The gigaflop that thinks it's a teraflop: A small British supercomputer company has captured one of America's most prized contracts with its 'monster machine'. Greg Wilson explains

Click to follow
ONCE upon a time, a small group of British engineers founded a company to manufacture and market a new type of supercomputer. The City ignored it, preferring, in the early Eighties, to put its money into safer ventures such as property development. The government of the time, while keen on entrepreneurial spirit, declined to support it.

Meanwhile, governments of less enlightened countries were backing their supercomputer manufacturers generously. Despite its lack of support, the tiny company prospered. Although founded in Britain, it became an important player in the American market. So much so, that last month, one of the leading US research laboratories decided to buy one of its machines, rather than a computer from an American supplier.

A fairy-tale? Given the British habit of failing to turn technological innovation into commercial success, it certainly sounds like one. However, the story is true. The Bristol-based company, called Meiko, was founded in 1985 to make supercomputers, which offer much faster and more powerful computing than conventional machines.

The founders of Meiko were part of a team recruited by the semiconductor manufacturing company Inmos. Set up by the last Labour government in 1979, Inmos created a completely novel 'computer on a chip' - the transputer.

The founders of Meiko had a simple idea. Instead of building a supercomputer in the traditional way, around a single, fast processor, they would combine dozens, or hundreds, of inexpensive transputer chips to achieve the same power at a fraction of the cost. Such designs are called 'parallel computers' because their processors work simultaneously to solve a single problem.

Several other firms, such as Parsys in the UK, Telmat in France and Parsytec in Germany, were set up to exploit the new transputer chips in the same way. In the years that followed, however, Meiko proved more successful than its competitors.

But it gradually became clear that Inmos was not going to be able to sustain its early momentum in designing and manufacturing commercially useful transputers. So Meiko started using other vendors' chips to increase their power. At the same time, the company slowly moved much of its operations to the US to be closer to its major markets.

Most important, despite the recession, Meiko has produced a second generation of the machine. This does not contain transputers, but uses the SPARC microprocessor designed by Sun Microsystems for their work stations.

There can be dozens or hundreds of SPARCs in Meiko's new machines, depending on the computing power required. Each can be coupled with two small arithmetic processors manufactured by the Japanese giant Fujitsu and with special-purpose chips produced by Meiko.

This combination has proved so effective that the Lawrence Livermore National Laboratory, outside San Francisco, has decided to make its next major supercomputer purchase from Meiko. It will be the first time the laboratory has bought a parallel computer, rather than a more traditional design, to support the bulk of its work. The Meiko design will be one of the fastest and most powerful supercomputers in the world.

The Livermore was originally set up for nuclear weapons research, but has since branched out into environmental modelling and energy production and conservation.

Along with Nasa, the National Security Agency and the Weather Bureau, US government laboratories have a tradition of buying the latest supercomputing hardware and squeezing the ultimate in performance out of it. They also have bought only from a select few firms, all US companies. .

The Livermore's surprise announcement brought fierce criticism in the US, most notably in an article in the New York Times last month, which accused Meiko of selling computers to Israel for use in nuclear weapons design. The article neglected to mention similar sales by American companies. In fact, Meiko's sale was not for weapons work and was approved by British authorities.

The article also accused the Livermore of spending American tax dollars on the lowering of national security. Less comical were allegations that Meiko lacked a sound track record (despite the fact that it has installed more than 400 machines in seven years). An unnamed executive from another manufacturer said: 'It's one thing to lose a sale in Europe, it's quite another to lose at a national laboratory.'

This name-calling may reflect a growing unease about the focus of supercomputing research in the US. The goal of many firms, and of much government funding, is production of a so-called 'teraflops' machine. The term 'flops' refers to 'floating point operation per second' - roughly speaking, the addition or subtraction of two decimal numbers in one second.

The goal of supercomputing in the US is to produce machines capable of adding together one million million numbers in less than a second. The most powerful supercomputer today can perform 20 to 25 gigaflops (a billion floating point operations a second). The teraflop machine would be at least 40 times more powerful than anything in existence and would allow scientists to model the Earth's climate, or design pharmaceuticals and aircraft, much faster and with greater accuracy.

However, an increasing number of computer scientists and industry figures doubt the importance of such 'monster machines' in the long term. Teraflop machines will cost tens of millions of dollars and few commercial firms are willing to pay such a price.

The 'traditional' design of supercomputers is inflexible and can cater only for the speed specified by a customer. A parallel computer uses dozens or hundreds of chips to split up the calculations, so while the manufacturers can provide supercomputers to companies that can afford them, they can also scale down the design to provide less powerful, but more affordable machines.

Encouraging industry to use supercomputers in large numbers, and developing software capable of harnessing their power effectively will ultimately have more economic impact than production of a few prestigious machines.

In the Eighties, parallel computers had to prove they could compete with 'traditional' supercomputer designs. The Livermore's decision makes it clear they have succeeded. Its purchase also shows that a small number of companies that once had a guaranteed customer in the US government will now have to compete on the same playing field as everyone else.

The author researches parallel computing at Edmonton University, Canada.

(Photograph omitted)