Record: Multi-GPU enumeration and reporting — 2026-08-23

Outcome

Extended the service to enumerate all AMD GPUs and control every MI25 found.

Changes:

gpu_interface now enumerates every AMD card under /sys/class/drm/card*/device/ and reports which are controllable (MI25 + amdgpu + usable hwmon).
run_check() prints one line per AMD card: controllable or not, PCI address, vendor, device, driver, hwmon path.
run_service() controls every controllable MI25, not just the first.
The control loop carries per-card state (pwm_last, pwm_want, temps, report cadence) and labels every report line with the PCI address.

A card is controllable when device is 0x6860 (MI25), driver is amdgpu, and the hwmon has pwm1_enable, pwm1, and the three temp inputs.
Non-MI25 AMD cards are reported but not controlled.
The feature requirement now states the service controls every MI25 found.

Two-MI25 fixture: both enumerated, both reported controllable, both ramped in the loop (verified 2026-08-23).
Non-MI25 AMD fixture card reported not-controllable.
Removing pwm1_enable flips the report to not-controllable.
make build, make test, ctest, clang-format all pass.