Analog monolayer SWCNTs-based memristive 2D structure for energy-efficient deep learning in spiking neural networks

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence

Background:

• Emerging memristor technology aims for energy-efficient computation by integrating storage and processing within memory crossbars.
• Artificial Intelligence (AI) systems benefit from advancements in materials science and memory devices.
• Memristor devices are crucial for developing next-generation computing architectures.

Purpose of the Study:

• To fabricate an analog memristor device using single-wall carbon nanotubes (SWCNTs).
• To investigate the device's analog switching ability, state stability, and tunable conductance/capacitance.
• To demonstrate the memristor's application in energy-efficient spiking neural networks (SNNs) for AI.

Main Methods:

• Fabrication of an analog memristor device with a planar structure using SWCNTs as the switching medium.
• Characterization of the memristor's analog switching behavior, state stability, and simultaneous tuning of conductance and capacitance.
• Deployment of the memristor in a spiking neural network (SNN) for computation-in-memory (CIM) tasks.

Main Results:

• The fabricated SWCNT memristor exhibits stable analog switching characteristics.
• The device demonstrates simultaneous tuning of conductance and capacitance, enabling multi-state storage and long-term memory.
• The CIM implementation in SNNs achieved 95% inference accuracy for Vector Matrix Multiplication with femtojoule-level energy efficiency per spike.

Conclusions:

• The SWCNT-based analog memristor is a promising candidate for bio-inspired computing and AI applications.
• The device's features facilitate energy-efficient analog computation in deep learning systems.
• This work highlights the potential of SWCNTs in advancing future AI hardware.