A cellular neural network (CNN) is an analog architecture that can significantly improve both the power and performance of various information processing functions (e.g., pattern recognition, motion detection, etc.) can be particularly efficient when compared to functional equivalents that are executed on a more traditional microprocessor. 

Current-voltage characteristics associated with tunnel field-effect transistors (TFET) and/or other LEAST Center devices could have a positive impact on the power, performance, and area associated with CNN architectures. As one example, by introducing a non-linearity in the CNN circuit architecture, one could eliminate the need for an output transfer function and the hardware associated with it. Simulations based on SPICE simulation program with integrated circuit emphasis) are used to verify correct functionality.

To consider how other CNN templates would be impacted by TFET-based CNN cells, Notre Dame researchers Mike Niemier and Sharon Hu, in conjunction with Tamás Roska’s group in Budapest, have developed a Matlab-based simulation toolset that can easily consider larger inputs and/or different templates. As an example, the labyrinth problem is illustrated at left. Settling times between 100 s of ps and ~10 ns appear possible with TFET-based CNNs.