Article: Power Hungry Processing – *Watts* Driving the Cost of AI Deployment?
This article is co-authored by Sasha Luccioni and Yacine Jernite (Hugging Face) and Emma Strubell (CMU).
The energy and carbon costs of deploying AI models have largely been unknown.. until now!
The authors tested 88 models on 30 datasets from 10 different tasks from different modalities and found some pretty cool stuff:
- Generative tasks and ones that involve images are more energy- and carbon-intensive compared to discriminative tasks and ones that involve text. We found that Stable Diffusion XL uses nearly 1 phone charge worth of energy per generation.
- Training remains orders of magnitude more energy- and carbon- intensive than inference. It takes between 200 and 500 million inferences from a BLOOM-family model to reach the quantity of energy used during training. But this can be reached pretty fast for a popular model used by millions of users, like ChatGPT.
- Using multi-purpose models for discriminative tasks is more energy-intensive compared to task-specific models for these same tasks. This is especially the case for, like sentiment analysis and question answering. The difference can be a factor of 30 times depending on the dataset.
Please click on this link to read the full article.
Image credit: Image by WangXiNa on Freepik