RQ4 · sustained vs initial throughput

Thermal: does it throttle?

The sanctioned long run — a bounded ~1-hour soak on 4B (sdpa), sampling temperature and power throughout to surface any throttling.

Bottom line: gfx1151 sustains 4B LoRA training throughput indefinitely at these power/thermal levels — there is no thermal-throttle penalty to budget for. The steady-state numbers from the short A/B windows hold under a full hour of load.

The harness records the per-step throughput series and a background temp/power sampler (amdgpu hwmon) for the whole run. The run is time-boxed at 60 min; it stops cleanly with the partial curve captured (time_boxed), not lost as a bare timeout.