Project TBD Memory Vault

EE 292P: Atoms, Bits, and National Interest (Jan 13)

3 min read

EE 292P: Atoms, Bits, and National Interest (Jan 13)

Attendees: Dustin J Ross Date: January 13, 2026 Type: Class Session

Summary

Course Schedule & Upcoming Speakers

  • Next week (Jan 20): Semiconductor technology at transistor level
    • Mark Johnson (K University) - only chief semiconductor officer in US academia
    • Paired with Garo from Applied Materials/Intel on industry perspective
  • Jan 27: Computing focus
    • Tom Lee presenting on 75th anniversary of transistor invention
    • CS department researchers on “intelligence per watt” paper (Chris Ray, John Hennessy, Azalim, Hosseini)

Power Consumption Crisis in AI/Computing

  • Google electricity usage surge: 12 TWh (2019) → 24 TWh (2023) → 30 TWh (2024)
    • 3x increase driven by AI adoption
    • Estimated $7B annual electricity cost at residential rates
    • Data centers get ~50% discount, still billions in operational costs
  • Training cost comparison (GPT-4 vs Grok-4):
    • Development: $100M vs $500M (5x increase)
    • Energy: 52M kWh vs 310M kWh (6x increase)
    • Newer chips likely more efficient per operation

Material Solutions for Power Efficiency

  • Current cooling limitations: ~100W/cm² without liquid cooling
  • Advanced materials being explored:
    • Gallium nitride (GaN) for power delivery/electrification
    • Diamond for heat transfer pathways
    • Carbon nanotubes (past research)
  • Key insight: Moving bits costs more energy than computing
    • 8-bit add: 30 femtojoules
    • Moving 1 bit 1mm: 200 femtojoules
    • DRAM access: 640 picojoules

Ferroelectric Technology Breakthrough

  • 2011 discovery at Nam Lab (Germany): CMOS-compatible ferroelectric material
  • Hafnium oxide (HfO2) doped with zirconium - same material as current transistor gates
  • Voltage scaling potential:
    • Current: ~0.7V supply voltage
    • Near-term: 0.5V achievable
    • Theoretical: 0.3V possible
  • Power scales quadratically with voltage reduction

Technical Implementation Challenges

  • Variability issues with polycrystalline grains
    • Each device may have different numbers of grain boundaries
    • Intel solved similar issues by making hafnium oxide amorphous
  • Endurance limitations for computational memory applications
  • Temperature sensitivity of ferroelectric properties
  • Gap between university prototypes and high-volume manufacturing

Industry Innovation Dynamics

  • Example: Micron developed 3D non-volatile DRAM but shelved it
    • Reason: Would cannibalize existing product sales
  • Small vs large player challenges:
    • Small customers + small vendors = unstable relationships
    • Large customers need substantial revenue justification for new lines
    • Medium-sized deals ($200M) can work with established players like Infineon

Energy Efficiency of Local vs Cloud Computing

  • Communication energy costs by distance (64 bits):
    • On-chip: picojoules
    • DRAM access: 640 picojoules
    • Wireless transmission: nanojoules (10,000x higher)
  • Local processing advantages:
    • Background subtraction: 100x less energy than raw transmission
    • Feature extraction: 1,000x less energy
    • Classification: 10,000x less energy than cloud processing

Computational Memory Applications

  • Ferroelectric devices enable multiple bits per transistor
  • Analog synapse capabilities for neural networks
  • Content addressable memory possibilities
  • Startups pursuing technology:
    • Mythic: $175M recent funding, using SONOS stack
    • Kepler Computing: ~$700-800M raised, Intel spinoff

Graph View