ITA Software's Technical Seminar Series

Talk Topic: Approximate Computing: Useful Results 10,000x Faster
Joseph Bates, Carnegie Mellon University
September 23, 2009

Abstract

Modern computers have accustomed us to computing with high precision arithmetic. But a variety of tasks are solvable using approximate arithmetic along with mildly clever algorithms. This is significant because useful approximate arithmetic can take very little silicon, e.g. O(100,000) processors will fit on a programmable co-processor chip. We sketch a machine of this kind along with several example algorithms. Compared with CPUs, this approach appears to be able to bring 10,000x speedup or power reduction to some important tasks in vision, pattern recognition, search, optimization, computational science, computational finance, and other domains. 

Bio

Bates spent 10 years as a regular faculty member in Carnegie Mellon's Computer Science Department, where he remains an Adjunct Professor. He's also worked at Cornell, Johns Hopkins, and MIT, and led a spinout company funded by Fujitsu Limited based on his CMU work. He believes speedups from approximate computing may greatly accelerate progress in fields important to humanity. Singular Computing is commercializing an approach to approximate computing.