thermal_monitor

A terminal telemetry daemon. It samples CPU load, package temperature and memory use twice a second, renders them as a live-updating table, and writes every tick to a local SQLite database so a run can be reviewed after the fact.

How it works

• One coherent sample per tick. HardwareSensors.sample() polls the CPU exactly once and threads that reading through the derived values. Sampling per-sensor meant three separate blocking cpu_percent calls per frame and three different load figures inside a single row.

• Load covers the whole run. Each tick reports CPU load since the previous tick rather than blocking for a 0.1 s poll, which measured one sixth of every loop and slowed it to 1.7 Hz.

• Thermal readings with a fallback. Package temperature is read from coretemp (Intel), k10temp or zenpower (AMD), or cpu_thermal (ARM) via psutil.sensors_temperatures. Those thermal zones are not exposed on Windows or inside WSL, so when the read fails the daemon models a curve from load instead of reporting nothing, labels the row (est.), and stores the tick with temp_measured = 0.

• Sampling, logging and rendering are separate. The table builder is pure. Only the loop writes rows, so a repaint never inserts a duplicate tick.

• Persistent history. Each tick lands in the system_metrics table of thermal_grid.db, next to main.py wherever you start it from, so you can query a session afterwards:

  SELECT MAX(cpu_temp), AVG(cpu_usage) FROM system_metrics
  WHERE timestamp > '2026-08-30' AND temp_measured = 1;

Thresholds

Rows turn red past the limits set at the top of main.py: 80% sustained CPU load, 65 °C package temperature.

Run

git clone https://github.com/apollo-2006/thermal_monitor.git
cd thermal_monitor

pip install -r requirements.txt
python main.py

Ctrl-C stops the run and closes the database cleanly.

Layout

main.py             loop, thresholds, table rendering
core/sensors.py     psutil reads, thermal fallback, one-shot sample()
core/storage.py     SQLite schema and per-tick inserts

Known limits

• VRAM clock is modelled, not measured. get_vram_clock_sim steps through the three clock states a card uses, keyed off system activity. Reading the real value needs PyNVML (NVIDIA) or PyAMDGPUInfo (AMD).
• The temperature fallback is an estimate. When no thermal driver is available the number shown is derived from load, not measured from the die. It is labelled as such on screen and in the database.
• Every tick commits. At 2 Hz that is two fsync-backed commits a second, which is fine for a foreground session and wasteful for a long-running daemon.

License

MIT. See LICENSE.

Author

Abir Deol · abirdeol.tech