At Rise of AI Summit 2020, Florin Manaila, Senior AI Architect and Inventor at IBM Systems Hardware Europe, addresses the carbon footprint of training AI systems on the Applied AI Day/Future Space track. Speech recognition for this recording failed, so no transcript is available, and this page is built from the YouTube description alone. According to that description, data scientists commonly train AI systems on huge quantities of data with almost no consideration of the resulting CO2 impact. Manaila's stated aim is to give data scientists visibility into the carbon footprint of their training choices and to explore how CO2 limits could be enforced at the AI data center level. No further detail, such as specific numbers, tools, or case studies discussed in the talk, could be confirmed from the available material, since the auto-generated transcript could not be recovered. Readers wanting the concrete methods or figures Manaila presented should watch the original 29-minute recording directly.
Key takeaways
- Manaila's stated premise is that data scientists train AI systems on large datasets with almost no consideration of the resulting carbon footprint.
- He aims to give data scientists visibility into the CO2 impact of their training decisions.
- He proposes exploring how CO2 limits could be enforced at the level of the AI data center.
- The talk was given on the Applied AI Day/Future Space track of Rise of AI Summit 2020.
- No transcript could be recovered for this recording, so specific figures, tools or case studies from the talk are not confirmed here.
Detailed notes
The problem: untracked carbon cost of AI training
Per the talk description, data scientists typically train AI systems on very large datasets with almost no thought given to the carbon footprint of that training. Manaila's talk sets out to change that by proposing ways to make data scientists aware of the CO2 impact of their work and by exploring how CO2 limits could be enforced at the level of the AI data center itself. Beyond this framing, no specific methods, metrics or examples from the talk could be confirmed, since the transcript for this recording could not be recovered.
Questions answered in this talk
What is the carbon footprint problem with AI training?
According to the description of Florin Manaila's talk at Rise of AI Summit 2020, data scientists train AI systems on huge quantities of data with almost no thought given to the resulting CO2 impact.
How could AI data centers limit their CO2 impact?
Florin Manaila's talk at Rise of AI Summit 2020 proposes exploring how CO2 limits could be enforced at the AI data center level, though the specific mechanisms are not detailed in the available material.