ZC Technologies Trust Per-node monthly license rev. 01 build 21da0fd6
A small daemon that watches GPU temperature and clamps the clock ceiling in 150 MHz steps when the silicon enters the warning band. Stock NVIDIA actuator, hysteresis controller, auto-relax when temperature returns to floor.
Hot band ≥ 78 °C → step max clock down. Warm band → hold. Cool band ≤ 72 °C for 3 consecutive samples → step back up. Floor 1800 MHz, ceiling 3000 MHz.
| Time | Temp | Clock | Util | Action |
|---|---|---|---|---|
| 07:46:28 | 82 °C | 2463 MHz | 94 % | STEP_DOWN |
| 07:47:28 | 83 °C | 2463 MHz | 94 % | STEP_DOWN |
| 07:56:29 | 76 °C | 1976 MHz | 95 % | HOLD |
| 08:13:44 | 72 °C | 2093 MHz | 94 % | HOLD (cool streak 1) |
| 08:14:14 | 72 °C | 2093 MHz | 94 % | HOLD (cool streak 2) |
Same Ollama workload, same util band. Temperature fell from 83 °C → 72 °C over ≈ 28 minutes. No thermal-throttle events.
curl -O https://thermal.zctechnologies.org/dl/zc-thermal-control.tar.gz tar xzf zc-thermal-control.tar.gz cd zc-thermal-control sudo bash install.sh
Writes /usr/local/bin/zc-thermal-control, a systemd unit, and a scoped sudoers rule for nvidia-smi -lgc. SHA-256 of the bundle: 781907430427eb6d90877b76999fa194da82434a0c427af82553107f18aee8fd.
| Flag | Default | Meaning |
|---|---|---|
| -hot | 78 °C | Step clock down at or above this |
| -cool | 72 °C | Step up after 3 consecutive samples at or below this |
| -floor | 1800 MHz | Minimum max-clock during cooling |
| -ceil | 3000 MHz | Maximum max-clock when relaxed |
| -step | 150 MHz | Adjustment per sample |
| -interval | 30 s | Sampling period |
Not an undervolt tool. Not a fan-curve controller (GB10 does not expose fan PWM). Not a workload scheduler. The single actuator is nvidia-smi --lock-gpu-clocks; the controller chooses when and how much.
nvidia-smi against the local driver.