Reduce the environmental footprint of your Python programs
summary
Claims and measurements are concrete: codecarbon and CarbonTracker and Green Algorithms are used to estimate emissions, training GPT‑3–class models generates ~300,000 kg CO2 (≈125 NY–Beijing round trips), and single queries cost ~2–7 g CO2. Practical tactics and results include quantization: 23–50% lower emissions with ~1–2% accuracy loss and ~4× memory savings, pruning: 20–60% energy reduction, 3× speedups, 30–90% memory cuts. Not just slides — Google Colab notebooks and experiments demonstrate these techniques and show caching, flash attention and distillation yielding major inference and energy wins.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:45The environmental impact
- 3:38Measurement and stakeholders
- 6:15AI carbon footprint analysis
- 8:11General Python optimization
- 9:42AI quantization techniques
- 11:54Model pruning strategies
- 14:15Caching methods for AI
- 17:26Flash attention optimization
- 19:46Model distillation
- 21:50The zen of green pythonic AI
- 23:41Technical debt and Q&A
related talks
31mTrack your code's CO2 emissions with Code Carbon
47mCodeCarbon: How to Measure the CO2 Impact of M...
34mPath to production: An environmentally sustainable approach
11mGreen Code 2020
34mCarbon Footprint–Aware Software Development
44mCloud computing and carbon emissions
25mBuilding a Sustainable Future
01h 02mJoint Keynote Address: Scaling AI Sustainably: An Uncharted Territory