[Improvement] Inaccurate core temps shown in monitor mode for AMD Ryzen chipsets #901
Description
Related:
Problem description:
Currently when running in monitor mode:
sudo auto-cpufreq --monitorWe use the following function to determine cpu temps:
This does a few things:
- grabs sensors from a list in a global
CPU_TEMP_SENSOR_PRIORITY = ("coretemp", "acpitz", "k10temp", "zenpower") - if there's not enough temps in all those lists combined, we just use the average over all for missing values.
This results in the CPU temps not being accurate, as e.g. acpitz takes priority over k10temp. acpitz shows various thermal zones, not necessarily at the CPU itself. k10temp on AMD systems exporses Tdie/Tctl, which are the hottest die temperature/control temperature (what's used for fan curves, based on Tdie) respectively. (This was different on zen1, as these were just delta's I believe, but I digress)..
AMD does not expose temp per core, so ideally we should just show the k10temp for these (either the Tdie or Tctl - some offset).
However doing this per model and figuring out the various offsets could be difficult.
There's also inconsistencies, as --debug uses an entirely different function that just uses the first found temp for all cores:
auto-cpufreq/auto_cpufreq/core.py
Line 853 in 25bd7d9
Concretely, this ends up in the monitor mode showing acpitz temps for my first 4 cores:
acpitz-acpi-0
Adapter: ACPI interface
temp1: +42.8°C
temp2: +42.8°C
temp3: +44.8°C
temp4: +46.8°C <= This is actually the same as the k10temp Tctl, the other values seem to be something else?
And the cores after show the average of these values.
This probably also applies to the GUI.
Proposed changes:
First and foremost, the debug and monitor mode seem to not be sharing their code, and use different functions, this should probably be unified into one working version.
As per the actual fix:
- Based on available sensors we determine the proper values (e.g. we know we're AMD and have Tdie available, show that).
- Possibly calculate the actual temp if it's not exposed (via Tctl)
- If that's too difficult, at least just use a singular value.
- On intel systems, use coretemp, as these show temp per actual core.
System information:
Add/paste output of:
auto-cpufreq --debug
Using settings defined in /etc/auto-cpufreq.conf file
-------------------------------------------------------------------------------
Linux distro: Fedora Linux 43 Workstation Edition
Linux kernel: 6.17.7-300.fc43.x86_64
Processor: AMD Ryzen 9 7940HS w/ Radeon 780M Graphics
Cores: 16
Architecture: x86_64
Driver: amd-pstate-epp
------------------------------ Current CPU stats ------------------------------
CPU max frequency: 4001 MHz
CPU min frequency: 403 MHz
Core Usage Temperature Frequency
CPU0 0.0% 40 °C 403 MHz
CPU1 0.0% 40 °C 403 MHz
CPU2 0.0% 40 °C 403 MHz
CPU3 0.0% 40 °C 403 MHz
CPU4 0.0% 40 °C 403 MHz
CPU5 0.0% 40 °C 403 MHz
CPU6 4.0% 40 °C 544 MHz
CPU7 0.0% 40 °C 403 MHz
CPU8 0.0% 40 °C 403 MHz
CPU9 0.0% 40 °C 403 MHz
CPU10 0.0% 40 °C 544 MHz
CPU11 5.0% 40 °C 971 MHz
CPU12 0.0% 40 °C 403 MHz
CPU13 0.0% 40 °C 544 MHz
CPU14 4.0% 40 °C 403 MHz
CPU15 0.0% 40 °C 403 MHz
CPU fan speed: 0 RPM
auto-cpufreq version: 2.6.0
Python: 3.14.0
psutil package: 7.0.0
platform package: 1.0.9
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: 0.07 %
Total CPU usage: 1.2 %
Total system load: 0.39
Average temp. of all cores: 37.00 °C
Currently using: performance governor
Currently turbo boost is: off
-------------------------------------------------------------------------------