Tokyo

Japanese groups are competing to reclaim for their country the honour of owning the world's fastest supercomputer.

Japan lost the title last autumn, when IBM's BlueGene/L computer — which has a maximum speed of 140 million million calculations per second (140 teraflops) — overtook the Yokohama-based Earth Simulator (see Nature 431, 618; 200410.1038/431618a).

Last month, a group that includes Toshiba, the University of Tokyo and the Institute of Physical and Chemical Research announced plans for a computer that would be 73 times faster — capable of 10,000 million million calculations per second (10 petaflops). It said that the project could cost between ¥80 billion (US$700 million)and ¥100 billion, and would be ready by March 2011.

But experts are already expressing doubts. Tetsuya Sato, director of the Earth Simulator, says the project will be held back by the limits on semiconductor technology, which is reaching the narrowest dimensions of circuits that can be etched. “I am very sceptical,” he says.

Erich Strohmaier, a computer scientist at Lawrence Berkeley National Laboratory, says difficulties with increased heat output and power consumption would also hamper the development of faster processors. But he predicts that technological advances could produce a machine of at least 3 petaflops by 2011, and doesn't rule out the Japanese team reaching their goal. “It's an ambitious project,” he says. “With an exceptional system design, 10 petaflops seems within reach.”

Meanwhile, Sato has his own plan — hooking the 40-teraflop Earth Simulator to a new 400-teraflop computer. The smaller machine would act as a ‘macro-structure’, performing large-scale calculations and directing the more intricate operations of the larger one, he told Nature.

Sato thinks that changing the networks' wiring in this way would allow the equivalent of 16 petaflops. “We are close to the technological limit,” he says. “But this solution is based on physics, not technology.” Strohmaier warns, however, that such hierarchical structures are a risk because they sacrifice a great asset of the Earth Simulator — “a tried and tested technology, which we understand and know how to program,” he explains.

This computing power, if achieved, could be used to model climate change, drug metabolism and galaxy formation, for example. But it would be very expensive. It is not clear which of the projects, if either, will make it into the Japanese science ministry's budget requests, due this month.