Naveen Rao: 4D Computing, AI's Energy Wall & Beating Biology

影片總結

Naveen Rao says Unconventional AI’s first dynamical chip generated images at about 500 nanojoules each, aiming for 1,000× power efficiency.

Naveen Rao, CEO of Unconventional AI, argues AI is a transformative force but warns that energy could constrain its growth: Google processes 3.2 quadrillion tokens monthly, which he estimates would require 12 gigawatts at 10 joules per token. Since energy is already about half the cost of serving a token, his startup is rethinking computer architecture to reduce data movement. Rao says its first physical dynamical computer was built in five months and generated images at roughly 500 nanojoules each, compared with millijoules for a GPU. The design combines computation and memory, using physical dynamics and time—an approach he calls 4D computing. He aims for 1,000× greater power efficiency and says a rack-scale data-center product could arrive within two years. Existing models will require porting at the model layer, while the team bridges theorists and chip designers using Python libraries. Rao’s ultimate goal: beat biology and enable widely distributed computing and robotics.

章節

1. 0:00Welcome Naveen Rao: From Nirvana Systems to Unconventional AI’s 1,000x Efficiency Goal
2. 4:45AI Energy Gap: Google’s 3.2 Quadrillion Tokens and an Estimated 12-Gigawatt Load
3. 10:24Dynamical Computing: Uno’s Oscillator Images and Sparsity’s n² Scaling Challenge
4. 14:58First Physical Dynamical Computer: 500 Nanojoules per Image and 4D Computing
5. 19:40Unconventional AI Product Path: Rack-Scale System Within Two Years

完整分析還有什麼

5 章 · 27 個重點 · 5 段段落總結 · 心智圖

其餘 24 個重點、段落總結與心智圖在完整分析裡。 分析這支影片,看完整內容 →