Could Model Partitioning Make Federated Learning More Sustainable?

Model partitioning and distributed training strategies can significantly enhance the sustainability of Federated Learning (FL), particularly in cross-silo and edge scenarios. Research indicates that while FL training energy is often comparable to Centralized Learning (CL), CL incurs higher total carbon emissions due to the substantial energy required for data transfer and storage across silos.

Key sustainability factors include:

• Reduced Data Redundancy: Cross-silo FL minimizes the need to replicate large datasets, thereby lowering the CO2 emissions associated with data centers and network transfer, which are often overlooked in traditional analyses.
• Optimized Communication: Techniques like Consensus-Driven Federated Averaging (CFA) eliminate the need for a central parameter server, reducing energy use in mesh networks and improving efficiency in low-power IoT environments.
• Adaptive Client Selection: Strategies such as EcoLearn and FedZero optimize carbon footprints by selecting clients based on high utility per unit of carbon or by leveraging renewable excess energy, potentially reducing emissions by up to 10.8x or achieving zero operational carbon.
• Local Epoch Tuning: Adjusting the number of local epochs can balance communication and computation costs; for instance, 5 local epochs often reduce total rounds and energy compared to 1 local epoch, though this depends heavily on data distribution (IID vs. non-IID).

Ultimately, sustainability in FL is not guaranteed by the paradigm itself but depends on system design, data distribution, and energy-aware scheduling to mitigate the higher communication overhead and slower convergence often seen in non-IID settings.

This material examines whether model partitioning can make federated learning (FL) more sustainable as the paradigm moves beyond low-power mobile devices toward cross-silo settings involving edge servers, data centres, and heterogeneous compute resources. As FL systems scale, their carbon footprint becomes a significant concern: many clients may be powered by high-carbon grid electricity, and training or inference workloads may run during periods of high grid demand. The paper frames sustainability not merely as an operational afterthought, but as a first-order design constraint, asking whether the structure of the federated training process itself can be reorganized to reduce emissions.

A central contribution is the analysis of model partitioning as a carbon-reduction mechanism. Rather than treating all FL participants as equivalent endpoints, the work considers how model components can be split across devices, edge nodes, or servers so that computation is placed where it is more energy-efficient or where the electricity supply is cleaner. This can allow training to be aligned with low-carbon energy availability or low-demand grid periods, while reducing the energy burden on high-carbon client devices. The key insight is that sustainability in FL may depend not only on when training occurs, but also on where different parts of the model are trained and how much work is delegated to different tiers of the infrastructure.

The material matters because it connects two increasingly important concerns in distributed machine learning: system efficiency and environmental impact. As FL is adopted in more compute-intensive and infrastructure-heavy deployments, carbon-aware design becomes essential for both practical deployment and responsible AI practice. By exploring model partitioning through a sustainability lens, the work suggests a path toward FL systems that optimize not only accuracy and communication cost, but also energy consumption and grid impact, making federated learning more viable in real-world, carbon-constrained environments.