Comparing Analog and Digital Processing for Ultra Low-Power Embedded Artificial Intelligence

In this paper, a comparison between analog and digital processing focused on ultra-low power embedded artificial intelligence is proposed. Several works developed before [1] –[6] demonstrate that features extraction and high sampling rate ADC are the most energy expensive tasks in fully digital embedded machine learning applications. Therefore, in this work analog and digital processing are compared, showing that under some conditions, analog processing is at least 30 times more efficient in terms of power consumption without taking into account the additional effect of the reduction of analog-to-digital sampling rate. Two case studies are presented: to set these ideas on a simple example, first order filter implementations using analog and digital circuits are first compared. Then, two techniques of spectrum analysis using digital FFT and analog filter benches are presented and discussed. Finally, a rule defining the situations where analog is more relevant than digital processing is proposed. This one can be used for intelligent Internet of Things (IoT) autonomous systems working on small batteries such as a single CR2032 coin cell for a very long time.