Generating just nine images with AI is expensive. In addition to fully recharging your mobile phone

The AI ​​revolution won’t be cheap. The vast computing resources required by generative AI models are accompanied by an uncomfortable reality: these training processes and the use of said models themselves cause significant carbon dioxide emissions. We already knew that ChatGPT and Bard are polluting, but the Midjourney and DALL-E 2 of the world are even more polluting.

the study. Research conducted by researchers at Carnegie Mellon University and startup HuggingFace reveals that using generative AI models requires enormous amounts of energy. In the study https://arxiv.org/abs/2311.16863 It revealed the use of a tool called Code Carbon to compare the CO2 emissions of the 16 most common models used in Face-hugging huban open platform that allows these models to be easily tested and used.

ChatGPT consumes, but DALL-E consumes much more. The conclusions showed that text-generating AI models – chatbots similar to ChatGPT – consume a lot of energy, but things were particularly worrying in the case of models dedicated to generating images.

Consumption comparison. In tests, generating 1,000 inferences meant a consumption of between 0.06 and 2.9 kWh, with an average consumption of 1.35 kWh. The researchers then highlighted how a smartphone consumes an average of 0.012 kilowatt-hours of energy to charge. What does it mean?

Charging your phone gives you nine photos. With these numbers, the study concluded that generating nine images using these AI models consumes the same cost as charging a full phone. This data puts the problem into perspective, because the norm in AI processes for generative images is often to generate only a few images until we find the one we’re looking for.

There are models that are largely ineffective. Although average, there were particularly power-hungry models, and according to the study, the worst of them all consumed 11.49 kilowatt-hours per 1,000 inferences, meaning each photo basically costs the same amount of energy as recharging a mobile phone. Once. .

And emissions? Again, the main issue is also CO2 emissions, which is notable: on average, each image produced generates between 0.1 and 0.5 grams of CO2. In the most polluting models in the study, the figure was stark, with each photo emitting 2 grams of carbon dioxide.

It is not comprehensive, but it gives clues. Those responsible for the study themselves stressed that discussions about energy consumption and emissions focus on the training phase, but must take into account (very) the later use phase. They hope that this data – even while acknowledging that it is not comprehensive – will help implement measures, good practices and regulations that help mitigate the problem.

Image | Lacey Jensen

In Chataka | “Make it more”: The latest trend on ChatGPT is requesting increasingly exaggerated surreal images