Generative AI tools such as OpenAI’s ChatGPT and DALL-E consume a massive amount of power, especially when generating images from text prompts.
A group of researchers from Hugging Face and Carnegie Mellon University have published a paper that examines the power requirements of AI tools for different tasks. The findings emphasize the significant carbon emissions produced when using them for image creation instead of employing a human artist.
Strikingly, the study revealed that the open source XL model from Stable Diffusion, considered the least efficient, consumed nearly the same amount of power per image as it takes to fully charge a smartphone.
The paper states that creating 1,000 images with this model produces the same amount of carbon emissions as driving an average gasoline-powered car for 4.1 miles.
However, the energy consumption for generating 1,000 images may vary depending on the model used, but on average it requires 2.907 kWh. This is approximately equivalent to charging a phone’s battery to 24 percent for each image.
According to the researchers, generating text consumes significantly less power compared to other processes, requiring only as much energy as three smartphone charges to handle 1,000 queries. But when the data is expanded to a global level, it becomes evident that companies such as OpenAI and Google are experiencing significant increases in energy costs as they strive to maintain their generative tools operational.
Just cooling these servers alone has an astonishing environmental footprint. According to Google’s 2023 Environmental Report, the company used an astronomical 5.6 billion gallons of water last year, a 20 percent increase over its 2021 usage.
In short, the AI industry’s carbon footprint will continue to be a big problem, especially as the world creeps ever closer to a climate catastrophe. The latest research serves as a reminder that even on an image-by-image basis, the energy costs of using these generative AI tools can be considerable.
Reference- Futurism, The Register, Hugging Face and Carnegie Mellon University Research Report