From Christopher Lazou
Will Knight highlights the cheapness of supercomputing clusters (6 December, p 28). But this is a myth.
The initial capital price of clusters is usually a tenth of the price of a supercomputer, but when running costs are included in the calculation, studies have shown that over a five-year life cycle these PC clusters turn out to be more expensive than their superior vector-processing cousins, such as the NEC SX6 and the new Cray X1.
In addition, modern vector parallel supercomputers are not dinosaurs: they are complex to build and expensive, as they use cutting-edge technology amortised in a minuscule market. But so long as scientific computing uses von Neumann computer architectures, vector parallel supercomputers will remain at the top of the computing productivity pyramid.
This is why the US government, after spending billions of dollars in the 1990s promoting PC-based systems, has changed tack in the last two years. It now puts the emphasis on the raw speed of vector processing, and it is this move that has helped revive Cray, to the relief of many US engineering companies, such as Boeing, that need the power of a supercomputer.
Advertisement
London, UK
