Binarized neural network of diode array with high concordance to vector-matrix multiplication.

Area of Science:

• Materials Science
• Electrical Engineering
• Computer Science

Background:

• Binarized neural networks (BNNs) offer computational efficiency but require specialized hardware for efficient operation.
• Traditional hardware implementations for BNNs often involve complex architectures and significant power consumption.
• Developing novel materials and device structures is crucial for advancing neuromorphic computing.

Purpose of the Study:

• To investigate the use of silicon diode arrays for performing vector-matrix multiplication (VMM) in BNNs.
• To evaluate the characteristics of p+-n-p-n+ diodes for synaptic applications.
• To demonstrate the feasibility of implementing BNNs using these diode arrays.

Main Methods:

• Fabrication of silicon diode arrays with p+-n-p-n+ structures.
• Characterization of diode electrical properties, including subthreshold swing and current ratio.
• Experimental demonstration of vector-matrix multiplication and matrix multiply-accumulate operations using diode arrays.
• Assessment of linearity, reliability, and uniformity of the diode arrays.

Main Results:

• Diode arrays achieved VMM with binarized weights and inputs.
• Diodes exhibited steep switching (subthreshold swing < 1 mV) and high current ratios (~10^8).
• Arrays demonstrated self-rectifying functionality and high linearity (R-squared = 0.99986).
• A 2x2 diode array successfully performed matrix multiply-accumulate operations with high concordance to VMM.

Conclusions:

• Silicon diode arrays are a viable component for implementing BNNs.
• The unique properties of these diodes enable efficient and compact synaptic cell design.
• The demonstrated performance supports the potential for disturbance-free, nondestructive readout and semi-permanent data holding in BNN hardware.