Recent research emphasizes that AI model training and deployment can significantly impact carbon footprints due to the massive computational resources required. A study published in the Journal of Machine Learning Research highlights that the majority of AI's carbon emissions arise from the energy consumed by data centers housing these resources.
Key finding:
A significant proportion of energy consumption during model training and inference comes from the data centers' 'infrastructure layer', including cooling and power distribution systems, which contribute up to 64% of the total energy consumption.
Practical impact:
To mitigate AI's carbon footprint, researchers recommend exploring alternative infrastructure designs, such as in situ training on edge devices, in addition to optimizing model architecture and hyperparameters. This shift towards sustainable AI development could lead to substantial energy savings in the long run, ultimately aiding in reducing overall carbon emissions.
Publicado automáticamente
Top comments (0)