Energy- and Area-Efficient Ionic-Switch Activation Neuron for Monolithic 3D Neural Network Architectures.

Area of Science:

• Materials Science
• Computer Engineering
• Neuroscience

Background:

• The von Neumann architecture presents a bottleneck for hardware-based neural networks.
• Advancements in hardware-oriented neural network design are crucial for efficient computing.

Purpose of the Study:

• To propose a monolithic 3D vertically integrated neural network architecture.
• To demonstrate an ion-based switching neuron device as a core component.

Main Methods:

• Conceptually proposed a 3D vertically integrated neural network architecture.
• Experimentally demonstrated an ion-based switching neuron device with ReLU-type output.
• Incorporated a resistance element for synaptic device compatibility.

Main Results:

• The ion-based neuron device exhibits rectified linear unit-type output.
• The proposed architecture supports vertical multiply-accumulate operations and horizontal data transmission.
• Projected reductions in area (≥10^3-fold) and energy consumption (≥10^5-fold) compared to conventional networks.

Conclusions:

• Established a device-level hardware basis for high-density, energy-efficient computing.
• Proposed a scalable architectural direction for parallel in-memory computing systems.
• The ion-based neuron device is suitable for monolithic 3D stacking.