What are the climate costs and benefits of machine learning?
Training for a large language model like ChatGPT-3 creates around 500 tonnes of greenhouse gas pollution due to massive computers chewing up gigawatt hours of energy.
Experts say even using generative artificial intelligence to respond to user prompts requires orders of magnitude more power than a Google search.
So while machine learning can be used for tasks like climate modelling and integrating higher shares of renewable electricity into the grid, researchers say the use of the technology comes with a climate cost.
The issue raises questions about the appropriate uses for the technology and the potential for smaller, more efficient models capable of delivering the same performance using less power.
Professor Simon Lucey, Director of the Australian Institute for Machine Learning at the University of Adelaide, describes the environmental impact of machine learning as a “double-sided coin”.
“There’s a lot of public good. But also, there are ways it could be worryingly used.”
Lucey says the issue of machine learning’s energy consumption is similar to concerns often raised about cryptocurrencies.
“Essentially you have these massive computers that are chewing up gigawatts of energy, and potentially – depending on where they’re sourcing the energy – leading to a lot of carbon in the atmosphere,” he explains.
For generative AI, like large language models, it’s the training phase that uses the most energy, because the system is trawling through data. But the system still uses a lot of energy when you interact with it, Lucey says.
Training and interacting with machine learning models is energy-hungry and emissions-intensive
Stanford University’s AI Index Report 2023 says greenhouse gas pollution emitted by machine learning models depends on factors such as the number of model parameters (data points), the energy efficiency of data centres, and the source of electricity to those centres.
In general, the larger and more complex the model and the dirtier the electricity used to power computers and data centres, the higher the resulting greenhouse gas emissions.
The report cites an analysis comparing the emissions from training four large language models (Gopher, BLOOM, GPT-3 and OPT) by research scientist and climate lead at Hugging Face, Sasha Luccioni.
GPT-3 released the most emissions (502 tonnes CO2), twenty times more than the smallest emitter BLOOM (25 tonnes CO2). But even training BLOOM used the energy equivalent of powering the average American home for 41 years.
Newer models like GPT-4 are even larger in terms of model size and the number of parameters.
| Model | Parameters | Datacentre PUE | Grid Carbon Intensity | Power Consumption | CO2 Equivalent Emissions (EE) | CO2 EE x PUE |
|---|---|---|---|---|---|---|
| Gopher | 280B | 1.08 | 330g CO2ee/kWh | 1,066 MWh | 352 tonnes | 380 t |
| BLOOM | 176B | 1.20 | 57g CO2ee/kWh | 433 MWh | 25 t | 30 t |
| GPT-3 | 175B | 1.10 | 429g CO2ee/kWh | 1,287 MWh | 502 t | 552 t |
| OPT | 175B | 1.09 | 231g CO2ee/kWh | 324 MWh | 70 t | 76.3 t |
Is the use of such an energy-intense algorithmic system appropriate to the task?
Friederike Rohde researches the social and environmental impacts of digitisation and machine learning and recently co-authored a paper on the sustainability challenges of artificial intelligence, published in Ökologisches Wirtschaften (an academic journal for social and ecological economics).
Rohde says we should be careful not to get locked into unsustainable infrastructure, and, given the high energy and emissions cost of machine learning, she wants people to think carefully about the appropriateness of AI’s use across different sectors and tasks.
“Do we really need all sectors and everything to have machine learning behind every tool?” she asks.
“Is the use of such an energy-intense algorithmic system appropriate to this task?”
These questions are timely, as technology companies, including Meta, Microsoft and Alphabet (which owns Google) say they are already integrating generative artificial intelligence tools into workplace tools and social media platforms.
“AI has reached a point where we really have to think about how it’s being used and how it’s being used responsibly,” Lucey says.
He says responsible AI is really about making sure the technology is applied where it can provide more public benefit than drawback.
But he anticipates the concerns about energy use and emissions will be short-lived. Technology companies and research institutions are increasingly thinking about where they are sourcing their power from, he says.
And ultimately the energy cost will drive technological developments to make software, and then hardware, more powerful and more efficient.
“Energy consumption is one reason a lot of companies in the machine learning space are making a loss at the moment. Because the cost of actually running the query is a lot greater than any income they get from servicing the query,” he says.
One way forward is smaller models with fewer parameters, an approach open source developers are already working on: “Getting it really small and really fast,” he says.
“I can run something equivalent to GPT-3 on my laptop now right now, which is something unthinkable even six months ago.”
Could machine learning be used to help solve environmental issues like climate change?
Lucey thinks so.
For instance, he says machine learning can be used to support the decarbonisation of the electricity grid, by using the technology’s ability to better forecast weather conditions, predict and integrate variable renewable energy sources, and time energy use more efficiently.
Another example is the NOBURN app developed by the Australian Institute for Machine Learning in collaboration with the University of the Sunshine Coast. The citizen science tool captures data from people living in fire-prone areas and uses machine learning to better predict the likelihood of bushfires, with the ultimate aim of reducing their impact.
Lucey sees machine learning as enhancing productivity rather than replacing the important role of people being the ones to make decisions in high-risk scenarios.
Rohde agrees there are applications for machine learning which could make a positive impact on sustainability and agrees increasing the uptake of renewable energy might be an effective use.
But she cautions, “I think we should be aware of this narrative that machine learning tools can help us make a really great step towards sustainability.”
“Energy consumption is one reason a lot of companies in the machine learning space are making a loss at the moment. Because the cost of actually running the query is a lot greater than any income they get from servicing the query,”
Professor Simon Lucey
In other areas the usefulness of the technology is unproven, she says.
A key issue is that machine learning trained on historical datasets cannot produce transformative solutions, and may be biased towards the status quo.
For example, AI algorithms might be able to help to optimise traffic flow in a city, but they aren’t designed to produce transformative solutions – such as changing the transport system as a whole to reduce car dependency.
Rohde adds many climate and environmental issues also rely on social processes, from society deciding to take a problem seriously, to implementing measures.
“We can see what the problem is through applying machine learning but to solve the problem, it’s a social process,” she says.
“We have to make political decisions or companies have to take responsibility for what they are doing.”
Originally published by Cosmos as Massive computers chewing up gigawatts of energy to support AI
Petra Stock
Petra Stock is a journalist and engineer. She has previously worked in climate change, renewable energy, environmental planning and Aboriginal heritage policy.
