Energy Is All We Need: Beyond FLOPs in Model-Heterogeneous Federated Learning

Most energy-efficient federated learning (FL) methods report FLOPs or parameter reductions as evidence of energy savings. However, real GPU energy consumption differs across model architectures in ways that these computational proxies do not capture. In this work, we conduct an empirical study on reducing real GPU energy consumption in model-heterogeneous FL, where clients maintain distinct architectures and coordinate solely through shared prototypes. Through systematic experimentation with hardware-level energy measurements, we propose and validate two complementary mechanisms to make model-heterogeneous FL efficient in terms of real energy. First, Prototype-Aware Structured Pruning (PASP) evaluates each channel by the product of its batch-normalization scale magnitude and its gradient with respect to the prototype alignment loss---rather than the task loss---preserving channels critical for cross-client feature alignment while physically removing the rest. Second, Energy-Aware Training Scheduling (EATS) adaptively allocates local epochs to each client based on epoch-wise prototype alignment quality per unit energy, using smoothed alignment--energy profiles with a knee-point cutoff to prevent energy waste on epochs that yield diminishing alignment gains or amplify local drift. We investigate each mechanism's impact on accuracy and real energy consumption, as well as explore the impact of using both mechanisms simultaneously. All energy figures are obtained from GPU power monitoring rather than computational proxies, providing an empirically grounded evaluation of energy efficiency in heterogeneous federations.