops. Nvidia Power stay at the lowest possible like if I'm on energy saving and on battery with power-profile-daemon. that results in half the available wattage to not being used by my GPU which is a deal breaker.

Here are the differences on my laptop :

Without auto-cpufreq (plugged in)

Capture d’écran du 2024-05-29 17-40-06

With auto-cpufreq (plugged in)

Capture d’écran du 2024-05-29 17-41-47

I've tried to :

Additional info : The screenshots you see have been taken while running Unigine Superposition on the Nvidia GPU.

If anyone could help that would be nice. Thanks and have a nice day :)


AdnanHodzic commented 1 month ago

Hm, not sure what's happening as auto-cpufreq shouldn't have any effect on GPU.

Curious to hear what happens if you setup config file, and set energy_performance_preference to other value then performance, would this change anything?

Will also ask on our Discord community if anyone else has Nvidia and has faced this problem.

Angel-Karasu commented 1 month ago

This may be due to the PowerMizer configuration in the Nvidia settings.

You can get the value using nvidia-settings -q GPUPowerMizerMode, 0 is adaptative, 1 is performance mode and 2 is auto

Try to set it in Performance mode using sudo nvidia-settings -a "[gpu:0]/GpuPowerMizerMode=1"

MiMillieuh commented 4 weeks ago

Unfortunatly I still encounter the issue...

by the way I can't querry the powerMizer on my GPU :

mimillie@tataru-os:~$ sudo nvidia-settings -a "[gpu:0]/GpuPowerMizerMode=1"

mimillie@tataru-os:~$ nvidia-settings -q GPUPowerMizerMode

ERROR: Error resolving target specification '' (No targets match target
       specification), specified in query 'GPUPowerMizerMode'.

mimillie@tataru-os:~$ nvidia-settings -q "[gpu:0]/GpuPowerMizerMode"

ERROR: Error resolving target specification 'gpu:0' (No targets match target
       specification), specified in query '[gpu:0]/GpuPowerMizerMode'.

What's even stranger is that the GPU is sitting at 55W limit while doing nothing but when I ask for the GPU to be used it limits itself to around 40W

Without auto-cpufreq I set via gnome the profile to performance and it can use up to 80W if the load is enough

Angel-Karasu commented 4 weeks ago

Try to get the value using the GUI of nvidia settings

MiMillieuh commented 4 weeks ago

Try to get the value using the GUI of nvidia settings

Where can I find it ? By the way I use Wayland so I don't have the full Nvidia settings.

Angel-Karasu commented 3 weeks ago

Normally in nvidia-settings, there should be a PowerMizer section under the GPU.

MiMillieuh commented 3 weeks ago

I know there is on Xorg but not on Wayland :

image

Angel-Karasu commented 3 weeks ago

I don't know, maybe wait for the 555 driver, which improves compatibility with Wayland.

MiMillieuh commented 3 weeks ago

Hm, not sure what's happening as auto-cpufreq shouldn't have any effect on GPU.

Curious to hear what happens if you setup config file, and set energy_performance_preference to other value then performance, would this change anything?

Will also ask on our Discord community if anyone else has Nvidia and has faced this problem.

I've tried multiple settings in the config but I still get the same results.

I managed to have my GPU using 80W with autocpu-freq with this command : sudo echo "performance" | sudo tee /sys/firmware/acpi/platform_profile

However that's unoptimal as I need root acces to change the profile of my GPU and once done it nevers goes back to its low power state... when using balanced it just keep the old profile and low-power is what I get with balanced in power-profile-daemon currently. And it may change something else I don't know I haven't tested it for too long since it's really not a great solution in my case.

There's probably something that trigger the GPU driver to change it's power limit but I can't find what.

All I know is that power-profile-daemon trigger it correctly.