快猫短视频

WHEN computer scientists at Virginia Tech鈥檚 Terascale Computing Facility needed a new supercomputer this year, they quickly realised they would have to cut a few corners. 鈥淲e talked to all the usual suspects,鈥� says associate director Kevin Shinpaugh. 鈥淏ut it was all out of our price range.鈥�

So they turned their attention to a less obvious piece of hardware: Apple鈥檚 Power Mac G5 desktop computer. If they could lash hundreds of these machines together, they reasoned, they might be able to build a formidable number cruncher at an affordable price.

That hunch came good last month, when the latest supercomputer 鈥淭OP500鈥� speed charts were released. Shooting in like a bullet was Virginia Tech鈥檚 G5 cluster 鈥� listed as the third most powerful computer in the world.

Virginia Tech鈥檚 Mac-based 鈥淴 System鈥�, also known as Big Mac, can perform 10.28 trillion calculations per second (10.28 teraflops). It tightly networks 1100 G5 machines, each containing two 64-bit microprocessors, and cost $5.2 million 鈥� a fraction of the cost of dedicated supercomputers of similar performance.

Big Mac鈥檚 success reflects a shift in supercomputer design. Colossal dedicated machines once ruled research labs around the world. But in the US at least, they have been steadily overtaken by less expensive machines built by networking clusters of cheap general-purpose computers. Nowhere is this trend more evident than in the TOP500 list. In the latest edition, clusters outnumber dedicated machines for the first time, and no fewer than seven of the top 10 machines are clusters. 鈥淣owadays off-the-shelf microprocessors are very clearly dominating,鈥� says Erich Strohmaier, a supercomputing researcher at the University of Tennessee who helps draw up the list twice a year.

Outside the US, however, what some see as the dinosaurs of supercomputing are far from extinct. At the Japan Marine Science and Technology Center, at the Yokohama Institute for Earth Sciences, NEC鈥檚 enormous Earth Simulator climate-modelling computer was switched on in April 2002. Its design appeared to be a throwback to an earlier age, but to the amazement of many supercomputing experts it recorded 35.86 teraflops in testing, making it more than five times as fast as the previous record holder. A year later it is still the world鈥檚 most powerful computer by far.

The Earth Simulator has 640 networked units each containing eight vector processors 鈥� chips that work on whole arrays of data at once rather than sequences. The complete machine, with its 5000-plus processors, occupies the area of three tennis courts and uses 2900 kilometres of cable for networking. Its power shocked the Americans. 鈥淭he Earth Simulator was a real wake-up call,鈥� recalls Steve Scott of Cray Research, the US dedicated supercomputer maker.

But Hamid Arabnia, editor-in-chief of the Journal of Supercomputing, was not surprised. Huge clusters of smaller computers do not necessarily guarantee high performance, he points out. The trick is in how they are wired together. Dedicated super-computers have architectures that maximise inter-processor communication speeds and memory accesses. This makes them better at tackling problems that cannot easily be 鈥減arallelised鈥� by dividing into smaller subsets. Many problems fall into this category, including biological modelling, climate prediction and some areas of mathematics. And clusters can be trickier to maintain and more prone to failure.

IBM is hoping to bridge the gap between the two approaches, however. Last month, it unveiled an ultra-compact supercomputer that can be bolted together to form a larger cluster. Each unit is the size of a small fridge and capable of 1.4 teraflops 鈥� putting it at 73 in the charts. By plugging together hundreds of these units, IBM plans to build a machine called Blue Gene that is 10 times as fast as the Earth Simulator. It will have 131,000 processors and deliver 360 teraflops, and is due to be switched on in 2005 at the Lawrence Livermore National Laboratory in California. It will model the way proteins fold into more complex structures.

But Blue Gene is venturing into uncharted territory. 鈥淲e don鈥檛 know if it can handle 100,000 processors,鈥� admits Strohmaier.