Practical AI Photonic computing for AI acceleration
Nov 2, 2021
Nicholas Harris, CEO of Lightmatter, leads a company pioneering photonic computing for AI acceleration. He discusses the energy constraints of traditional transistor-based technology and the necessity for innovative solutions. The podcast dives into how photonic chips, utilizing lasers, can greatly enhance both performance and efficiency in deep learning. Harris highlights advancements that tackle issues like heat management and signal interference. He emphasizes the crucial difference between training and inference in AI and the powerful potential of photonics in overcoming current computational challenges.
AI Snips
Chapters
Transcript
Episode notes
AI Hardware Energy Limits
- Transistor-based AI hardware faces fundamental energy efficiency limits.
- Energy scaling hasn't kept up with transistor shrinking, causing overheating and hindering progress.
Advanced Cooling Methods
- High-power chips like NVIDIA's A100 and Intel's Ponte Vecchio require advanced cooling.
- Azure uses immersion cooling with edible oil, showcasing the lengths data centers go to.
Energy Consumption of AI
- Compute and interconnect will consume a significant portion of global energy by 2030.
- This unsustainable trend may slow AI progress due to financial and logistical constraints.