AI-generated images: Requires as much power as charging a phone

Large machine learning models will make a decisive breakthrough in 2023.

The trend is toward greater use of large models that are trained to solve many different tasks, rather than small, specialized models. It could be bad news for the climate.

The ambition to “universal models” comes at a high cost to the environment, given the amount of energy the systems require and the amount of carbon they emit.

This was written by Alexandra Sascha Lucioni, Yacine Gernet and Emma Strobel Article published on Arxiv. Researchers from Hugging Face and Carnegie took a closer look at the energy use of a number of machine learning models.

Research shows that our new popular sport, generating memes with artificial intelligence, is one of the most energy-intensive things we do online.

Record He wrote about the article first.

Out of authority

Generative AI is known to be thirsty and energy intensive.

Previous research has shown that Chat with Chat GPT costs a pint of water just to cool data centers.

From 2017 to 2021, data center giants Google, Amazon, Microsoft, and Meta doubled their electricity consumption. Data centers now use between 1 and 1.3 percent of the world's electricity. The share is growing.

Although the energy cost per task in most models is small, their utilization is enormous. The Google Translate machine learning model is used billions of times every day.

Just like charging your phone

Researchers have calculated how much energy different AI systems need to perform different tasks. They asked different AI models to perform thousands of tasks in areas such as text and image classification and image and text generation.

Not surprisingly, classification models are the best. Using machine learning to identify text or images costs between 0.002 and 0.007 kWh per 1,000 times. Forms that will generate require a lot more content, and image generating forms use more energy than text forms.

The least efficient image models used as much power as needed to charge 950 smartphones in 1,000 connections, roughly one full charge for each image generated, the researchers wrote.

They warn against uncritical use of large, multi-purpose models.

Use specialized models

The trend, especially after the introduction of Chat GPT and Copilot, is that companies are moving towards models that are trained to solve a variety of tasks.

They are more energy efficient per function in the development and training phase, but require more electricity than specialized models during use.

Over time, generic models can be more climate costly per use than smaller, specialized models, especially when there is “simpler” AI work to do, such as text and image classification.