Monolithically Integrated Complementary Ferroelectric FET XNOR Synapse for the Binary Neural Network

Area of Science:

• Neuromorphic computing and artificial intelligence hardware.
• Advanced semiconductor device physics and fabrication.

Background:

• Neuromorphic computing mimics the brain for efficient AI hardware.
• XNOR synapse-based Binary Neural Networks (BNNs) offer compact size and low cost.
• Existing XNOR synapses face trade-offs between cell density and accuracy.

Purpose of the Study:

• To develop nonvolatile XNOR synapses with high density and accuracy.
• To overcome limitations of previous XNOR synapse designs.
• To advance hardware implementation for AI and neuromorphic systems.

Main Methods:

• Utilized monolithically stacked complementary ferroelectric field-effect transistors (C-FeFETs).
• Employed a dual-gate configuration and unique operation scheme for n-type ferroelectric TFTs.
• Performed array-level simulations (512x512 subarray) and system-level analysis.

Main Results:

• Achieved 60F² per cell density with high accuracy using 2C-FeFETs.
• Demonstrated improved image recognition accuracies (MNIST +3.17%, CIFAR-10 +14.07%) compared to other synapses.
• Exhibited high throughput (717.37 GOPS) and energy efficiency (196.7 TOPS/W).

Conclusions:

• The developed C-FeFET nonvolatile XNOR synapses offer superior density and accuracy.
• This approach significantly enhances performance for AI hardware and neuromorphic systems.
• The technology holds promise for high-density memory, logic-in-memory, and neural network hardware.