A group of researchers from Hugging Face and the University of Carnegie-Mellon published the results of their analysis on energy consumption in various machine learning models. The study revealed that machine learning models used for image generation consume the most energy, while text classification models are the least energy-intensive. On average, image generation models consume about 1500 times more energy than text classification models and 60 times more than text generation models.
As an illustration, the most energy-consuming model for image generation required 11.49 kwh of energy to complete 1000 iterations. This amount of energy is equivalent to the energy consumption of approximately 950 smartphone charges (source).
/Reports, release notes, official announcements.