EE 292P: Atoms, Bits (Jan 27)

Attendees: Dustin J Ross Date: January 27, 2026 Type: Class Session

Summary

Course Overview & Context

• EE 292P: Atoms, Bits, and National Interest lecture session
• Building from energy band gap concepts toward compute efficiency challenges
• Focus on transistor scaling limits and next-generation computing approaches

Transistor History & Scaling Crisis

• Bell Labs Origins (1930s-1940s)
  • Marvin Kelly identified exponential growth in technician requirements
  • Projected every adult American would need to be Bell technician by 2000
  • Conscious decision to pursue solid-state solutions over vacuum tubes
• Shockley’s Team Achievement
  • December 1947: First transistor discovery (actually accidental)
  • Original MOSFET design failed by 5 orders of magnitude
• Industry Evolution
  • Fairchild Semiconductor: “Traitorous Eight” left Shockley (1957)
  • Integrated circuits solved discrete component assembly inefficiency
  • Moore’s Law: Cost per transistor optimization, not just density

Current Computing Energy Crisis

• Exponential Growth Problem
  • AI model parameters growing super-exponentially
  • Computing energy consumption ~10% of global electrical production
  • Extrapolation: Would require Dyson sphere by ~2140
• Hardware Response Limitations
  • Nvidia Blackwell: 200B transistors, >1kW power, 4 teraflops/watt
  • Moore’s Law dead - cost per transistor no longer decreasing

Brain-Inspired Computing Solutions

• Biological Efficiency Benchmark
  • Human brain: 20W total (12W for compute after heating)
  • Composes symphonies, poetry with vanishingly small power budget
• Neocortex Architecture Insights
  • ~100K cortical columns as universal computing elements
  • Same substrate handles motor, linguistic, visual processing
• Hamiltonian Computing Approach
  • Focus on energy flows rather than forces/masses
  • Geometric problem solving for path finding, optimization
• Phase-Based Implementation
  • Oscillator networks with energy-minimization dynamics
  • Program by defining energy costs, not algorithms

Local AI Efficiency Research

• Intelligence Per Watt Metric
  • 88.7% of queries handleable by local models (vs frontier models)
• Efficiency Improvements
  • 2x intelligence/watt improvement year-over-year
  • 18x intelligence/joule improvement over 16 months
• Smart Routing Benefits
  • 5x energy reduction through intelligent query routing
  • 80% router accuracy sufficient for substantial gains

PC Era of AI Transition

• Mainframe to Personal Computing Analogy
  • Current: Centralized data centers like 1940s-50s mainframes
  • Future: Distributed local inference like PC revolution
  • Consumer GPUs approaching data center performance