Jack Dongarra, a computer science professor at the University of Tennessee and the keeper of the list of the world's top 500 supercomputers, says these elite machines typically cluster 500 to 1,000 processors. The current top gun, the IBM Blue Gene at Lawrence Livermore National Laboratory, can process 280 trillion floating-point operations per second (TFLOPS) and clusters a staggering 131,072 processors. Those clusters must rapidly move huge quantities of information from memory to all those processors and back again. The bandwidth needed for this data flow, along with the latency caused by the sheer distances involved, effectively caps the amount of work the computer can do.

Dongarra says that more than 60 percent of the top 500 computers are clustered rather than relying on the traditional exotic architectures most commonly associated with Seattle-based Cray Inc. "Clusters have completely changed the scientific computing landscape," he says, because they offer a price/performance ratio that exotic machines can't touch.

Moreover, as clusters have become popular, users have found "a surprisingly large number of real-world applications that do not require the extreme latency and bandwidth capabilities of the exotics," says Justin Rattner, a distinguished fellow at Intel Corp.

However, as users call for more powerful tools, Cray executives believe the supercomputing pendulum is swinging back their way -- and some research scientists agree. Indeed, they say, it's possible that before 2015, exotic supercomputers, with their benefits of low latency and high bandwidth, will join clusters, with their price advantage, in hybrid architectures suitable for a variety of applications.

According to Jan Silverman, a senior vice president of corporate strategy at Cray, the company is working on compilers that can distinguish code best suited to its vector processors -- which can operate on a whole string of numbers at once -- from code best run on a more pedestrian scalar processor. These compilers, which Silverman says will be in production by 2009, will be able to schedule vector and scalar work for the hybrid supercomputer. Otherwise, this slow and difficult scheduling work would fall to programmers.