Joint Keynote Address: Scaling AI Sustainably: An Uncharted Territory
summary
A concrete account of AI’s footprint: operational vs embodied carbon are measured across model life cycles, showing a 2021 translation training cost of 45.2 metric tons and recommendation training roughly 9× higher, Llama 3 using 3.75× more resources than its predecessor and Meta deploying 340,000 H100 for generative AI. Work includes an open MLPerf DRM benchmark, a performance model (mmax) that finds ~65% higher throughput paralyzation strategies, the ACT architectural carbon tool and the Carbon Explorer framework for joint operational/embodied tradeoffs.
AI-written summary · about
chapters (13)
- 0:00Scaling AI sustainably
- 6:05AI energy demand growth
- 12:30AI carbon footprint lifecycle
- 22:28Investment directions for sustainable AI
- 24:13Optimizing deep learning models
- 29:28LLM training challenges
- 37:20Jevons Paradox in AI
- 39:04Resilient AI training systems
- 41:55Flexible data center workloads
- 44:01Carbon-aware data center design
- 49:25AI for grid decarbonization
- 51:57Reducing AI hardware emissions
- 57:02Sustainable AI future
related talks
01h 02mJoint Keynote Address: Scaling AI Sustainably: An Uncharted Territory
44mThe Invisible Monster: The Carbon Footprint of AI
14mThe Real Cost of a Token: the Hidden Bill of Generative AI
45mGreen AI: Making Machine Learning Environmentally Sustainable
43mGenerative AI and Global Warming: co…
57mThe Next Decade of Software Is About Climate - What Is the Role of ML?
14mGreener Intelligence: Reducing AI’s Carbon Footprint
20mBuilding GenAI That Doesn't Cost the Earth