Many of my former colleagues at Cray should be very proud of their achievements when they read about the latest and greatest supercomputer announced by IBM. Without their efforts in working out precisely how thousands of processors could be made to work as if they were one, we would still be struggling to get computer power much above where it was two decades ago when single-processors dominated. It's much like bureaucracy on a huge scale. Putting two bureaucrats together doesn't double the amount of work that gets done, and adding a thousand more doesn't increase it a thousand fold either.

Lead engineer Don Grice of IBM inspects Roadrunner. Copyright IBM

Some very clever people had to figure out how to split one job between thousands of processors; how to get input data in to each of them and the output data back out; and keep every single one simultaneously running flat-out. If you think running one processor or one bureaucrat flat-out would be hard enough, imagine the challenge of doing that for the 116,640 processors in the new IBM machine. The methods devised by Cray hardware and software engineers are being used today thanks to some far-sighted decisions, such as adopting Unix as the core operating system and making the solutions they developed freely available to other developers.

I guess the only regret I have about the new machine is that it is so ugly. It has none of the elegance of a Seymour Cray designed system, which in fact was merely a by-product of his engineering brilliance. Cray systems were designed with an optimum physical layout so cable runs, for instance, would be as short as possible. The closer together components were, the quicker data could be passed between them. The rest of the machine was simply packaged around the core components in the most efficient way possible and the final shape emerged as a result. And what a result. The Cray-1 and the Cray-2 were visually stunning pieces of engineering design, albeit expensive. A top-end Cray system would set you back up to $35 million, and, as Dudley Moore quipped in "Arthur", that was when $35 million was a lot of money. The new IBM supercomputer cost $133 million.

But if you wait twenty years, that power will be available to you as a desktop computer at a desktop price, just as the power of a Cray-1 is available to you now. In fact, a typical PC with a dual processor chip, 1 gigabyte of RAM, and a hard disk with 160 gigabytes of storage has, on the face of it, a better spec than a single processor Cray-1 with a mere 8 megabytes of RAM and a few hundred megabytes of disk space. The difference is that the Cray-1 was something called a vector processor and had extra circuits for floating point arithmetic and higher memory bandwidth. This meant it could produce two floating point answers per clock period whereas a PC would take several tens of clock periods for each result. The nearest supercomputer with the same spec on paper would have been a later generation two processor Cray X-MP which would have set you back around $5 to $10 million.

Performance is measured in what are called FLOPS, or FLoating-point Operations Per Second. A Cray-1 ran at a peak speed of 160 MegaFLOPS, or 160 million floating point operations per second. It was by far the most powerful computer when it was launched just over thirty years ago. Almost twenty years ago Cray broke through the gigaflop barrier with the Cray Y-MP, that is one billion floating point operations. The next target was the teraflop or a thousand billion flops which they broke about ten years ago with a Cray T3E. The new IBM supercomputer has just broken the petaflop barrier. That's a thousand trillion flops. Compared with a Cray-1, it is more than 6 million times more powerful.

Why is it worth spending so much money, why don't these establishments simply wait for performance to match their requirements? The short answer is that these powerful computers are like time machines. They give researchers the ability to solve problems that their competitors - another company or another country - will not be able to solve for a further five years or until their own computer capacity catches up. Having a new drug on the market five years ahead of your competitors, for example, or having the most fuel-efficient engine on the market for five years, will give a company a lucrative market advantage. Having a weapons system that is five years more advanced than a potential enemy gives you a battle-winning capability. The corollary is that allowing yourself to fall five years behind your competitors is a significant disadvantage to be avoided at all costs. So there is a big potential market for top-end computers.

If you want to think about it, imagine trying to do anything you take for granted today on a twenty year-old desktop computer. Then imagine what you could do if you had a desktop computer beamed back to you from twenty years into the future. That's what the lucky users of this new computer have right now. Top-end supercomputers really are time machines.

Full story on the IBM web site here

By the way, this one is called "Roadrunner" because it is going to be installed at the Los Alamos National Laboratory in New Mexico, and the Roadrunner is the state bird of New Mexico. I think, since it is so ugly, it should be named after the state flower instead - the Yucca.

And finally, just to remind you what supercomputers used to look like, here is a Cray X-MP of the sort I used to work on: