Sustainability of AI
summary
AI training carbon depends on location, datacenter PUE, model complexity and training time: CIFAR-10 48s vs ImageNet 11,000s to reach ~50% accuracy shows training time scales CO2 linearly. Neural architecture search can emit about 5× a US car lifetime. Example ImageNet estimates: ~1,500 g CO2e in a high-emission Chinese site, falling to ~1,000 g with good PUE, ~700 g in Germany and ~85 g in France. Practical levers: pick low-emission regions, check PUE, simplify models, trim data, fine-tune pre-trained (Hugging Face) models or train on edge.
AI-written summary · about
chapters (8)
- 0:00Start
- 1:44Reduce the Carbon Footprint of Ai
- 2:03Calculate the Carbon Footprint of the Ai Training
- 2:38Emission Factors for Different Countries
- 3:27The Power Use Efficiency of the Cloud Facility
- 4:34Training Time
- 5:50Power Usage
- 8:27Choose Your Data Wisely
related talks
45mGreen AI: Making Machine Learning Environmentally Sustainable
14mGreener Intelligence: Reducing AI’s Carbon Footprint
44mThe Invisible Monster: The Carbon Footprint of AI
01h 02mJoint Keynote Address: Scaling AI Sustainably: An Uncharted Territory- 01h 02mJoint Keynote Address: Scaling AI Sustainably: An Uncharted Territory
25mBuilding a Sustainable Future
20mBuilding GenAI That Doesn't Cost the Earth
20mHuman-Centered AI for Sustainable Development