AI summary by TwinMind

Memristors for Analog AI Chips

Asianometry

16m 24s

Historical Background and Theory

  • 1971: UC Berkeley Professor Leon Chua published paper proposing memristor (memory resistor)
    • Strange theoretical proposal, mostly forgotten for 30+ years
    • Now explored for analog AI accelerators due to AI revolution
  • Chua's paper analyzed three fundamental two-terminal passive circuit elements:
    • Capacitor (invented 1745, Leyden jar): relates charge and voltage, stores/discharges electrical energy
    • Resistor (1827, George Simon Ohm's law): relates voltage and current, limits electrical flow
    • Inductor (1831, Michael Faraday electromagnetic induction): relates current and magnetic flux linkage, stores energy in magnetic field
  • Chua identified missing fourth element connecting magnetic flux linkage and charge
    • Proposed memristor to fill the gap
    • Student Sung Mo Steve Kang later clarified concept in second paper

Memristor Characteristics

  • Resistive device with inherent memory
  • Similar to resistor but resistance depends on history of electrical current/voltage
  • Nonlinear electronic component - double input doesn't yield double output
  • Water metaphor: sand filter in pipe that changes flow resistance based on accumulated debris
  • Remembers how much current passed through without storing it (unlike capacitor)

Controversy Over Fundamentality

  • Critics argued memristor and resistor too similar - both about resistance values
  • Memristor just has nonlinear output - questioned if truly fundamental
  • Published as curiosity among other nonlinear circuit work
  • No actual memristor existed initially - appreciated as theoretical concept

HP's 2008 Implementation

  • HP researchers working with oxide materials and resistive switching physics
  • Created two-terminal, two-layer semiconductor:
    • Two platinum electrodes separated by thin titanium dioxide film
    • Titanium dioxide normally high resistor, used in white paint/sunscreen
    • Withholding oxygen during deposition creates oxygen vacancies (doping)
    • Current passes through doped (low resistance) and undoped (high resistance) regions
    • Boundary drift between regions changes final resistive value based on current history
  • 2008: HP team (R. Stanley Williams) announced memristor implementation
  • HP equations broadly (not precisely) fit boundary drift behavior

Implementation Criticism

  • Mixed reception - devices invented frequently, need to match ideal equations
  • Critics: HP memristor doesn't match Chua's equations, ideal memristor unachievable
  • Blaise Mouttet (Xerox patent officer) called HP claims "propaganda campaign"
  • Historical precedent: 1960 Bernard Widrow coined "memistor" (memory resistor)
    • Three-terminal device with memory and resistance
    • Physical implementation: two-terminal wire in chemical bath
    • Not solid-state, not scalable for semiconductor production
    • Used for neural network research in 1960s, then forgotten
  • Other historical examples: carbon arc lamp discharge experiment (1801)
  • Chua declared 2008 HP device valid memristor but changed definition
    • Now includes all memory devices based on resistive switching regardless of charge/flux relationships

Modern Memristor Types

  • Most share metal-insulator-metal (MIM) sandwich structure similar to HP 2008 design
  • Metal options: copper, platinum, tungsten, silver, gold, titanium nitride, ceramics, carbon nanotubes
  • Two insulator types:
    • Organic: cheaper, flexible (like OLEDs)
    • Inorganic: better resistive switching, compatible with traditional semiconductor manufacturing

AI Applications - Crossbar Architecture

  • Memristors connected in crossbar pattern: rows of MIM sandwiches
  • Top: horizontal metal bars, bottom: vertical metal bars
  • Reading data: sensing voltage through top wire, measure bottom wire for resistance value
  • Writing data: activate top/bottom wires at half threshold, only target memristor receives full voltage
  • Programmed with unique resistive values for AI calculations

Analog AI Computation

  • Performs vector-matrix multiplication operations for AI inference
  • Model parameters obtained through training
  • Analog computation: continuous values vs digital ones/zeros
  • Uses electrical engineering laws for multiplication/summation across memristor rows
  • Each memristor has special resistive value representing model parameters
  • Voltage applied across top row interacts with unique resistance values creating current
  • Current = voltage × conductance (inverse of resistance)
  • Single readout value = literal physical manifestation of multiply-add action

Compute-in-Memory Benefits

  • Calculations performed inside memory cells where parameters stored
  • Skips Von Neumann bottleneck power/latency issues
  • Single low-power step for vector-matrix multiplication
  • Significant time/energy savings vs digital AI accelerators
  • Constant time concept: calculations done in same time regardless of matrix size
  • Best for inference (real-world model running)
  • Training requires more precision than current analog systems provide

Manufacturing and Integration

  • CMOS compatible: can be made on silicon wafers using traditional semiconductor techniques
  • Deposition, photolithography, etching processes
  • Enables CMOS-memristor hybrids
  • March 2023 Nature article: CMOS part stabilizes readable resistance levels
  • Major benefits:
    • Very small memristors (~10nm thick)
    • 3D stacking capability for more memory capacity
    • Leverage existing semiconductor ecosystem investments
    • Quick scaling of advanced product production

Manufacturing Challenges

  • Defectivity: normal manufacturing defects cause below-expectation performance
  • Breakdown after few cycles
  • Analog systems fundamentally depend on physical interactions
  • Defects matter more than in digital systems
  • Need new metrology and test tools for detection/fixing throughout production

Competition and Market Outlook

  • Memristors don't monopolize analog computation for neural networks
  • IBM analog AI chip uses phase change memory (heat-induced material phase changes)
  • Other analog AI accelerators use light interactions
  • Memristor potential due to CMOS ecosystem compatibility
  • Question: 10x performance leap needed for paradigm-shifting technologies to overcome limitations
  • Market exists for memristor-based accelerators in near future

TwinMind

Get summaries like this for your own meetings

TwinMind takes notes in the background during meetings, lectures, and calls, then writes the summary for you.

Free plan to get started

No bot joins your calls

Phone, desktop & Chrome apps

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome