AI-Reference Modell for Energy- and Ressource Efficiency and its Industrial Application (KIRA)

The development of the reference model for energy- and resource-efficient AI is the main objective of the project. It provides a comprehensive framework to guide stakeholders throughout the lifecycle of AI to improve the sustainability of AI systems and minimize their environmental impact. In particular, the model aims to significantly reduce the carbon dioxide equivalent (CO2e) emissions of AI systems and conserve critical resources such as water, rare earth elements (REE), etc. The model is not only committed to resource efficiency, but also to the introduction of practices that reduce the environmental footprint of AI activities.

A literature review has produced numerous guidelines and methods for measuring and assessing the environmental impact of AI systems. However, these are often based on flawed assumptions or inaccurate data, which can lead to incorrect conclusions. The GrAIn Model addresses these problems. It is based on scientific principles and provides a comprehensive list of factors that influence the sustainability of AI technologies. These are then assigned criteria to enable the assessment of the environmental impact of AI, providing stakeholders with a more informed and effective basis for decision-making.

A prototype of the model (so far without a list of criteria) can be found at https://green-ai-model.github.io. An initial publication on the topic is available [WGB+23].

The “Software Energy and Resource EfficieNcy Analysis” (SERENA) method for evaluating the energy and resource consumption of software has been developed and continuously expanded since 2009 by the “Green Software Engineering” research group at Trier University of Applied Sciences, Environmental Campus Birkenfeld. The method is suitable for collecting and evaluating data on the resource and energy consumption of software and is the result of research work carried out as part of several research projects. The development of the measurement system and the implementation of energy and resource efficiency measurements and their evaluation with the analysis tool “Open Source Consumption Analysis and Reporting” (OSCAR) was extended to the application with AI-based systems as part of KIRA.

In principle, AI-based systems are also software systems. Therefore, many of the methods developed for measuring and evaluating their consumption, as well as the criteria for assessing their environmental impact, can also be applied to these systems. However, not all criteria, metrics and measures are applicable or useful to all types of software and some systems may require the development of additional metrics. Examples of such adaptations are the definition of the “useful work” of the systems for calculating energy efficiency and the addition of GPU metrics for AI-based systems. Furthermore, the measurement method needs to be adapted, e.g. with regard to usage scenarios or tool support, as AI systems require different approaches for logging, performance measurement and evaluation. Distributed AI systems may require different usage scenarios for the different types of clients and for embedded devices (AIoT) the automation of usage scenarios must be remotely controlled as they do not have a graphical user interface.

The methods are available at https://gitlab.rlp.net/green-software-engineering/serena nd https://gitlab.rlp.net/green-software-engineering/oscar.