Using AI at scale requires huge amounts of energy. According to scientists, generating a single image costs the same amount of energy as it takes to fully charge your smartphone.
Scientists from Carnegie Mellon University in the United States and AI developer Hugging Face have calculated how much energy artificial intelligence needs to perform certain tasks. They looked at the emissions of the 10 most common AI tasks on the Hugging Face platform. For each task – for example, creating text or an image – they conducted experiments with a large number of models.
who are they research, which has been published on arXiv but has not yet been peer-reviewed, shows that creating images is the most wasteful task of all. The heaviest model is the Stable Diffusion XL Base 1.0, which emits the same amount of emissions per thousand images as driving a car 6.5 kilometres. Generating an image with this model uses the same amount of energy as charging a smartphone, although researchers acknowledge that there is significant variation between models and the size and quality of the image they generate.
Copywriting is the least CO2-intensive activity. For every thousand copies with the most economical model, DistillBERT, the same amount of CO2 is released as a car after just one meter of emissions. Generating texts with this model is about 7000 times less polluting than Stable Diffusion XL Base 1.0. Striking: even if you compare this model with the most economical model that generates images, emissions are 700 times lower.
As much electricity as the Netherlands
Earlier this year, economist Alex de Vries published that AI energy consumption would rise to around 85 to 134 terawatt hours by 2027, more than the entire Netherlands. Our country consumed 108 terawatt hours in 2022. AI will require half a percent of global energy consumption if this scenario becomes a reality, according to De Vries.
In his calculations, the economist did not take into account some important factors. In his model, the growth rate of AI remains unchanged. There are also no improvements in energy efficiency. De Vries also did not incorporate the amount of energy needed to cool the servers into his study.
Read more news about science.
