Hybrid functional 3D artificial synapses for convolution and reinforcement learning.

Area of Science:

• Computer Engineering
• Materials Science
• Artificial Intelligence

Background:

• Conventional 2D compute-in-memory (CIM) systems excel in neural networks but are inefficient for convolutional neural networks (CNNs).
• There is a need for more efficient CIM architectures to support complex AI tasks.
• Vertically stacked 3D flash memory offers potential for novel computing paradigms.

Purpose of the Study:

• To develop a 3D compute-in-memory architecture for efficient convolution and reinforcement learning.
• To reduce area overhead and enhance energy efficiency in convolution operations.
• To demonstrate the co-integration of 3D convolution blocks (3D CB) and 3D fully connected blocks (3D FCB).

Main Methods:

• Designed a novel 3D convolution block (3D CB) utilizing vertically stacked 3D flash memory.
• Developed a compatible 3D fully connected block (3D FCB) with minimal structural modifications.
• Integrated 3D CB and 3D FCB on a single wafer for a complete system.

Main Results:

• The 3D CB significantly reduces area overhead and improves reliability and energy efficiency for convolution.
• The 3D CB and 3D FCB demonstrated distinct functionalities with seamless co-integration.
• The integrated system achieved precise and consistent path planning for autonomous driving with high efficiency.

Conclusions:

• The proposed 3D CIM architecture enables efficient convolution and reinforcement learning.
• This advancement paves the way for next-generation compute-in-memory technologies.
• The system shows promise for energy-efficient autonomous driving applications.