Well over a decade ago, many believed that an engine of growth driving the semiconductor and computing industries, captured nicely by Gordon Moore’s remarkable prophecy (Moore’s law), was speeding towards a dangerous cliff-edge. When faced with this issue, I decided to consider a different approach back in 2002—one which suggested falling off the metaphorical cliff as a design choice, but in a controlled manner. This resulted in devices, circuits and architectures that switch and produce bits of information that deviate from the intended ``correct’’ specification—the results would be approximate or inexact. Despite being inexact by trading accuracy away, these hardware designs have been shown to be significantly more efficient in the energy they consume, their speed of execution, and physical area needs. This in turn makes them attractive for a range of large-scale data-centric applications such as analytics, search, learning and climate modeling, as well as embedded applications involving video and audio. In this talk, I will start with our beginnings with inexactness from 2002—one that Technology Review labeled as being heretical in their TR10 citation—and give an overview of a range of ideas that our group as well as other groups around the world have been developing since. I will also outline opportunities for applying inexact design in the context of realizing extremely energy-efficient platform architectures including memory systems, for computing with big data at one extreme, and battery-constrained embedded systems with strong ties to neural processing at the other. We intend to use this talk as a basis for developing partnerships to collaboratively help grow this exciting frontier in inexact hardware, and eventually software systems.
Tuesday, February 18, 2014
Free and open to the public