Real-time reconfigurable subthreshold CMOS perceptron

Area of Science:

• * Electronics
• * Computer Engineering
• * Materials Science

Background:

• * Perceptron circuits are fundamental building blocks in neural networks and signal processing.
• * Existing designs often face limitations in reconfigurability, power efficiency, and transistor functionality.
• * Low-voltage operation is crucial for portable and energy-constrained electronic systems.

Purpose of the Study:

• * To introduce a novel, real-time reconfigurable perceptron circuit element.
• * To demonstrate its functionality and performance through chip measurements and simulations.
• * To evaluate its potential for future applications in neural networks and mixed-signal circuits.

Main Methods:

• * Design and fabrication of a six-transistor perceptron circuit element.
• * Chip measurements to verify functionality for multiple threshold levels (T=1, 2, 3).
• * Subthreshold operation analysis at low supply voltages (100-350 mV).
• * Performance comparison with standard static complimentary metal-oxide-semiconductor (CMOS) circuits using ring oscillators.
• * Statistical simulations to assess yield under process variations and mismatch.

Main Results:

• * Successful demonstration of a reconfigurable perceptron circuit with a fan-in of three.
• * Achieved adequate outputs for thresholds T=1, 2, and 3.
• * Demonstrated viable subthreshold operation with supply voltages as low as 100 mV.
• * Competitive performance in terms of speed and energy delay product compared to static CMOS.
• * Highest reported functionality per transistor for comparable circuits (excluding floating-gate techniques).
• * Statistical simulations indicate high probabilities for working circuits despite process variations.

Conclusions:

• * The presented perceptron circuit element offers high reconfigurability and efficiency at low voltages.
• * Its high functionality per transistor and competitive performance make it a promising candidate for advanced electronic designs.
• * The circuit serves as a potential building block for future defect-tolerant mixed-signal circuits and neural networks, leveraging redundancy.