By name: strix-halo-monitoring description: Monitor CPU/GPU temps, power draw, and clock speeds on Bosgame M5 (Strix Halo APU gfx1151) during inference. Includes hwmon sensor map, monitor script, and power safety guidelines. version: 1.0
Strix Halo APU Thermal & Power Monitoring
Monitor CPU/GPU temps, power draw, and clock speeds on Bosgame M5 (Strix Halo gfx1151, 128GB unified) during inference workloads.
Context
The Strix Halo APU can pull significant wattage under sustained GPU load. In performance BIOS mode, running large models (Qwen 122B, GLM 4.7 Flash) for 30+ minutes on a shared 16A circuit tripped a home breaker. Monitoring power draw in real-time is essential to avoid this.
Hardware Sensor Map (hwmon)
| Sensor | hwmon | Key readings |
|---|---|---|
| k10temp (CPU) | hwmon2 | temp1_input (Tctl) |
| amdgpu (GPU) | hwmon5 | temp1_input (edge), power1_input (PPT mW), freq1_input (sclk Hz) |
| acpitz (ACPI) | hwmon0 | temp1_input |
| nvme | hwmon1 | temp1-4_input (Sensor 2 runs hot ~80C, crit 89.8C) |
| mt7925 (WiFi) | hwmon4 | temp1_input |
| r8169 (NIC) | hwmon3 | temp1_input |
Fan control is exposed via ec-su_axb35 kernel module at /sys/devices/virtual/ec_su_axb35/. Three fans with RPM, level, mode (curve/manual), and ramp curves. Also provides EC temperature and APU power mode.
ec-su_axb35 Sensor Map
| Path | Reading | Notes |
|---|---|---|
temp1/temp |
EC temperature (°C, integer) | Min/max in temp1/min, temp1/max |
fan{1,2,3}/rpm |
Fan RPM (integer) | 0 = off/stopped |
fan{1,2,3}/level |
Fan level (integer) | 0-based |
fan{1,2,3}/mode |
Fan mode (string) | "curve" or "manual" |
fan{1,2,3}/rampup_curve |
Temp thresholds to ramp up | e.g. "60,70,80,88,95" |
fan{1,2,3}/rampdown_curve |
Temp thresholds to ramp down | e.g. "50,60,70,78,85" |
apu/power_mode |
APU power mode | "balanced" or "performance" |
Tools
lm-sensorspackage providessensorscommand -- shows all readings formatted- Custom monitor script at
~/llm-server/monitor.sh-- logs CSV + live terminal output- Usage:
~/llm-server/monitor.sh [interval_sec] [logfile] - Default: 2s interval, auto-named log with timestamp
- Logs: timestamp, CPU C, GPU C, GPU watts, GPU MHz, NVMe temps
- Usage:
Power Safety Guidelines
- 16A circuit at 230V = 3,680W max total
- Strix Halo sustained full load: 120-150W+ (check with power1_input during inference)
- BIOS performance mode raises PPT limits = higher sustained draw = breaker risk on shared circuits
- Recommendation: use balanced/default BIOS mode for inference; move other devices to separate circuits for heavy workloads
Pitfalls
sensors-detectis not needed -- all sensors auto-detected via k10temp and amdgpu driverspower1_inputreports microwatts (µW); divide by 1,000,000 for wattsfreq1_inputreports Hz; divide by 1,000,000 for MHz- NVMe Sensor 2 (temp3_input) runs ~80C at idle -- normal for this platform but watch near 89.8C crit threshold