Energy-efficient Circuits and Systems for Intelligent Edge Devices

Edge devices continue to be a fundamental aspect of daily life including healthcare, smart image systems, and smart industrial monitoring. An edge device not only collects data but also tends to perform data processing within the device. As this trend continues to grow, the devices have gotten more intelligent to enable novel applications at the edge. For example, a smart image node might use artificial intelligence (AI) to recognize persons’ faces in a real-time image stream. One of the biggest challenges of intelligent edge devices is lengthy operating time with a limited-sized battery. This hurdle is aggravated by the increased computational complexity of tasks and the larger amounts of data to process and transfer accordingly.  Thus, improving the energy efficiency of circuits and systems is critical to addressing this challenge.  This dissertation presents three different energy-efficient edge computing systems. Combinations of different methods were explored according to different application settings: (1) a 170µW image signal processor for a low-power smart imaging node, (2) a 581µW low-power neural signal processor for decoding neural signal into continuous finger movement, and (3) a neural network accelerator leveraging bit-sparsified sign-magnitude multiplier and dual adder tree.

Chair: Professor Dennis M. Sylvester