Energy measurement for edge inference
This study examines when operation counts cease to be reliable predictors of neural-network inference energy. The completed Raspberry Pi phase contains 258 valid trials across 43 models and two precisions. MAC count tracked broad energy scaling, but closely matched standard/depthwise pairs differed by 32.0%–49.5% in measured board-input energy. ESP32-S3 measurement is the next cross-platform phase.
Raspberry Pi experiment complete; microcontroller phase in progress.