Exploiting Dual-Gate Ambipolar CNFETs for Scalable Machine Learning Classification.

Area of Science:

• Electronics
• Materials Science
• Computer Engineering

Background:

• Ambipolar carbon nanotube based field-effect transistors (AP-CNFETs) offer unique electrical properties like tri-state operation and bi-directionality.
• These characteristics are suitable for complex and reconfigurable computing systems.

Purpose of the Study:

• To design and demonstrate a mixed-signal machine learning logistic regression classifier utilizing AP-CNFETs.
• To evaluate the performance, accuracy, power consumption, and physical size of the proposed AP-CNFET based classifier.

Main Methods:

• The logistic regression classifier was designed using SPICE simulation with a 15 nm feature size.
• The system was tested on the MNIST digit dataset.
• Performance was benchmarked against Python implementations and state-of-the-art CMOS and memristor based classifiers.

Main Results:

• The AP-CNFET based classifier achieved 90% accuracy on the MNIST dataset.
• No accuracy degradation was observed compared to a Python implementation.
• The system demonstrated lower power consumption and a smaller physical size than existing mixed-signal classifiers.

Conclusions:

• AP-CNFETs are a viable technology for developing efficient and compact mixed-signal machine learning classifiers.
• The proposed design offers a promising alternative to conventional CMOS and memristor based approaches for hardware machine learning acceleration.