Source-linked AI summary
Quantifying the Carbon Emissions of Machine Learning
Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, Thomas Dandres
TL;DR
ML training emissions vary with server location and grid, training duration, hardware, and training practices, but their magnitude is difficult to assess. The paper introduces a Machine Learning Emissions Calculator using public data and offers actions for estimating and mitigating emissions. It concludes that these choices provide some control over direct emissions, while calculator estimates remain approximate because of load balancing and limited organizational transparency.
Problem
The environmental impact of ML training is difficult to assess because emissions depend on location, energy grids, training procedures, and hardware, while prior studies often focus on large fossil-fuel-powered models.
Method
The paper develops a Machine Learning Emissions Calculator using CO2eq and public data on hardware, provider-region grids, emissions per kWh, and offsets.
Results
The calculator estimates approximate training emissions from server region, GPU type, and training time, while the paper identifies location, experiment design, and hardware as actionable factors.
Takeaways & Limitations
Practitioners can use the calculator and select lower-carbon locations, efficient hardware, and less wasteful training practices to reduce direct ML-training emissions.
Takeaways & Limitations
Calculator outputs remain approximate because global load balancing can reduce aggregate gains and organizations do not transparently report exact energy use and emissions.
Abstract
from arXiv · showhide
From an environmental standpoint, there are a few crucial aspects of training a neural network that have a major impact on the quantity of carbon that it emits. These factors include: the location of the server used for training and the energy grid that it uses, the length of the training procedure, and even the make and model of hardware on which the training takes place. In order to approximate these emissions, we present our Machine Learning Emissions Calculator, a tool for our community to better understand the environmental impact of training ML models. We accompany this tool with an explanation of the factors cited above, as well as concrete actions that individual practitioners and organizations can take to mitigate their carbon emissions.
1 Introduction
As GPU-based ML training becomes more widespread and resource-intensive, its energy use and climate impact are growing. The paper introduces a calculator and practical guidance to estimate and reduce emissions from model training.
- 1 Introduction: Growing GPU accessibility and increasingly large, long training runs are raising the energy costs of ML research.Recent models often use more GPUs, larger datasets, and longer training procedures.
- 1 Introduction: The paper presents the Machine Learning Emissions Calculator to estimate carbon emissions produced by training ML models.It also explains emissions-related factors and recommends practices for researchers and organizations.
2 Quantifying Carbon Emissions in Neural Network Training
The paper quantifies training emissions with CO2-equivalents and examines how energy grids, computing infrastructure, training time, and training choices affect them. It combines public grid information with server data to support emissions estimation.
- 2 Quantifying Carbon Emissions in Neural Network Training: CO2-equivalents provide a standardized single measure for comparing greenhouse-gas emissions from ML training.The measure expresses different gases through their equivalent global-warming impact as CO2.
- 2 Quantifying Carbon Emissions in Neural Network Training: Estimating cloud-server emissions requires assumptions because providers rarely disclose the energy grid connected to a given server.The calculator assumes servers use local grids and combines public grid-emissions data with known GPU-server information.
- 2.2 Computing Infrastructure and Training Time: GPU performance increased from 100 Giga FLOPS per second in 2004 to up to 15 Tera FLOPS per second in recent hardware.Despite more capable hardware, deeper models are often trained on multiple GPUs for weeks or months, increasing energy requirements.
- 2.3 Training Procedures: Pre-trained models with task-specific fine-tuning have matched training from scratch while being more robust in image recognition and NLP.The paper also discusses random hyperparameter search as a way to avoid superfluous computing and energy costs.
3 ML Emissions Calculator and Actionable Items
The paper’s calculator estimates training emissions from server location, GPU type, and training time, while its guidance targets provider choice, infrastructure efficiency, wasted experiments, and hardware selection. These choices give practitioners some control over direct emissions, though practical and estimation limits remain.
- Quantify Your Emissions: The alpha calculator takes server region, GPU type, and training time as inputs and outputs approximate CO2eq emissions.Its data covers hardware energy use, provider-region location, regional emissions per kWh, and potential provider offsets.
- Quantify Your Emissions: The calculator uses publicly available, editable data and accompanies it with educational material and mitigation actions.The authors intend to update the data as more information becomes available.
- Choose Cloud Providers Wisely: Data-center PUE accounts for cooling, power conversion, and auxiliary energy, and can vary substantially across facilities.Google Cloud Services is given as an example with an average PUE of 1.1.
- Select Data Center Location: Selecting a low-carbon data-center location can substantially reduce direct training emissions compared with a fossil-fuel-powered server.The paper describes several hundreds of kilograms of CO2eq emitted when training on fossil fuels instead of hydroelectricity.
- Reduce Wasted Ressources: Random search can replace inefficient grid search and potentially accelerate hyperparameter search, reducing emissions from extended experiments.Testing, debugging, literature review, and noise analysis can also reduce failed or uninformative experiment cycles.
- Choose More Efficient Hardware: TPU 3 can be 4 to 8 times more efficient than GPUs, while CPUs can be 10 times less efficient than GPUs.For embedded applications, the Jetson AGX Xavier can be 10 to 20 times more efficient than traditional GPUs.
4 Discussion
The calculator provides an approximate starting point for estimating direct ML-training emissions, while practical constraints and methodological limitations restrict how precisely or broadly those estimates apply.
- Small choices in model training can affect direct carbon emissions, although server location and data requirements are not always controllable.Privacy considerations can constrain server location, while large datasets may be needed for robust models.
- Global load balancing means that concentrating workloads in low-carbon locations would produce a smaller global reduction than location-level estimates suggest.Servers in preferred locations could become saturated, requiring other servers to remain in use.
- The calculator has a large margin of error because organizations’ actual energy consumption and carbon emissions lack transparency.The authors use the best publicly available sources and remain open to additional data.
- The current tool focuses on training emissions, while continuous, large-scale inference can also consume substantial energy.Deployment emissions should therefore be considered in real-world products.
- The authors propose efficiency as an evaluation criterion and concrete emissions reductions while acknowledging broader debates about ML’s environmental value.They do not propose a complete solution but aim to support further discussion and positive changes.
- Empirical GFLOPS/W measurements would improve accuracy, but approximate values suffice for comparing device classes.
Appendix A: Energy Grid Data Used for the ML Emissions Calculator
The appendix presents a reduced version of the energy-grid data used by the calculator and records the database state at publication.
- The appendix contains fewer columns than the public database for clarity.The omitted columns include the data sources.
- The appendix data excludes source information that remains available in the public database.
- The public database’s commit hash at publication was e692e28.
Google Cloud Platform
The Google Cloud Platform data lists cloud regions alongside their countries, localities, and associated emissions values.
- Finland’s Hamina region is listed with an emissions value of 211.
- Switzerland’s Zürich region is listed with an emissions value of 16.
- India’s Mumbai region is listed with an emissions value of 920.
Amazon Web Services
The Amazon Web Services data lists regions by provider identifier, country, locality, and associated emissions values.
- Canada’s Montreal region, ca-central-1, is listed with an emissions value of 20.
- India’s Mumbai region, ap-south-1, is listed with an emissions value of 920.
- Australia’s Sydney region, ap-southeast-2, is listed with an emissions value of 802.
Microsoft Azure
Microsoft Azure entries span regions across Asia, the United States, Europe, South America, and Australia, with each location paired with a numeric value.
- Regional coverage: Azure entries include Hong Kong, Singapore, Osaka-shi, Tokyo, and multiple regions across the United States.The listed United States locations include Des Moines, Blue Ridge, Boydton, San Francisco, Chicago, and San Antonio.
- Regional coverage: European entries cover Dublin and Amsterdam, while additional entries identify São Paulo, Sydney, and Melbourne.