Energy Allocation Between Computation and Wi-Fi Transmission in TinyML Federated Learning on IoT Devices: A Measured Crossover

Abstract Energy models for federated learning (FL) on battery-powered Internet of Things (IoT) devices often assume that radio transmission costs far more than on-device computation. We test this assumption on a Wi-Fi microcontroller: an ESP32-S3 trains the output head of a keyword-spotting model and exchanges its updates over 2.4 GHz Wi-Fi, while an independent power monitor and hardware phase markers separate the energy of computation from that of communication. Across 105 runs, computation costs 1.00 mJ per local epoch with no detectable fixed overhead, and communication costs 7.4 to 9.7 mJ per client turn in two sessions, regardless of the number of epochs. The two are therefore equal at 7.4 to 9.7 local epochs. Training started immediately after a radio burst absorbs about 0.5 mJ of tail energy, which a short pause removes. Shannon-capacity and switching-energy models underestimate the measured energies by 95% to 99% at their nominal parameters, because they omit protocol overhead and memory activity, and neither contains the power drawn while the device is awake. We show that sleeping between rounds moves the crossover toward the ratio of exchange time to epoch time, which yields 11 to 19 epochs with sleep powers measured on the same board. In deep sleep the board draws 17.2 mW, but each wake-up, dominated by Wi-Fi re-association, costs 0.68 J, 35 times the marginal energy of a ten-epoch turn. From these results, we derive a measurement-calibrated procedure for selecting the number of local epochs and design guidelines for energy-constrained IoT clients.