Issue #1 | Summer 2022
Tools for Measuring CO2 Consumption of AI Applications
Over their life cycles, AI systems cause both direct and indirect CO2 emissions. On the one hand, greenhouse gases are generated by the hardware required during raw material extrac- tion, production, transport and disposal. These emissions, however, cannot be precisely quan- tified in most cases due to a lack of informationavailable and the complexity of their supply chains. It is also difficult to directly link them to specific AI models. Most of the CO2 emissions caused by AI systems are the product of the electricity consumed by the hardware when the systems are developed and deployed. The extent of the emissions depends mainly on two factors: the amount of electricity consumed by the hardware and the CO2 intensity of the underlying power mix. Some tools can be inte- grated into the source code of the applications (CodeCarbon, experiment-impact-tracker, carbontracker) to determine the emissionvalue. It is also possible to make projections regarding the expected emissions by considering the system specifications (with the Machine Learning Emissions Calculator or with carbontracker, for example).