Standardized Methods and Recommendations for Green Federated Learning

The article introduces a standardized carbon-accounting methodology for federated learning (FL) that enables tracking of CO2 emissions during collaborative model training. Utilizing NVIDIA NVFlare and CodeCarbon, the method accounts for phase-aware tasks and communication emissions, revealing significant carbon footprint variations based on system efficiency and hardware configurations, such as a 1.7x runtime difference between H100 and V100 GPUs. This framework is crucial for practitioners aiming to evaluate and mitigate the environmental impact of FL systems, promoting reproducibility in green AI practices.