John Gustafson – Winner of inaugural Gordon Bell Prize, Former Senior Fellow, Chief Product Architect, AMD; Former Director, Intel Labs (USA)

John Gustafson

Keynote – III Multicore World

A New Kind of Parallelism: The Ubox Method

John Gustafson, CTO Ceranovo Inc.

——————

“It is time to overthrow a century of numerical analysis. Current methods are based on the acceptance of rounding error and sampling error using numerical representations that were invented in 1914 and algorithms designed for a time when transistors were expensive. The pursuit of exascale floating point is ridiculous, since we do not need to be making 10^18 sloppy rounding errors per second; we need instead to get provable, valid results for the first time, by turning the speed of parallel computers into higher quality answers instead of more junk per second. The ubox method, based on a new numerical format that uses metadata to store more information using fewer bits, creates a the richest source of data parallelism since the Monte Carlo method, and redefines what is meant by “high performance”. Examples are given for practical application to structural analysis, radiation transfer, the n-body problem, linear and nonlinear systems of equations, and Laplace’s equation that suggest that the ubox method is general and can replace the ‘bag of tricks’ we currently use to solve technical computing problems.”

——————————————————————————-

Bio

“Dr. John Gustafson is an applied physicist and mathematician best known for his work in High Performance Computing, having introduced cluster computing in 1985 and having first demonstrated scalable parallel performance on real applications in 1988, for which he won the inaugural Gordon Bell Award. ‘Gustafson’s Law’ is taught in computer science courses everywhere. He was Director of the Extreme Research Lab at Intel, where he led their exascale computing program and other energy-efficient computing research. At Sun Microsystems, he led a $50 million supercomputer project centered on breaking “the memory wall”. He is the recipient of the IEEE Computer Society’s Golden Core Award. An honors graduate of Caltech and Iowa State University, he has held a number of CTO and CEO positions at publicly held and startup companies, and most recently was Chief Product Architect at AMD, before accepting the CTO role at Ceranovo Inc. in November 2013.”

——————————————————————————-

———————————————————————————

Preliminary work re Ubox at the Silicon Valley IEEE Computer Society Chapter, March 2013 – video – slides

Comments from John Gustafson about the discussion at the Multicore & Parallel Computing LinkedIn Group (~9,000 members) that presented those slides:

“I hope people realize that they are going to see material that has never been presented anywhere else, and not just another presentation of the unum storage format. This is going to be about applications and parallelism.”

Just one example of something I think will really startle people: Imagine doing a nonlinear ordinary differential equation with time stepping, like the (exact) equation of a pendulum swinging.”

“You get rounding error as you march through the time steps, and sampling error since you approximate variable acceleration and velocity with constants, but worst of all, you HAVE to perform the calculation sequentially. With my new method, you turn the problem sideways and compute the time spent in each region of space, since velocity and acceleration are functions of spatial position. Time is a function of position, instead of the usual other way around. Guess what that lets you do? Compute all the behavior in parallel! I can solve a nonlinear ODE with a rigorous bound, no rounding error, no sampling error, and DO IT WITH A MILLION PROCESSORS for more speed and accuracy, if I have that many.” (January 28, 2014)

———————————————————————————————————-

Advertisements