parallel 2013 » Agenda »

// Keynote: Exascale Arithmetic

The reason we use computers is that they're fast, and the reason we use parallel computers is that they are faster than serial computers. But this begs the question of what it means to be "fast." The traditional definition of how many double-precision floating-point operations the system can do per second is becoming quite disconnected from the goal of being able to solve the largest computing problems. What if we were to re-examine the most fundamental aspect of computing, that of performing arithmetic on numerical data, and discover that speed and parallelism could be exploited to improve the quality of the results, not just the rate at which we churn out rounded calculations? We are attempting ever more ambitious calculations with ever less attention to numerical analysis, and blindly using double-precision as the panacea; sometimes double-precision is insufficient, but most of the time it is excessive, wasting storage, bandwidth, and electrical energy and power that make it harder to achieve an exascale supercomputer. We similarly often use excess precision in our integers, simply because of programmer convenience. A new numerical format, the "unum" (universal number) can achieve higher accuracy with fewer bits and less programmer effort than the century-old floating-point arithmetic, and bring the goal of exascale (not exaflops) much closer.

// Referent

// Dr. John L. Gustafson

is Senior Fellow at AMD, where he serves as Chief Product Architect. He is a 35-year veteran of the computing industry, and comes to AMD from Intel, where he directed the company’s Extreme Technologies Lab. Prior to that, he served as CEO at Massively Parallel Technologies, and CTO at ClearSpeed Technology, a high-performance computing company. John has also held key management and research positions at numerous companies including Sun Microsystems, Ames Laboratory, and Sandia Laboratories. John holds an M.S. and Ph.D. in applied mathematics from Iowa State University, and a B.S. in the same from Caltech. He holds numerous patents and has authored an extensive array of technical publications.