Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Standardized Methods and Recommendations for Green Federated Learning

Austin Tapp, Holger R. Roth, Ziyue Xu, Abhijeet Parida, Hareem Nisar, Marius George Linguraru

TL;DR

This paper tackles the lack of standardized carbon accounting in federated learning by introducing a pragmatic, phase-aware methodology that combines CodeCarbon with NVFlare. It defines clear measurement boundaries for compute, idle overhead, and communication emissions, and includes a simple network-energy model to estimate model-update transmission costs. The approach is validated on CIFAR-10 and multi-site retinal segmentation under controlled efficiency tiers, revealing substantial emissions differences driven by idle time, coordination overhead, and hardware/grid variability. By providing an open-source instrumentation layer and minimal reporting fields, the work enables reproducible green-FL evaluations and a path toward more energy-conscious distributed learning practices.

Abstract

Federated learning (FL) enables collaborative model training over privacy-sensitive, distributed data, but its environmental impact is difficult to compare across studies due to inconsistent measurement boundaries and heterogeneous reporting. We present a practical carbon-accounting methodology for FL CO2e tracking using NVIDIA NVFlare and CodeCarbon for explicit, phase-aware tasks (initialization, per-round training, evaluation, and idle/coordination). To capture non-compute effects, we additionally estimate communication emissions from transmitted model-update sizes under a network-configurable energy model. We validate the proposed approach on two representative workloads: CIFAR-10 image classification and retinal optic disk segmentation. In CIFAR-10, controlled client-efficiency scenarios show that system-level slowdowns and coordination effects can contribute meaningfully to carbon footprint under an otherwise fixed FL protocol, increasing total CO2e by 8.34x (medium) and 21.73x (low) relative to the high-efficiency baseline. In retinal segmentation, swapping GPU tiers (H100 vs.\ V100) yields a consistent 1.7x runtime gap (290 vs. 503 minutes) while producing non-uniform changes in total energy and CO2e across sites, underscoring the need for per-site and per-round reporting. Overall, our results support a standardized carbon accounting method that acts as a prerequisite for reproducible 'green' FL evaluation. Our code is available at https://github.com/Pediatric-Accelerated-Intelligence-Lab/carbon_footprint.

Standardized Methods and Recommendations for Green Federated Learning

TL;DR

This paper tackles the lack of standardized carbon accounting in federated learning by introducing a pragmatic, phase-aware methodology that combines CodeCarbon with NVFlare. It defines clear measurement boundaries for compute, idle overhead, and communication emissions, and includes a simple network-energy model to estimate model-update transmission costs. The approach is validated on CIFAR-10 and multi-site retinal segmentation under controlled efficiency tiers, revealing substantial emissions differences driven by idle time, coordination overhead, and hardware/grid variability. By providing an open-source instrumentation layer and minimal reporting fields, the work enables reproducible green-FL evaluations and a path toward more energy-conscious distributed learning practices.

Abstract

Federated learning (FL) enables collaborative model training over privacy-sensitive, distributed data, but its environmental impact is difficult to compare across studies due to inconsistent measurement boundaries and heterogeneous reporting. We present a practical carbon-accounting methodology for FL CO2e tracking using NVIDIA NVFlare and CodeCarbon for explicit, phase-aware tasks (initialization, per-round training, evaluation, and idle/coordination). To capture non-compute effects, we additionally estimate communication emissions from transmitted model-update sizes under a network-configurable energy model. We validate the proposed approach on two representative workloads: CIFAR-10 image classification and retinal optic disk segmentation. In CIFAR-10, controlled client-efficiency scenarios show that system-level slowdowns and coordination effects can contribute meaningfully to carbon footprint under an otherwise fixed FL protocol, increasing total CO2e by 8.34x (medium) and 21.73x (low) relative to the high-efficiency baseline. In retinal segmentation, swapping GPU tiers (H100 vs.\ V100) yields a consistent 1.7x runtime gap (290 vs. 503 minutes) while producing non-uniform changes in total energy and CO2e across sites, underscoring the need for per-site and per-round reporting. Overall, our results support a standardized carbon accounting method that acts as a prerequisite for reproducible 'green' FL evaluation. Our code is available at https://github.com/Pediatric-Accelerated-Intelligence-Lab/carbon_footprint.
Paper Structure (19 sections, 1 equation, 4 figures, 2 tables)

This paper contains 19 sections, 1 equation, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Carbon Emission Measurement
  4. Federated Learning Experiments
  5. Federated orchestration.
  6. Client-side emissions tracking.
  7. cuDNN initialization.
  8. CIFAR-10 Classification
  9. Non-IID partitioning.
  10. Model and training loop.
  11. Accuracy target.
  12. Retinal Optic Disk Segmentation
  13. Communication Emissions Estimation
  14. Efficiency Tests (High / Medium / Low)
  15. Results
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Mean emissions over time by client-performance efficiency tier (high, medium, low) for the CIFAR-10 workload.
  • Figure 2: Cumulative (aggregate) CO$_2$e over time by site for the retinal segmentation task (H100 vs. V100).
  • Figure 3: Energy vs. emissions under alternative grid carbon intensities (CI), illustrating how identical energy use leads to different CO$_2$e outcomes depending on location.
  • Figure 4: Global variation in grid carbon intensity, motivating location-aware and time-aware carbon accounting. Source: Electricity Maps live map.