IBM breaks petaflop record - now, what about the exaflop?

Now that IBM has broken supercomputing's petaflop barrier with its RoadRunner system, capable of more than one thousand trillion (one quadrillion) sustained floating-point operations per second, attention among supercomputer developers turns next to a new performance goal: an exascale system.

An exaflop is a million trillion calculations per second, or a quntillion, and is a thousand times faster than a petaflop. It is the next obvious headline-making goal for the developers of the world's fastest supercomputers.

IBM announced this weekend that it had broken the petaflop barrier with a system at the Los Alamos National Laboratory, dubbed the RoadRunner. This system cost US$100 million.

To get some sense of what's next in line in terms of compute performance, you need to turn back the clock 11 years to the development of Sandia National Laboratories' ASCI Red supercomputer. That system, which was as heralded in its day as the RoadRunner is now, was the first computer to break the teraflop barrier -- one trillion calculations per second. That system cost US$55 million. Today there are blade systems priced in the low six figures that are capable of speeds close to ASCI Red.

At next week's International Supercomputing Conference in Dresden, Germany, Jack Dongarra, a professor of computer science at University of Tennessee and a distinguished research staff member at Oak Ridge National Laboratory, will be giving a presentation on exaflop systems "in the year 2019."

In an telephone interview, Dongarra said the performance gains are so far following a predictable path, with the first gigaflop system (a billion calculations per second) arriving 22 years ago (the Cray Y-MP computer developed for Lawrence Livermore National Laboratory), followed by ASCI Red.

"The projection is very clear -- in 11 years we will have an exaflop," said Dongarra, who believes by then every system on the Top500 computing list will be at least a petaflop. Indeed, the performance reached by the ASCI Red system isn't sufficient to make the most recent iteration of that list, which is updated twice yearly. (Dongarra is a co-compiler of that well-known list.)

RoadRunner is expected to be on the top of the Top500 list when it is released next week.

But Dongarra said there's some question about whether an exaflop-capable system can be developed at the same pace the petaflop system was. "It will take an incredible amount of effort," he said.

Dongarra said that IBM's exceptional achievement in RoadRunner is the programming that allowed the system to utilize different processor technologies -- AMD Opteron, and Cell. "It's a tour de force of programming," said Dongarra.

But to achieve exascale systems, Dongarra said developers will have to create new programming languages and algorithms that can calculate at high degrees of concurrency to complete calculations rapidly. The difficulty in reaching that level of programming, and in moving to new methods, could well be the stumbling block that prevents exaflop systems from being realized in a similar timeframe, he said.