A Framework for Energy and Carbon Footprint Analysis of Distributed and Federated Edge Learning

TL;DR

This paper presents an analytical framework to evaluate the environmental impact of federated learning compared to traditional centralized methods, highlighting significant energy savings and trade-offs in convergence speed.

Contribution

It introduces a comprehensive framework for analyzing energy and carbon footprints of federated learning, considering communication costs and decentralized policies, with real-world industrial evaluation.

Findings

Federated learning achieves 30-40% energy savings in wireless systems.

Consensus-driven FL reduces emissions further in mesh networks.

FL policies are slower to converge with uneven local data distribution.

Abstract

Recent advances in distributed learning raise environmental concerns due to the large energy needed to train and move data to/from data centers. Novel paradigms, such as federated learning (FL), are suitable for decentralized model training across devices or silos that simultaneously act as both data producers and learners. Unlike centralized learning (CL) techniques, relying on big-data fusion and analytics located in energy hungry data centers, in FL scenarios devices collaboratively train their models without sharing their private data. This article breaks down and analyzes the main factors that influence the environmental footprint of FL policies compared with classical CL/Big-Data algorithms running in data centers. The proposed analytical framework takes into account both learning and communication energy costs, as well as the carbon equivalent emissions; in addition, it models both vanilla and decentralized FL policies driven by consensus. The framework is evaluated in an industrial setting assuming a real-world robotized workplace. Results show that FL allows remarkable end-to-end energy savings (30%-40%) for wireless systems characterized by low bit/Joule efficiency (50 kbit/Joule or lower). Consensus-driven FL does not require the parameter server and further reduces emissions in mesh networks (200 kbit/Joule). On the other hand, all FL policies are slower to converge when local data are unevenly distributed (often 2x slower than CL). Energy footprint and learning loss can be traded off to optimize efficiency.