CMOS-compatible flash-gated thyristor-based neuromorphic module with small area and low energy consumption for...

Area of Science:

• Computer Science
• Electrical Engineering
• Materials Science

Background:

• Traditional von Neumann architectures face performance bottlenecks.
• In-memory computing (IMC) aims to overcome these limitations, particularly for analog vector-matrix multiplication (VMM).
• Integrating neuromorphic modules with memory is crucial for expanding IMC capabilities.

Purpose of the Study:

• To propose a novel complementary metal-oxide semiconductor (CMOS)-compatible flash-gated thyristor-based neuromorphic module (FGTNM).
• To demonstrate the FGTNM's ability to integrate multiple neural network functions (quantization, nonlinear activation, max pooling) into a single module.
• To showcase system-level IMC performance using the FGTNM.

Main Methods:

• Development of a CMOS-compatible FGTNM.
• Integration of flash memory, flash-gated thyristor, n-type metal-oxide semiconductor, and p-type metal-oxide semiconductor devices on a single wafer.
• Evaluation of the FGTNM's footprint, energy consumption, and accuracy on CIFAR-10 classification.

Main Results:

• The FGTNM exhibits a small footprint (53 μm²) and low energy consumption (9.1 fJ/operation).
• It outperforms previous CMOS-based neuromorphic modules.
• System-level IMC using the FGTNM achieved 89.97% accuracy on CIFAR-10 classification.

Conclusions:

• The FGTNM effectively integrates diverse neural network functions, enabling efficient IMC.
• The proposed module offers superior performance in terms of size and energy efficiency.
• This work highlights the potential for co-integrating various devices for broader IMC applications beyond VMM.