In the evolution of biological organisms, the ability to see has been a popular feature for millions of years. However, the majority of machines built today remain essentially blind to the photons that continually bombard them. Effective real-time computer vision has countless applications and would have a transformative effect on the economy and safety. Enabling machines to see relies not only on the advancement of computer vision research, but also the design of the hardware platforms that will execute these algorithms with sufficient energy efficiency.
Until recently, the conventional wisdom was that the ability to manufacture chips with exponentially increasing numbers of transistors will bring us processors that are fast enough for whichever algorithm it is we want to run. Unfortunately, power limitations are limiting how many of these transistors a chip can use at any one time. As a result, energy efficiency is the critical metric which will influence whether future vision algorithms are viable solutions to the real-time vision problems.
This research focuses around the design of Stingray, a chip with many massively specialized, diverse kinds of processing cores, which is tuned for maximal energy efficiency in vision processing applications. The project explores these and other architectural challenges that arise in designing effective low power Stingray systems. It exposes students to a broad range of issues including parallel compilation, vision, and hardware design. It seeks to capture the properties of vision applications to build a complete, prototype vision processor.
This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).
