Title: When Close is Good Enough: Exploiting Randomness for Highly Reliable Approximate Computing
Speaker: David J. Lilja, Professor and Head, Department of Electrical and Computer Engineering, University of Minnesota
Abstract: The continued scaling of device technologies to smaller and smaller feature sizes introduces greater variability, defects, and noise into the circuits. As a result, it is becoming increasingly challenging to maintain the conventional deterministic Boolean computation model that rigidly transforms binary inputs into binary outputs, such as integers into integers or floating-point values into floating-point values. While this abstraction effectively hides variability and errors at the circuit level from the application program, it is expensive to detect and correct bit-level errors using system-level techniques, such as modular redundancy and error-correcting codes. Instead of forcing rigid, deterministic operations at each level of the abstraction hierarchy, our approach treats randomness as a valuable computational resource by conceptually transforming probability values into probability values. Our representation operates on stochastic unary bit streams using conventional logic gates to naturally tolerate substantial noise and circuit variability. While techniques for probabilistic analysis are well established, we propose new synthesis techniques to approximately compute complex operations directly in the stochastic domain. We demonstrate with several image processing applications how this conceptual shift produces circuits that are remarkably tolerant of errors while maintaining cost and performance comparable to – and in some cases even better than – conventional deterministic implementations.
Speaker Bio: David J. Lilja received a Ph.D. and an M.S., both in Electrical Engineering, from the University of Illinois at Urbana-Champaign, and a B.S. in Computer Engineering from Iowa State University in Ames. He is currently the Louis John Schnell Professor of Electrical and Computer Engineering at the University of Minnesota in Minneapolis, where he also serves as the ECE department head, a member of the graduate faculties in Computer Science and Scientific Computation, and a Fellow of the Minnesota Supercomputer Institute. Previously, he worked as a research assistant at the Center for Supercomputing Research and Development at the University of Illinois, and as a development engineer at Tandem Computers Incorporated in California. His main research interests include computer architecture, parallel processing, computer systems performance analysis, and high-performance storage systems. He is a Fellow of both the Institute of Electrical and Electronics Engineers (IEEE) and the American Association for the Advancement of Science (AAAS) for contributions to the statistical analysis of computer performance.