TechMan's interest was piqued recently when he saw a Web headline reporting that the Pentagon was buying 2,200 Sony PlayStation 3 gaming machines.

Entertainment for our troops overseas? No, expansion for a Department of Defense supercomputer.

It turns out the military already has a supercomputer subsidized by Sony and consisting of 336 PS3s. But when Sony dropped the price of the gaming machines to $299, the DOD decided to go for a big expansion.

PS3s contain Cell chips developed by a consortium of Sony, Toshiba and IBM. It turns out the Cell processor excels at several types of scientific processing.

Advertise with us

When the DOD did a study, it found that a server containing two Cell chips would cost $8,000, while two Sony PS3s with a slightly less powerful Cell chip would cost $600. Even for the thrift-challenged Department of Defense, this was a no-brainer.

The use of a gaming-machine chip in a supercomputer illustrates how thinking about powerful computers has changed. Specialized chips have given way to what is called "massively parallel computing" -- linking a large number of off-the-shelf processors together.

The current holder of the fastest supercomputer title is the Jaguar, a Cray XT5 machine at Oak Ridge National Laboratory, with 224,162 computer processor cores using AMD six-core Opteron chips. It computes at a speed of 1.759 petaflops or quadrillions of calculations per second.

Just last month it dethroned Los Alamos National Laboratory's IBM Roadrunner system with a speed of 1.04 petaflops. Roadrunner is built from AMD Opteron and IBM Cell chips.

So how far can this go? Talk at this year's supercomputing conference is of an exaflop system, or a million trillion (one quintillion) calculations per second. The time frame for this is around 2018, and such a computer could have up to 100 million computer processing cores.

Advertise with us

But some significant problems remain to be solved, a big one being power consumption. One estimate is that to operate an exaflop machine at today's levels of efficiency might take 2 gigawatts or 2 billion watts, roughly the output of a medium-sized nuclear power plant.

Obviously building a supercomputer out of readily available chips is not just a matter of throwing them into a box and plugging it in. The real tricks to parallel computing are adding the technology to allow the chips to communicate and dividing up the task so that as many chips as possible can work simultaneously on it. Programing for a supercomputer is a bear.

Supercomputers are best used for solving big problems with lots of data, such as weather forecasting, climate research and simulations of airplanes in wind tunnels, detonation of nuclear weapons and nuclear fusion. In fact, some supercomputers are built to specialize.

In the region, the Pittsburgh Supercomputing Center, a joint project of Carnegie Mellon, Pitt and Westinghouse Electric, makes supercomputing resources available to business, university and government researchers.