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
