AI and environment (1/3): alarming numbers

By The GenerIA Team  |  August 6, 2024  |  9 m.

Share this article:

Insatiable for energy and a major producer of CO2, conventional artificial intelligence looks more and more like an environmental dead end. Is there any hope of sustainability? Everywhere, the numbers suggest otherwise...

Today's AI is voracious. Everywhere in the Western world, its energy consumption and carbon footprint are increasing significantly. The industry's giants, whose efforts at transparency can be applauded, are clearly the main culprits. Google, for example, has increased its CO2 emissions by 48% in 2023 compared with 2019. Logically, according to their annual Sustainability reports, Microsoft, Meta and Amazon face similar challenges.

This general trend is cause for concern. EPRI (Electric Power Research Institute), a non-profit energy research organization, estimates the electrical cost of a single ChatGPT query at around 2.9 Watts/hour, compared with 0.3 W/h for a standard Google or Bing search, i.e. around 10 times more. Given the global surge in AI, MIT estimates that a data center today consumes on average as much electricity as 50,000 American homes. By extrapolation, their study estimates that “the cloud” now emits as much CO2 as the entire global airline industry.

But it's when these figures are compared with everyday values that the problem reveals its true scale. A round trip between New York City and San Francisco generates 0.9 tonnes of CO2 equivalent (eqCO2) per passenger. A year of human life, 4.5 tons (16.4 tons in the USA). For an American car with all its fuel over its entire life cycle, that's 57 tonnes. But for a modest Transformer model with 213M parameters, as used for a long time by many organizations without any particular optimization of the training and inference machines, we reach 284 tonnes eqCO2, again over the whole life cycle.

Generative AI on trial

As the previous example implicitly shows, most of the blame lies with generative AI. Let's take the example of GPT-3, a model whose age allows analysts a certain hindsight. According to Goldman Sachs Research, its training alone has released 552 tonnes of CO2 into the atmosphere.

In use, several studies come to the same conclusion: it requires around 4 W/h to generate a page of content. OpenAI does not disclose the number of daily requests at the time when GPT-3 was ChatGPT's main engine. But one can always infer.

AWS estimates that 90% of the energy cost of a large language model comes from its use, versus 10% for its training. Schneider Electric, for its part, sees this split at 80/20 on average, given the evolution in model size. Finally, according to Meta, this same split varies according to use cases, with around 65/35 for standard LLMs and up to 50/50 for more “recommendation”-oriented models, whose parameters need to be updated frequently due to the constant arrival of new data. Averaging these averages, we can therefore say (with caution!) that GPT3 has already contributed 1,800 tonnes to the CO2 that surrounds us.

Scaling up

GPT-3 was 175 billion parameters. GPT-4 is 1,700 billion parameters (estimated), hence its better contextual understanding and the overall coherence of its answers, which we all noticed when it was released. Everyone is free to derive figures from this new order of magnitude, but be warned: it's not that simple.

While it is well established that the environmental cost of training increases in proportion to the size of the models, this cost exploded when a large number of additional tuning steps were required to gain a few percent in accuracy. Neural architecture search stages, in particular, are all the more energy-intensive as they are essentially based on trial-and-error approaches. Their benefits in terms of performance are not always immense, but in the battle for the top positions on leaderboards that research teams are waging, they remain unavoidable.

To diversify our examples and put things in perspective, here's a table comparing the environmental costs of training a few well-known models, since the inception of generative AI:

The figures in the Params. column are expressed in billions of parameters. The Chips column indicates the type and number of processors used to train the model. The Hours column represents the number of training hours. The Energy column represents the electricity consumed by training, in MW/h. The CO2 column represents estimated training-related emissions only, in metric tons.

It doesn't matter whether CO2 estimation methodologies differ, or whether the processors used are more or less optimized (NVIDIA V100 or A100, TPU v3 or v4...). This is one of the major problems with artificial intelligence as it is invented today. Statistically, the process of designing and testing a model with a view to scientific publication requires about 4,700 successive training sessions, often for a limited period of relevance. If proof were still needed, reading these publications one after the other, one does not get the impression that the majority of researchers are overly concerned about the ecological consequences of their work...

Glimmers of hope

In the face of such heavy and negative trends, is there any cause for optimism? Efforts are being made on all fronts to improve system efficiency. Between 2010 and 2018, there was a 550% increase in computing instances and a 2,400% increase in storage capacity in the world's data centers, while energy consumption rose by just 6%. Admittedly, these figures predate the advent of generative AI and the ubiquity of gluttonous GPUs, but efficiency gains from improved hardware, virtualization and data center design are also growing.

In addition, low-emission sources (in other words, anything other than fossil fuels) are expected to account for almost half of the world's electricity generation by 2026, compared with less than 40% in 2023. Some studies conclude that the overall volume of CO2 emissions linked to AI should decrease by around 3.5% per year until 2026, and between 4 and 4.5% until 2030. Others, however, see them doubling over the same period, given the numerous reactivations of gas and coal-fired power plants to meet immediate demand.

Finally, although this article is not the place to extol the virtues of GenerIA, we can also hope that consumers of inferences, both private and professional, will become aware of the problem and prefer frugal or even carbon-neutral AIs, which fully satisfy their needs while respecting a strict ethic of preserving natural resources.

Let's face it, it's the arms race that's responsible for the environmental mess we're in. No single model, however broad, can meet everyone's requirements. For the general public, no problem. Standard general-purpose AI, which is virtually free, is an undeniable vector of efficiency and freedom. In the enterprise, on the other hand, it is inapplicable for a number of reasons: poor adaptability to specific data and vocabularies, leak or predation risks on sensitive data, difficulties integrating with existing applications, lack of dedicated support and, more generally, strategic control...

By constantly refining their “big” models, the “big” players are aiming first and foremost for General Artificial Intelligence, synonymous not only with a cognitive big bang but also with economic and geostrategic domination. These objectives are widely debated but that's not what's going to stop them. On the other hand, more and more authorities (China, Singapore, Germany, Ireland, the city of Amsterdam...) are imposing restrictions on the creation of new data centers or computing facilities. Several projects have been abandoned as a result. In Ireland, for example, which benefits from a low tax rate and high-capacity submarine cables, several operators have been refused building permits. And while Ireland already allocates 17% of its electricity consumption to data centers, this figure is set to rise to 32% by 2027, given the number of new sites already approved.

Conclusion

Although the figures are alarming, progress in energy efficiency, the growing adoption of renewable energies, administrative control of resources and the operational efficiency of frugal AI offer encouraging prospects. It is crucial, however, that technology companies, policymakers and consumers work together to put an end to the current environmental havoc. It is clearly imperative to rethink AI if we hope to make it truly sustainable.

References

Google's 2024 Environmental Report

Microsoft's 2024 Environmental Sustainability Report

Meta's 2023 Sustainability Report

EPRI: Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption

Center for Data Innovation: Rethinking Concerns About Energy Use

MIT Technology Review: Training a single AI model can emit as much carbon as five cars in their lifetimes

Columbia University: Projecting the Electricity Demand Growth of Generative AI Large Language Models in the US

Goldman Sachs: AI is poised to drive 160% increase in data center power demand

Sources

[1] Strubell, Ganesh, McCallum, Energy and Policy Considerations for Deep Learning in NLP

[2] Facebook Research, Llama 2 Model Details

[3] Luccioni, Viguier, Ligozat, Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model

[4] Chowdery et al., PaLM: Scaling Language Modeling with Pathways

[5] Patterson et al., Carbon Emissions and Large Neural Network Training