AI Video Generation‘s Energy Consumption Far Exceeds Expectations,New Study Finds
PARIS – Generating videos with artificial intelligence consumes considerably more energy than previously understood,potentially posing challenges to climate goals and electricity grids,according to a new study published on Arxiv by researchers from hugging Face. While the energy demands of AI text and image generation have been assessed,this is among the first research to quantify the power usage of AI video creation.
The study reveals that video generation’s energy consumption isn’t linear; doubling video length quadruples the energy required, and doubling resolution also results in a fourfold increase in energy use. The most advanced model tested consumed 415 Wh for a five-second video at 1,280 x 720 pixels – equivalent to 81 images. A smaller version of the same model used 90 Wh for the same clip.
For comparison, a recent study cited in the research estimated text generation at 0.047 Wh and image generation at 2.9 Wh.
Researchers warn that this exponential increase in energy use could lead to soaring electricity costs and a ample carbon footprint. They suggest companies like Google may struggle to meet climate objectives due to the energy demands of AI video generation, and point to the need for model optimization through techniques like bright caching, low-precision inference, and improved attention mechanisms.
The study focused on open-source models, and the authors note the results may not reflect the energy consumption of Google’s VEO 3, a leading model in the field – a potential description for Google’s pursuit of powering its servers with a nuclear reactor.
