Title: A Probabilistic, Asynchronous System for Ultra-Efficient Performance

Speaker: Harry Marr, PhD, Georgia Institute of Technology

Date/Time: Monday, January 12, 2009, 11:00 am – 12:00 noon

Location: Bldg. 898, Room 1020

Brief Abstract: Motivated by the many stochastic processes found in biology that allow for ultra-efficient computing, this talk explores circuit implementations for ultra-efficient probabilistic and asynchronous computation in hardware. Several novel contributions are presented, namely a dynamically controllable network of probabilistic CMOS (PCMOS) gates that not allow ultra-efficient stochastic computation. It will also be shown that for the first time, probability will be linked to energy consumption in digital circuits. Furthermore, asynchronous techniques will be shown that show the same effect. A system with these technologies is implemented on a reconfigurable chipset allowing for the first time ever the calculations of a complex, dynamical system to be accelerated with an asynchronous, stochastic hardware system. The utility of this system is demonstrated by implementing the well-known Gillespie algorithm for simulating an arbitrary system of chemical reactions where over a $127X$ performance improvement over current software approaches is shown. These results are then extended to show how over a 1500X performance gain can be attained.

CSRI POC: James A. Ang, (505) 844-0068