Power grids creak as demands for AI increase

There’s a big problem with generative AI, says Sasha Luccioni of Hugging Face, a machine learning company. Generative AI consumes energy.

“Every time you query the model, everything is activated, so it is extremely computationally inefficient,” she says.

Consider the Large Language Models (LLMs) at the heart of many generative AI systems. They have been trained on vast stores of written information, which helps them produce texts in response to virtually any query.

“When you use generative AI… you are generating content from scratch, essentially inventing answers,” explains Dr. Luccioni. This means the computer has to work a lot.

A generative AI system can use about 33 times more energy than machines running task-specific software, according to a recent study by Dr. Luccioni and colleagues. The work has been peer-reviewed but has not yet been published in a journal.

However, it is not your personal computer that uses all this energy. Or your smartphone. The calculations we increasingly rely on happen in gigantic data centers that are, for most people, out of sight and out of mind.

“The cloud,” says Dr. Luccioni. “You don’t think about these huge metal boxes that get hot and consume so much energy.”

The world’s data centers are using more and more electricity. In 2022, they consumed 460 terawatt hours of electricity, and the The International Energy Agency (IEA) expects this will double in just four years. Data centers will be able to use a total of 1,000 terawatt hours annually by 2026. “This demand is approximately equivalent to Japan’s electricity consumption,” says the IEA. Japan has a population of 125 million people.

In data centers, huge volumes of information are stored for retrieval anywhere in the world – from emails to Hollywood movies. The computers in these faceless buildings also power AI and cryptocurrency. They sustain life as we know it.

But some countries know very well how energy-intensive these installations are. There is currently a moratorium preventing the construction of new data centers In Dublin. Almost a fifth of Ireland’s electricity is consumed by data centres, and this figure is expected to grow significantly in the coming years – meanwhile, Irish households are reducing their consumption.

The head of National Grid said in a speech in March that demand for electricity in UK data centers will increase six times in just 10 years, fueled in large part by the rise of AI. However, National Grid expects the energy needed to electrify transport and heating to be much greater in total.

Utilities in the US are starting to feel the pressure, says Chris Seiple of Wood Mackenzie, a research firm.

“They are being hit by data center demands at exactly the same time as there is a renaissance – thanks to government policy – ​​in national production,” he explains. Lawmakers in some states are now rethinking the tax benefits offered to data center developers due to the enormous pressure these facilities are placing on local energy infrastructure. according to reports in the US.

Seiple says there is a “land grab” for data center locations near power stations or renewable energy centers: “Iowa is a hotbed of data center development, there’s a lot of wind generation there.”

Some data centers can afford to go to more remote locations these days because latency—the delay, usually measured in milliseconds, between a data center sending information and the user receiving it—is not a major concern. to the increasingly popular generative AI systems. In the past, data centers that handled emergency communications or financial trading algorithms, for example, were located within or very close to large population centers to achieve the best response times.

There is no doubt that energy demand from data centers will increase in the coming years, but there is huge uncertainty about the amount, highlights Mr. Seiple.

Part of this uncertainty is due to the fact that the hardware behind generative AI is constantly evolving.

Tony Grayson is general manager of Compass Quantum, a data center company, and points to Nvidia’s recently released Grace Blackwell supercomputer chips (named after a computer scientist and a mathematician), which are designed specifically to power cutting-edge processes. , including generative AI. , quantum computing and computer-aided drug design.

Nvidia says that in the future, a company will be able to train AIs several times larger than those currently available in 90 days. It would require 8,000 previous-generation Nvidia chips and an electricity supply of 15 megawatts.

But the same work could be done at the same time for just 2,000 Grace Blackwell chips, and they would need a supply of four megawatts, according to Nvidia.

This still results in 8.6 gigawatt hours of electricity consumed – approximately the same amount as the entire city of Belfast consumes in a week.

“Performance is increasing so much that the overall energy savings are huge,” says Grayson. But he agrees that energy demands are defining where data center operators locate their facilities: “People are going where the cheap energy is.”

Dr. Luccioni notes that the energy and resources required to manufacture the latest computer chips are significant.

More business technology

Still, it’s true that data centers have become more energy efficient over time, argues Dale Sartor, a consultant and affiliate at Lawrence Berkeley National Laboratory in the US. Their efficiency is often measured in terms of power usage effectiveness, or PUE. The lower the number, the better. High-end data centers have a PUE of about 1.1, he notes.

These facilities still create significant amounts of waste heat, and Europe is ahead of the US in looking for ways to utilize that waste heat – such as heating swimming pools – says Mr. Sartor.

Bruce Owen, managing director of Equinix, a data center company in the UK, says: “I still think demand will grow even more than the efficiency gains we see.” He predicts more data centers will be built including on-site power generation facilities. Equinix was planning permission denied for a gas-powered data center in Dublin last year.

Sartor adds that costs may ultimately determine whether generative AI is worth it for certain applications: “If the old method is cheaper and easier, then there won’t be much of a market for the new method.”

Dr. Luccioni stresses, however, that people will need to clearly understand how the options before them differ in terms of energy efficiency. She is working on a project to develop energy ratings for AI.

“Instead of choosing this GPT-derived model that is very clunky and consumes a lot of power, you can choose this Energy Star A+ model that will be much lighter and more efficient,” she says.