NIS News
-translated from Dutch by google translate
Today, 17:51
AI guzzles energy: 'Could use as much power as the Netherlands in four years'
If the demand for artificial intelligence (AI) continues to grow at the current rate, by 2027 AI will consume as much energy as the whole of the Netherlands. Data scientist Alex de Vries calculates this in the scientific journal Joule.
De Vries used chip production from the company Nvidia, the leading producer of chips for artificial intelligence, among others, for his calculations.
So-called generative artificial intelligence, which can be used to produce texts, images, music and video, among other things, has been experiencing explosive growth since last year. Training this AI requires large amounts of data. That requires a lot of energy.
Microsoft's search engine Bing, meanwhile, uses the artificial intelligence of the advanced text generator ChatGPT. If Google were also to generate search results with artificial intelligence, it would cost 29.2 terawatt hours of energy annually, De Vries calculated. That's the equivalent of Ireland's energy consumption over a year.
Sustainable AI
Recently, attention has already been focused on the danger of discrimination by algorithms and the generation of fake news. For example, fake versions of actors and soccer players recently surfaced in advertisements on social media. De Vries also describes high energy consumption - and thus the consequences for sustainability - as one of the risks.
Artificial intelligence is simply very energy intensive," he says.
De Vries previously conducted research into the energy consumption of blockchain, the technology that underlies bitcoin. There he saw a similar hype developing as currently surrounding artificial intelligence: "Everyone wants to put it in everything, regardless of whether it is necessary. It often adds nothing at all," De Vries explains.
Limited information
The data scientists acknowledge that his estimates are accompanied by uncertainties. Companies are not generous with sharing information. For example, it is not always known how much energy companies need to train their artificial intelligence.
De Vries: "At Google, about 40 percent of the energy consumption is probably in training and 60 percent in the applications. But at Bloom (an advanced text generator, comparable to ChatGPT, ed.) a lot of energy goes into training. while with ChatGPT the focus is overwhelmingly on the applications side.”
In addition, Nvidia's production figures are known, but it is not public to which companies the chips are sold. And the energy consumption of the chips is partly determined by how efficiently they are cooled. "Google is very efficient with cooling, but not all companies are. That is why I did not include cooling in my research," says De Vries.
Electronic waste
In the discussion about the energy consumption of AI, it is regularly stated that new equipment will be developed in the near future that uses energy more efficiently.
The rapid development of equipment also brings with it a new problem. According to De Vries, the lifespan of that equipment is probably shorter than the standard depreciation period of six years: "A new server may already be obsolete in a few years, which can lead to additional electronic waste."
