A Low Power Mixed Signal Convolutional Neural Network for Deep Learning SoC
Abstract
Convolutional Neural Networks (CNNs) are getting fame due to their simpler design and higher performance. However, CNNs suffer from large area and power consumption constraints. The multiply-and-accumulate (MAC) unit, which perform the convolution operation inside a CNN, consumes a significant amount of power consumption. In this study we propose a mixed-signal approach for implementing analog MAC unit that can replace the digital MAC units in CNNs. The Analog MAC unit architecture is constituted from binary weighted current steering digital-to-anlaog (DAC) circuit and capacitors. A digital parallel interfaced is designed to provide input image and filter values to the MAC unit. To realize a complete CNN model a low power analog-to-digital (ADC) is then employed at the output to convert the final value back to digital value. When a 3×3 convolution is performed, the analog MAC unit offers a 10.7% reduction in area and a 59.2% reduction in power consumption compared to its fully digital counterparts.