CodeCarbon is an open-source Python package that estimates the carbon footprint of computer programs by measuring the energy consumption and carbon emissions of hardware. It was created 5 years ago and has become the de-facto standard for measuring the carbon emissions of different kinds of computer code. AI researchers and data scientists have been among the main users of CodeCarbon, as they use it to estimate the carbon footprint of their AI models, but CodeCarbon is also used more broadly as it comes with an optional command line interface and does not require coding skills.
The workshop will cover the environmental impact of computing and AI, ranging from the electricity it uses and how it is generated and the growing environmental footprint of AI due to the growing size of AI models and the quantity of computation they require. We will do a quick overview of work in the field and talk about the factors that are at play when calculating the carbon footprint of a piece of code- namely, the hardware being used, the compute time, and how the electricity being used was generated.
We will follow up with an interactive demo of CodeCarbon, showing how it works in a Jupyter notebook and via a command line interface. We will prepare several pieces of code that participants can run, both locally on their laptops and via Google Colab in order to compare different types of hardware, such as GPUs and TPUs. We will also demonstrate how to use the dashboard to visualize emissions and how they compare to common equivalents, like driving a car and heating a house.
Finally, we will conclude with a set of best practices for reducing the carbon footprint of code - such as choosing a low-carbon compute server, doing experimentation before running compute-intensive processes, and optimizing scripts to make them more lightweight.
Via this session, we hope to inspire our audience to both quantify the carbon emissions of their code and take steps to reduce them, in order to contribute to a more sustainable computing future.