ML CO2 Impact
Machine Learning has a carbon footprint.
We've made a tool to help you
estimate yours:
Compute your GPU's carbon emissions
Push for more transparency in our field by including the results in your publication (research paper, blog post etc.)
Install codecarbon to integrate carbon estimations in your Python workfow.
Publish
We believe that an important step towards reducing carbon emissions is the generalization of emissions reporting in papers, blog posts and publications in general.
To that end, here is a LateX template you can use in your publication to report the emissions you have calculated with the Calculator:
Social Media
Learn
This section will present some of the important concepts to better understand the carbon footprint of your ML research, including what providers do to help you offset it
CO2
At the center of the climate crisis is a commonplace but very important concept: that of carbon dioxide (CO2), low amounts of which occur naturally in the Earth's atmosphere, but its concentration has been rapidly increasing due to human activity. This increase is dangerous because of CO2'effect as a greenhouse gas, meaning that it absorbs and emits infrared radiation in the wavelength range emitted by the Earth, which contributes to the global warming of the planet ( IPCC Glossary ).
CO2 eq.
Since other gases such as methane, nitrous oxide or even water vapor also have this warming effect, a standardized measure for describing how much warming a given amount of gas will have is often provided in CO2-equivalents (CO2eq), for simplification purposes. For instance, the carbon intensity of transportation is measured in grams of CO2-equivalent per person-km, whereas the carbon intensity of energy is measured in grams of CO2-equivalent per kilowatt hour.
Renewable vs. Nonrenewable Energy
A key distinction to be made is that between renewable and nonrenewable energy sources. Renewable energy is collected from sources which are naturally replenished on a human timescale: this includes wind, sunlight, tides, geothermal heat, etc. In contrast, nonrenewable energy sources such as coal and petrol are not naturally replenished after their usage, and can take millions of years to be formed again. Finally, not all renewable energy sources are carbon-free, since the burning of biomass (such as plants or algae) still produces CO2-equivalents.
Carbon Offsetting
Carbon offsetting is a reduction in emissions of CO2-equivalents that is made in order to compensate for the emissions made by another actor. The money paid via carbon offsetting is invested towards different types of projects, including renewable energy or energy efficiency, carbon sequestration, methane abatement, among others. These projects are run by private companies, more often than not in a different place than where the offsetting is carried out. A large part of carbon offsetting goes towards funding renewable energy projects such as building wind or solar farms, hydroelectric dams, and extracting biofuel, in the hopes of reducing the cost of renewable production.
Renewable Energy Credits
An indirect form of offsetting that is often used by large companies is the purchasing of Renewable Energy Credits (RECs), which involves directly purchasing quantities of energy produced by renewable sources. This is considered to be an indirect of offsetting because in order to convert RECs into offsets, the energy purchased is translated into carbon reductions under the assumption that the clean energy is displacing an equivalent amount of electricity produced by non-renewable methods. Other commonly used forms of carbon offsetting include reforestation, sequestration, replacement of legacy equipment such as coal-burning stoves by more modern ones, etc.
Carbon Neutral
Carbon neutral is a term used to indicate a net zero carbon footprint of an individual or organization. While a small part of carbon-neutral entities rely entirely on renewable energy that is zero-carbon, the large majority of entities balance their emissions of CO2-equivalents, using approaches such as carbon offsetting or the purchase of RECs. Many major technology companies such as Google and Microsoft are carbon-neutral, matching 100% of their electricity use with renewable energy purposes (Google, 2018; Microsoft, 2018). This does not mean, however, that their operations do not produce CO2-equivalents, since this is dependent on the energy sources of the locations in which their infrastructure is located.
Featured Maps
Here are a few maps from Our World In Data which illustrate the complexity of carbon emissions mitigations. For a complete analysis, visit their webstie
Carbon Efficiency
Carbon efficiency: the amount of CO2 emitted per unit energy (grams of CO2 emitted per kilowatt-hour). This is largely related to a country�s energy mix. An economy powered by coal-fired energy will produce higher CO2 emissions per unit of energy versus an energy system with a high percentage of renewable energy. As economies increase their share of renewable capacity, efficiency improves and the amount of CO2 emitted per unit energy falls.
There are things you can do
With great computing power comes great responsibility
Choose your cloud provider wisely
Choose your region
Buy Carbon Offsets
Don't Do Grid Search
Choose Clean Energy
Push for more transparency
Publish
Fill Out Our Survey
Survey
About ML CO2 Impact
Progress in Machine Learning (ML) in recent years has been meteorical, with major breakthroughs happening in domains such as machine translation, image recognition and generation. While these advances are having widespread applications in many domains, detecting cancers on X-rays and improving the prediction of supply and demand,the carbon impact of ML training has not been a central part of the conversation until recently.
Therefore, as ML practitioners who are also aware of the overall state of the environment, we find that it is important to develop this conversion further and work towards building the tools we need to assess the carbon emissions generated by the models we train, as well as ways to reduce those emissions.
Read our paper
We have published this short paper to the Climate Change AI workshop at NeurIPS 2019.
Look at the Data
The data is hosted on Github to allow for external contributions
The sources are available in the csv file above. Use issues and pull-requests to suggest modifications/additions. The data used for the providers' offset ratio (how much of their emissions they offset) comes from these sources: Google, Microsoft, Amazon. For the latter we assumed non-carbon neutral regions were not offset at all.
We have looked for data about the main cloud providers. If you'd like to contribute and add more providers, please follow these guidelines.
Ongoing work: how to help
We're currently working towards enabling you to compute your emissions when using a private infrastructure. This would require either mapping cities to carbon emissions of local grids or expecting our users (you) to input such a value. The mapping does not (?) exist for now (and for free), and we feel like most users would not know the carbon efficiency of their electricity grid. This is why this is not yet available.
Any help with this issue is welcome, please get in touch.
If you wish to add a cloud provider to our list, here's how.
Generated Latex Template
\usepackage{hyperref} \subsection{CO2 Emission Related to Experiments} Experiments were conducted using {cloud provider} in region {region of compute}, which has a carbon efficiency of kgCO$_2$eq/kWh. A cumulative of {hours used} hours of computation was performed on hardware of type {hardware type} (TDP of W). Total emissions are estimated to be {emission} kgCO$_2$eq of which {percentage offset} percents were directly offset by the cloud provider. %Uncomment if you bought additional offsets: %XX kg CO2eq were manually offset through \href{link}{Offset Provider}. Estimations were conducted using the \href{https://mlco2.github.io/impact#compute}{MachineLearning Impact calculator} presented in \cite{lacoste2019quantifying}. @article{lacoste2019quantifying, title={Quantifying the Carbon Emissions of Machine Learning}, author={Lacoste, Alexandre and Luccioni, Alexandra and Schmidt, Victor and Dandres, Thomas}, journal={arXiv preprint arXiv:1910.09700}, year={2019} }