China has at last surpassed the US with the world's fastest supercomputer. The Tianhe-1A achieved 2.507 Petaflop, surpassing Cray's 2.3 petaflops Jaguar, built at the Oak Ridge National Lab in Tennessee.
An important part of this news was the design of the system. As with many recent Supercomputer efforts, designers moved away from either custom CPU's or 10's of thousands of generic CPUs. Instead the designers concentrate on melding traditional CPUs and NVIDIA's Tesla M2050 compute GPUs. This moved allowed then to use only 14,336 CPUs and 7,168 NVIDIA GPU's. It's estimated (by NVIDIA) that without the GPU horsepower, the system would have required over 50,000 CPUs.
Again this shows that design considerations are as important as simply throwing money at a problem. This also shows a newer trend in supercomputer design; Power. By using the GPU/CPU combination, Tianhe-1A uses 4.04Mw of power.`It's estimated that a similar system with only CPUs would have consumed over 12Mw. That makes it 3 times as energy efficent that it's nearest competitor.
We'll be interested in learning more about it's design in the coming months, and what it brings to the future design of Supercomputers. You can find more coverage at Engadget and NVIDIA.