In order to more accurately identify its impact, we need the ability and knowledge to measure it. That is no small task. Moreover, there are as-yet-unresolved challenges in researching AI’s energy consumption. Generative AI sets itself apart from other digital services by the significantly greater computing power it requires: for instance, the training of models can take several weeks and involve tens of thousands of GPUs.
According to France’s electronic communications and postal regulatory authority (Autorité de régulation des communications électroniques et des postes — Arcep), digital technologies already account for around 4% of global greenhouse gas (GHG) emissions. The meteoric rise of AI is expected to cause that figure to double over the coming years. Data centres, which are digital factories in every sense of the word, are at the forefront in that regard: in the United States, they generated 105 million tonnes of CO2 equivalent in one year, with a carbon intensity 48% higher than the national average. Generative AI is a major factor behind the rise of these concrete behemoths that require a huge amount of electricity — contemporary generative AI models use much more energy than their simpler, earlier counterparts. In two decades, we have gone from simple regressions with around ten parameters to models with several thousand billion parameters. In monetary terms, it’s the equivalent of going from the price of a coffee to the GDP of France.
Assessing the environmental impact of AI is a complex task, but a crucial one for guiding innovation towards greater frugality
Life cycle assessments: the cornerstone of measuring AI’s environmental footprint
As with any product or service, a life cycle assessment ( ) is necessary to properly gauge the impact of AI. This process is standardised by under standards 14040 and 14044. The key feature of an LCA is tracing impacts over the entire lifespan of a product so as not to overlook significant impacts that occur outside of its actual use. Such impacts relate in particular to energy and water consumption—whether for storing the data needed to train models, or for actually training and deploying them—but also to the production of equipment (mining, manufacturing, transport etc.) needed for computation and storage. Direct energy consumption and the production of equipment both contribute to greenhouse gas emissions, with the production of equipment also depleting abiotic resources. So, the focus here will be on analysing the design of AI systems, their uses and their end-of-life. AI is characterised by extensive resource sharing across certain phases (data collection, data preprocessing, training, and inference), which complicates the matter of measuring the impact of each individual system. It is also characterised the diversity of its life cycles, meaning that each research project may differ greatly from the last.
For example, looking at the data collection phase: how should we assess the environmental footprint of a model like GPT-4, which uses data from across the Internet? What methods can we apply to accurately quantify its impact? Work is under way, particularly at the , to standardise considerations like these and make it possible to meticulously select a model based on its impact, and incorporate that impact into broader studies. But measuring the impact of a digital service like AI is not limited to looking solely at how it works. It must be possible to study the consequences of its use, be they positive or negative first order effects, as well as the rebound effects it can have. This is stipulated in ITU Recommendations L.1480 and L.1410.
Moreover, as with a life cycle assessment, the key data needed for these calculations are often missing or intertwined, thus complicating any measurements. Focusing on the classic AI issue, namely its energy consumption, this indicator is closely linked to others since the systems that require the fewest computations, and therefore the least energy, do not, ipso facto, require the purchase of new equipment or the construction of new data centres.
How can measurements be made precisely?
AI processes do not run “in a vacuum”; if they did, measuring their consumption would be simple. Typically, the code is distributed across several processes contained in a virtual machine and run on a physical machine with several / , which in turn sits in a server rack within a data centre. Even though there are tools for estimating the consumption of each of these various components more or less accurately, those currently available only offer approximations of consumption. As such, we need to fully understand each tool to know whether it increases or decreases consumption, and to know what correction factor to use to approximate the actual consumption (considering that, in an LCA, impacts must never be underestimated).
There are three broad categories of measurement:
- External measurements: power meters connected to the hardware.
- Energy profiling of components or algorithms.
- Integrated measurements using sensors or specific software tools (CPU, GPU etc.).
Each method has limitations, e.g. cost, precision, granularity. For example, an external meter cannot isolate the consumption of a specific algorithm, while software tools often rely on estimations.
However, there are several software tools that allow the GHG emissions linked to an AI system to be estimated:
- CodeCarbon: Python library that measures CPU, GPU and consumption and then converts it into CO2 emissions. However, it will estimate consumption to be lower than what it is in reality, and it does not enable measurement of consumption from calls to external models.
- ML CO2 Impact: estimates emissions based on the power of the components and the duration of use, taking into account the cloud provider and location.
- EcoLogits: assesses electricity consumption, GHG emissions, resource depletion and primary energy usage for the inference of large language models. However, this library tends to overestimate the consumption of closed models because it disregards the optimisations achieved by scaling them.
Although these tools are not perfect, they offer useful orders of magnitude for comparing models, infrastructures and algorithms, and shaping impact reduction efforts.
Towards more responsible AI
Assessing the environmental impact of AI is a complex task, but a crucial one for guiding innovation towards more responsible consumption. The choice of tools and methods depends on the aim: assessing the overall impact of a service or optimising a specific component. However, care must be taken not to shift the problem: reducing the impact of one element should not increase the impact of another.
There are three research challenges to overcome in relation to estimation of AI’s energy consumption:
- Defining standardised measurements for the energy consumption of different algorithms and services. This is because, without a common methodology, no comparison is genuinely feasible.
- Adapting these measurements as new AI methods emerge.
- Determining correlations between measurable variables (e.g. energy consumption, carbon footprint and greenhouse gases) and the main efforts being made in the political and industrial spheres.
AI, as a source of both challenges and opportunities, must now combine performance with environmental responsibility. That is a pre-requisite for it to contribute to a sustainable digital future.
