CO2 tracker
CAFEIN CO₂ Tracker & Optimiser
Artificial Intelligence and big-data analytics are transforming industries, but they also bring a hidden cost to our planet. The massive energy demands of large centralized data centers — powering GPUs, CPUs, cooling systems, and network infrastructure — already account for 15% of the ICT sector’s total greenhouse gas emissions and around 0.3% of all global CO₂ emissions. With AI adoption accelerating, this share is projected to grow rapidly in the coming years.
At CAFEIN, we are building the tools to ensure that AI innovation is not just powerful, but also sustainable. Our Federated Learning CO₂ Tracker & Optimiser is a pioneering platform that measures, analyses, and helps reduce the environmental footprint of AI training.
How the Tracker Works
The CAFEIN CO₂ Tracker & Optimiser continuously collects and integrates:
- Hardware and system specifications — GPU/CPU energy draw, cooling requirements, data storage and transfer costs.
- Model characteristics — size, parameter count, number of training rounds, and compression strategies such as quantization and sparsification.
- Regional energy data — real-time carbon intensity (kg CO₂/kWh), energy source mix (solar, wind, coal, nuclear, etc.), and energy costs for each client location.
- Federation architecture details — centralized vs decentralized FL, communication efficiency (bits/Joule), and number of participating devices.
The tool estimates the cumulative carbon footprint of a Federated Learning deployment, taking into account multiple sites and their regional energy profiles. Results can be aggregated to provide an overall view of emissions and used to compare different training configurations.
Image 1 – Example view of the CAFEIN CO₂ Tracker & Optimiser dashboard, showing a simulated federation with clients across Europe, their energy sources, and estimated emissions.
From Measurement to Optimisation
Our research shows that optimising communication and computation strategies can drastically cut emissions.
- Quantization reduces the number of bits needed to represent model parameters.
- Sparsification sends only the most relevant parameters during training rounds.
- Architectural choices — such as consensus-driven decentralized training — can reduce network overhead when communication energy efficiency is low.
Image 2 – Impact of quantization (number of bits) and sparsification on validation accuracy, under different carbon footprint limits. Grouped curves indicate configurations leading to the same total CO₂ emissions.
Source: A Carbon Tracking Model for Federated Learning: Impact of Quantization and Sparsification (arXiv:2310.08087).
By fine-tuning these factors, researchers can reduce their Federated Learning carbon footprint by up to 50%, while maintaining or even improving model accuracy.
Impact and Vision
The CAFEIN-FL CO₂ Tracker & Optimiser is more than a measurement tool — it’s a decision-making engine for sustainable AI. It empowers researchers, enterprises, and policymakers to:
- Compare training strategies based on carbon cost.
- Select greener energy sources or relocate compute nodes to low-carbon regions.
- Set and monitor CO₂ reduction targets for AI projects.
- Report sustainability metrics transparently to stakeholders.
By making carbon tracking a first-class metric in AI design, we aim to help organisations align their technological ambitions with global climate goals — ensuring that the AI revolution is not just intelligent, but also environmentally responsible.
Collaborators
