The Limited Times

Now you can see non-English news...

The dirty secret of artificial intelligence


The rise of tools like ChatGPT triggers forecasts of global data center energy consumption, which could increase fivefold

Already daily actions such as consulting the best route to go somewhere or translating a text require large amounts of energy, hydraulic and mineral resources.

Those applications run in the cloud, a euphemism for millions of powerful computers arranged in sprawling data centers.

For mobile applications to work, legions of computers are needed that store trillions of data and perform operations in fractions of a second (for example, calculating distances taking traffic into account).

It is estimated that the energy consumption of data centers represents between 1% and 2% of the world total.

But everything indicates that these figures are going to skyrocket.

Generative artificial intelligence (AI), the one that makes intelligent chatbots

like ChatGPT possible

, as well as tools that generate original artwork or music from text, needs a lot of computing power.

The big technology companies, with Microsoft and Google at the helm, have decided to integrate these functionalities into search engines, text editors or email.

Our relationship with commonly used programs is going to change: until now, we pressed a series of commands to carry out certain activities;

Soon we will find ourselves talking to the machine, asking it for tasks that we used to do before.

What effect will this paradigm shift have on the environment?

Nobody knows, but all estimates are upward.

“AI may seem ethereal, but it is physically shaping the world,” says Kate Crawford in

Atlas of AI


The Australian, principal investigator of Microsoft Research and director of the AI ​​Now Institute, warned two years ago that the "planetary costs" associated with this technology do not stop growing.

Some scientists calculated four years ago that the technology sector would account for 14% of global emissions by 2040;

others, that the energy demand of data centers will multiply by 15 until 2030.

All those forecasts may fall short.

They are from before the irruption of ChatGPT.

Google and Microsoft accumulate hundreds of millions of users.

What happens if they all start using tools supported by generative AI?

Canadian Martin Bouchard, co-founder of Qscale data centers, believes that at least four to five times as much computing power would be needed for each search.

Asked about their current consumption levels and their growth forecasts in the era of generative AI, Google and Microsoft have preferred not to provide this newspaper with specific data, beyond reiterating their intention to achieve carbon neutrality by 2030. Crawford, that "means they offset their emissions by buying people's credit" through environmental makeup actions,

One of the corridors of the data center that Google has in Douglas, Georgia (USA).

"Generative AI produces more emissions than an ordinary search engine, which also consumes a lot of energy because, after all, they are complex systems that dive into millions of web pages," says Carlos Gómez Rodríguez, professor of Computing and Artificial Intelligence at the University of La Coruna.

"But AI still generates more emissions than search engines, because it uses architectures based on neural networks, with millions of parameters that need to be trained."

How much does AI pollute?

A couple of years ago the carbon footprint of the computer industry caught up with that of aviation when it was at its peak.

Training a natural language processing model is equivalent to as many emissions as five gasoline-powered cars will expel over their lifetime, including the manufacturing process, or 125 round-trip flights between Beijing and New York.

Beyond emissions, the consumption of water resources for cooling systems (Google spent 15.8 billion liters in 2021, according to a Nature study, while Microsoft declared 3.6 billion liters), as well as



on rare metals. to make electronic components, make AI a technology with great repercussions on the environment.

Training a natural language processing model is equivalent to as many emissions as five gasoline-powered cars will expel in their lifetime.

There are no data on how much energy and of what type are consumed by large technology companies, the only ones with an infrastructure robust enough to train and feed the large language models on which generative AI is based.

There are also no specific figures on the amount of water used to cool the systems, an issue that is already causing tensions in countries like the US, Germany or the Netherlands.

Companies are not required to provide such information.

“What we have are estimates.

For example, training GPT3, the model on which ChatGPT is based, would have generated about 500 tons of carbon, the equivalent of going to and from the Moon by car.

It may not be much, but it must be taken into account that the model has to be periodically retrained to incorporate updated data”, says Gómez.

OpenAI has just introduced another more advanced model, GPT4.

And the race will continue.

Another estimate says that the electricity use that had been made in January 2023 at OpenAI, the company responsible for ChatGPT, could be equivalent to the annual use of about 175,000 Danish families, which are not the biggest spenders.

“These are projections with the current numbers of ChatGPT;

if its use becomes even more widespread, we could be talking about an equivalent electricity consumption of millions of people”, adds the professor.

Aerial view of the Google data center in Saint-Ghislain, Belgium.

The data opacity will begin to dissipate soon.

The EU is aware of the increasing energy consumption of data centers.

Brussels has a directive underway that will begin to be discussed next year (and, therefore, it would take at least two years to enter into force) that sets requirements for energy efficiency and transparency.

The US is working on a similar regulation.

The expensive training of algorithms

“AI carbon emissions can be broken down into three factors: the power of the hardware being used, the carbon intensity of the power source that powers it, and the energy used in the time it takes to train the model. ”, explains Álex Hernández, a postdoctoral researcher at the Quebec Institute of Artificial Intelligence (MILA).

It is in training where most of the emissions are concentrated.

This training is a key process in the development of machine learning models, the type of AI that has grown the fastest in recent years.

It consists of showing the algorithm millions of examples that help it establish patterns that allow it to predict situations.

In the case of language models, for example, it is that when you see the words “the Earth is” you know that you have to say “round”.

The electricity use in January 2023 at OpenAI, the company responsible for ChatGPT, is equivalent to the annual use of about 175,000 Danish families

Most data centers use advanced processors called GPUs to perform training on AI models.

GPUs need a lot of power to run.

Training large language models requires tens of thousands of GPUs, which need to operate around the clock for weeks or months, according to a recent Morgan Stanley report.

“Large language models have a very large architecture.

A machine learning algorithm to help you choose who to hire might need 50 variables: where you work, what salary you have now, previous experience, and so on.

GhatGPT has more than 175 billion parameters,” illustrates Ana Valdivia, postdoctoral researcher in computing and AI at King's College London.

“You have to retrain all that kind of structure, and also host and exploit the data on which you work.

That storage also has a consumption ”, she adds.

Hernández, from MILA, has just presented an article in which he analyzes the energy consumption of 95 models.

“There is little variability in the hardware used, but if you train your model in Quebec, where the majority of electricity is hydroelectric, you reduce carbon emissions by a factor of 100 or more relative to places where coal, gas or others predominate” , emphasizes the researcher.

Chinese data centers are known to be powered by 73% coal-generated electricity, which resulted in the emission of at least 100 million tons of CO₂ in 2018.

Directed by Joshua Bengio, whose contribution to deep neural networks earned him the Turing Prize (considered the Nobel Prize in Computer Science), MILA has developed a tool, Code Carbon, capable of measuring the carbon footprint of those who program and train algorithms.

The goal is for professionals to integrate it into their code to know how much they emit and that this helps them make decisions.

More computing power

There is the added problem that the computing power required to train the largest AI models doubles every three to four months.

This was already revealed in 2018 by an OpenAI study, which also warned that "it is worth preparing for when systems need much greater capacities than they currently do."

It is a speed much higher than that set by Moore's Law, according to which the number of transistors (or power) of microprocessors doubles every two years.

“Taking into account the models that are currently being trained, more computational capacity is needed to guarantee its operation.

Surely, the big technology companies are already buying more servers”, predicts Gómez.

For Hernández, the emissions derived from the use of AI are less of a concern for several reasons.

“There is a lot of research aimed at reducing the number of parameters and the complexity of the energy that the models need, and that will improve.

However, there aren't that many ways to reduce them in training: that's where weeks of heavy use are needed.

The former is relatively easy to optimize;

the second, not so much.

One of the possible solutions to make training less polluting would be to reduce the complexity of the algorithms without losing efficiency.

“Do you really need so many millions of parameters to get models that work well?

GhatGPT, for example, has been shown to have many biases.

The way to achieve the same results with simpler architectures is being investigated”, reflects Valdivia.

You can follow

EL PAÍS Tecnología





or sign up here to receive our

weekly newsletter


Subscribe to continue reading

Read without limits

Keep reading

I'm already a subscriber

Source: elparis

All tech articles on 2023-03-23

Similar news:

You may like

Life/Entertain 2023-04-12T06:29:27.175Z
News/Politics 2023-02-25T18:45:16.014Z

Trends 24h

Tech/Game 2023-05-29T15:02:30.371Z


© Communities 2019 - Privacy

The information on this site is from external sources that are not under our control.
The inclusion of any links does not necessarily imply a recommendation or endorse the views expressed within them.