China’s Tianhe-9B has trumped America’s Oak Ridge Labs in the race to exascale computing, with a full 1,000 petaflop record.
Despite all their tweaking, Zeus, the Oak Ridge supercomputer, was able to clock only 887 petaflops. But this is already a quantum leap beyond the 33 petaflop benchmark of Trinity, which held the trophy as world’s fastest computer just five years ago.
This type of exponential increase in speed required a completely different base architecture, as well as the latest multicore GPUs or graphic processing units. Originally designed for smooth gaming applications on PCs, these GPUs are not particularly smart, but can fly through repetitive bit crunching at remarkable speed, without melting, thanks to their nanotech fabrication.
China’s winning design relies not only on brute size at any cost, but turns traditional computer topology on its head, placing the massive 3D memory bank in the center, and the processor nodes on the periphery. Perhaps counter-intuitively, this radical innovation has resulted in higher overall performance.
“They appear to have almost bent the laws of physics,” says Mike McCoy of Lawrence Livermore National Laboratory, “dialing up the speed without melting the cores. At some point you can’t get rid of the heat fast enough, no matter what cooling system you use!”
Besides the unusual layout of China’s exaflop computer, little is known about the design specifics; but the performance claim has been independently verified. On the other hand, Google’s chief technologist has dismissed the giant machine as “purely academic,” claiming that they process vastly more data on a daily basis, using cheap and efficient modular server farms.