Analytical Models for Calculating Power and Performance of a CNN System

Cellular neural networks (CNNs) are a powerful analog architecture that can outperform the traditional von Neumann architecture for spatio-temporal information processing applications – e.g., image processing, speech recognition, etc. Much existing work reports energy dissipations for CNNs at the chip level, which includes dissipation of sensors, actuators, accompanying digital components, etc. As such, the impacts of various system variables – e.g., application templates, characteristics of the resistive element, etc. – on the energy profile of a CNN cannot be easily determined. In this work, we propose analytical models to calculate CNN power dissipation, and performance (measured by settling time). Settling times obtained via the model for different non-linear characteristics are verified through simulation. By using these models, we show that a Tunneling FET (TFET) circuit based non-linear CNN demonstrates 4x, and 1.47x energy savings (assuming equal settling times), when compared to a conventional linear resistor based CNN for a tactile sensing, and a pattern recognition problem, respectively. Furthermore, a TFET-based operational transconductance amplifier (OTA) is proposed, and shown that it can provide additional 1.5x improvement in energy when used to replace a CMOS-based OTA in CNN.