View source for How to keep the GPU (and CPU) cooler

[[category:howto]]
[[category:hardware]]


nouveau errors co-occur with high temperatures (seen from {{cmd|journalctl}}), 
therefore, as an experiment,
I want to try to keep the GPU cool (and see whether the errors will stop occuring).

My graphical errors lead to almost complete freezing of X (say, during Skype calls).

My system is ALT p7 with linux 3.10.21-std-def-alt1 on an HP laptop.

    # lspci | fgrep -i vga
    02:00.0 VGA compatible controller: NVIDIA Corporation C77 [GeForce 8200M G] (rev a2)
    # 

== Keep the Nvidia GPU cooler ==

By reading documentation on the web and by experiments, I have found out that
setting the threshold temperatures (in sysfs) to lower values does cause the GPU
to become cooler!

Current temperatures (and the threshold values except for "fanboost") are
shown by {{cmd|sensors}} command.

I use {{pkg|sysfsutils}} package for setting the values in sysfs on boot.
({{cmd|service sysfs restart}} will reset the parameters to the new values afterwards, when needed, but not {{cmd|service sysfs start}}!)

    #
    # /etc/sysfs.conf - Configuration file for setting sysfs attributes.
    #

    # (temperature is in millidegrees)

    # "fanboost"; default: 90 degrees C
    class/drm/card0/device/hwmon/hwmon0/temp1_auto_point1_temp = 85000
    # "downclock"; default: 95
    class/drm/card0/device/hwmon/hwmon0/temp1_max = 83000
    # "critical"; default: 105
    class/drm/card0/device/hwmon/hwmon0/temp1_crit = 95000
    # "emergency"; default: 135
    # class/drm/card0/device/hwmon/hwmon0/temp1_emergency = 135000

"downclock" seems to bring most effect. Because I want the fan to be silent sometimes,
and "fanboost" turned out not to be a really effective way to cool down the GPU,
I raised the "fanboost" threshold above the "downclock" threshold
(unlike in the default settings).

== And keep the CPU cooler ==

The heat from CPU goes to GPU, too, therefore cooling down the GPU without 
cooling down some really high CPU temperatures is not effective.

I do not know yet of a similar downclock-policy setting mechanism for CPU
which would be based on temperature thresholds, therefore I simply use the brute
force of a governor.

I'm unhappy that it will be active all the time though, of course.
Usually I'm happy with the "ondemend" governor, and only when the temperature 
rises, I want to scale down...

TODO: find out how to do this!

    #
    # /etc/sysfs.conf - Configuration file for setting sysfs attributes.
    #

    devices/system/cpu/cpu0/cpufreq/scaling_governor = powersave
    devices/system/cpu/cpu1/cpufreq/scaling_governor = powersave

==See also==
* [[:ru:cpufreq]]