4K-memristor analog-grade passive crossbar circuit

Area of Science:

• Materials Science
• Electrical Engineering
• Computer Science

Background:

• Passive analog-grade memristive crossbar circuits offer superior density for on-chip neural network storage, reducing off-chip communication costs.
• Low memristor variation is critical for efficient operation in neuromorphic systems.

Purpose of the Study:

• To develop and characterize a high-performance passive memristive crossbar circuit suitable for large-scale neuromorphic applications.
• To demonstrate the feasibility of using such circuits for accurate neural network computations.

Main Methods:

• Fabrication of a 64x64 passive crossbar circuit using a foundry-compatible, low-temperature process with etch-down patterning.
• Characterization of memristor nonvolatility, switching voltage variation, and analog properties.
• Experimental demonstration of vector-by-matrix multiplication and modeling of neural network classifiers.

Main Results:

• Achieved ~99% functional nonvolatile metal-oxide memristors with a coefficient of variance in switching voltages below 26%.
• Successfully programmed a 4K-pixel gray-scale pattern with <4% average tuning error.
• Demonstrated 64x10 vector-by-matrix multiplication with 1% average relative conductance import accuracy, modeling MNIST image classification.

Conclusions:

• The developed passive crossbar circuit meets the stringent requirements for neuromorphic computing, enabling efficient on-chip storage and processing of large neural network models.
• The low device variation and high functionality pave the way for practical, large-scale neuromorphic systems with improved performance and reduced power consumption.