Green AI: Energy-Efficient and Sustainable AI
🖱️ Interaktiv: Ziehen zum Drehen · Lektion 14
Learning objectives
After this lesson you will be able to:
- Realistically assess the energy consumption of AI applications
- Apply green AI principles in everyday work
- Evaluate the costs and benefits of different model sizes
- Use efficiency levers (caching, small models, batch processing)
1. AI and energy consumption
Every AI query consumes electricity — massively during training, moderately during inference (answering), but relevant in aggregate.
What consumes how much?
| Process | Energy | Comparison |
|---|---|---|
| One chat query (inference) | ~0.003 kWh | like 1/10 of a phone charge |
| Training a large LLM | ~1,000 MWh | like 100 households/year |
| Human research (10 min) | ~0.0002 kWh | significantly less |
Insight: the individual query is cheap. The lever lies in repetition (the same question, a thousand times) and training.
💡 The 80/20 of AI efficiency
Most costs don't come from complex questions, but from:
- The same queries repeatedly (no caching) → caching saves 30-60%
- Models that are too large for simple tasks → small models are often enough
- Failed attempts in a loop → better prompts = fewer runs
2. Green AI and AI costs
Green AI = research and practice of running AI with as few resources as possible. This saves money and CO2.
Levers for efficient AI use
| Lever | Measure | Effect |
|---|---|---|
| Model choice | Small model for simple tasks (classification, extraction) | −70% cost |
| Caching | Cache answers to identical prompts | −40% queries |
| Batch processing | Many documents in one pass | −50% overhead |
| Quantization | 8-bit instead of 16-bit models locally | −50% storage/RAM |
| Local models | With data protection/data volume, on-premise pays off | Independent of cloud costs |
🏢 Praxis-Szenario: Example: support ticket triage
3. Efficient AI in the company
A practical checklist for sustainable AI use:
Quick check
- Do we use small instead of large models for simple tasks?
- Do we cache recurring answers?
- Do we have prompt templates to minimize failed attempts?
- Do we check local/on-premise models for data volumes?
- Do we measure AI costs per process (not just as a lump sum)?
✅ Wichtige Erkenntnisse
Haken setzen, um deinen Lernfortschritt zu markieren: