Graphene memristive synapses for high precision neuromorphic computing.

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence

Background:

• Memristive crossbar architectures are key for in-memory computing in artificial neural networks (ANNs).
• Limited non-volatile states in current memristors cause weight rounding errors, reducing ANN inference accuracy and hindering on-chip training.
• This necessitates novel memristive devices with higher precision for advanced AI applications.

Purpose of the Study:

• To introduce graphene-based memristive synapses with a high number of non-volatile, arbitrarily programmable conductance states.
• To address the limitations of conventional memristors in ANNs, specifically concerning weight quantization errors and training efficiency.
• To demonstrate the potential of these advanced memristors for improved AI hardware.

Main Methods:

• Fabrication and characterization of graphene-based memristive devices.
• Demonstration of multi-level (>16) and non-volatile conductance states with high programming endurance and retention.
• Implementation of vector-matrix multiplication using these memristors with k-means clustering for weight assignment.

Main Results:

• Graphene memristors exhibit >16 programmable non-volatile conductance states with excellent retention and endurance.
• K-means clustering-based weight assignment with graphene memristors significantly improves computing accuracy compared to uniform quantization.
• The developed memristors effectively enable precise weight representation crucial for efficient ANN operation.

Conclusions:

• Graphene-based memristive synapses overcome the state limitations of conventional memristors for ANNs.
• Arbitrarily programmable conductance states enhance accuracy and enable efficient on-chip training in AI hardware.
• These advanced memristors represent a significant step towards more powerful and accurate in-memory computing systems.