David Patterson on AI's Carbon Footprint: Bad Advice and Best Practices

The "Bad Advice" Approach to AI Carbon Footprint

• 💡 The presentation begins with "bad advice" for maximizing an AI's carbon footprint: pick the biggest model, use older GPUs (less energy per second, but slower), and use a local data center.
• 🎯 This approach, using the Llama 405B model and an older GPU, results in approximately 240,000 metric tons of emissions for training.

Optimizing AI for a Lower Carbon Footprint

• 🚀 To minimize emissions, the advice shifts to picking the most efficient model (e.g., Google's Gamma 327B), using the latest GPUs/TPUs (faster, thus less total energy), and leveraging efficient cloud data centers.
• 🌍 Training in a location with clean energy sources (like hydroelectric power) significantly reduces the carbon impact.
• ✅ By applying these principles, the emissions for training a comparable model can be reduced to about 150 tons, a reduction of over 1,500 times.

Fallacies and Accounting in Carbon Emissions

• 🧠 Fallacy 1: Newer hardware is worse. Life cycle analyses show that newer TPUs, despite higher operational energy, are more efficient overall due to speed and manufacturing improvements.
• 🎭 Fallacy 2: Unbundled Energy Attribute Certificates (EACs). The allegory of Max highlights how buying unbundled EACs (clean energy certificates from elsewhere) provides a false sense of environmental progress without encouraging local clean energy investment.
• 📊 Fallacy 3: AI is a major consumer of global electricity. Data centers currently consume a small percentage of global electricity (around 1.2-1.5%), and AI is a fraction of that. Other factors like economic growth and air conditioners will have a larger impact.

AI's Potential for Environmental Upsides

• 💡 AI can contribute positively, such as using AI to optimize flight paths to avoid contrails, which represent a significant portion of aviation emissions.
• ⚡ A single AI innovation like contrail avoidance could potentially reduce global CO2 emissions by a factor of 5-10 compared to the total emissions of major hyperscalers.

Personal Reflections and Advice

• 🔍 Surprises: The significant impact of geography on energy intensity and Google's proactive approach to avoiding unbundled EACs and improving data center efficiency were notable.
• 🔌 Personal Changes: Reducing flights (using Google Flights' carbon emission data) and switching from wireless to wired chargers to avoid vampire power and improve efficiency.
• 🌱 Advice for Developers: Be optimistic, find something you love to do, cherish family, and optimize for happiness over wealth.
• ❤️ The