Title: IEEE P7100 Standard for Measurement of Environmental Impacts of Artificial Intelligence Systems

Scope: This standard defines a measurement framework for reporting on environmental indicators for training models and deriving inference on Artificial Intelligence (AI) systems. This includes harmonized measurements of compute intensity (e.g. energy use) with associated environmental impacts (e.g. carbon dioxide (CO2) emissions, or water consumption). This standard describes methodologies to separate the measurement of AI-specific compute (i.e. data centres used for AI training or inference) from general purpose compute (e.g. data centres used for other purposes like cloud services).

Need for the Standard: The arrival of Generative AI models and the proliferation of AI applications across industries has increased the demand for AI compute resources by several orders of magnitude. The associated increase in energy consumption, carbon emissions, water use, and other lifecycle impacts, especially in the training and inference phases, has made it imperative to measure and mitigate the environmental impacts of AI models. Given the exponential nature of AI compute requirements and the difficulty of separating general-purpose compute from AI-specific environmental impacts, many governments, international organisations and technology providers have recognised the need for a dedicated and standardised reporting framework to estimate AI-related sustainability impacts.

The compute needs of advanced AI systems have grown significantly over the past decade. The widespread introduction of generative AI tools has further increased the demand for AI compute, in particular moving focus from AI training to AI inference. As generative AI becomes more widely used, policy makers and technology providers need accurate measures of their compute intensity. Some experts anticipate that generative AI will increase the demand for AI compute by several magnitudes, alongside increased energy use, carbon emissions, and water consumption.

There is currently a significant and well-documented gap in publicly available data on the environmental impacts of AI such as energy use, emissions, water consumption, and other lifecycle impacts. This includes lack of harmonised measurement standards, limited availability of multi-level data, and the challenge of separating AI-specific measurement from general purpose compute.

Developing a standard to measure AI’s environmental impacts would help assist in the implementation and potential enforcement of AI regulations and intergovernmental standards currently emerging around the world.

Participate in the IEEE P7100 EIAI Working Group:

Please subscribe to the IEEE P7100 EIAI Working Group public e-mail reflector and register an interest in the IEEE SA myProject system.