The soul of a sino-supercomputer

Heavy-duty processing

Graphics chips are credited with providing heavy-duty processing for relatively little power. The Nebulae system consumes about 2.55 megawatts, compared with about 7 MW for Jaguar, the fastest system on the Top 500 list.

Interestingly, Jaguar requires more than twice the power of Nebulae but provides less than 50 percent more performance. In addition, Nebulae sports theoretical maximum performance of 2.89 Pflops, compared with 2.3 Pflops for Jaguar.

"The performance per power is one of the biggest parts of the graphics story," said Sumit Gupta, manager of Nvidia's Tesla products, in comments on the trend toward high-end computing with graphics chips.

Click on image to enlarge.

Currently, the average power consumption of a Top 500 system is 397 kW; average power efficiency is 195 Mflops/W, up from 150 Mflops/W a year ago. The most power-efficient system on the latest list also uses graphics chips; the so-called QPace Cluster, based on IBM PowerXCell 8i processor blades, was rated at 774 Mflops/W.

Nebulae and Tianhe-1 helped catapult China into second place in installed supercomputer performance, at 9.2 percent, though that's still far behind the United States' 55.4 percent. China now has 24 systems in the Top 500 and is tied for fourth in the systems count with Germany, whose numbers have been declining steadily.

The Chinese researchers got plenty of support from U.S. semiconductor giants. Intel provided its Server System Infrastructure, a mechanical and software framework, for Nebulae. Nvidia provided tools to tune the Linpack performance of its Tesla chips.

China's engineers want to build even more powerful systems-next time with their own homegrown silicon. Dawning has long promised a series 6000 system using a version of the China-designed Longsoon-3 (aka Godson-3) processor, which packs 16 cores and 8 Mbytes of L2 shared cache.