The Fujitsu-designed K Computer, announced as winner of the biannual Top500 supercomputer list on Monday at the International Supercomputing Conference (ISC) in Hamburg, is three times faster on the Linpack high-performance computing (HPC) benchmark than last November's winner, the Tianhe-1A.
K Computer is capable of over eight quadrillion calculations per second with an Rmax score of 8.16 petaflops. This makes it more powerful than the next five systems in the top 10 combined.
Unlike previous winners, K Computer does not use graphical processing units (GPUs), such as those made by Nvidia, nor does it use x86 processors from Intel and AMD; instead, is it based on Fujitsu-designed Sparc processors.
K Computer is based at the independent Riken research lab in Japan. Riken caries out research in physics, chemistry, medical science, biology and engineering. Japan was last home to the world's most powerful Top500 system in 2004, with the 36-teraflop Earth Simulator.
When it comes to power consumption, the Japanese system is incredibly hungry yet surprisingly efficient. It pulls down over 9.8MW of power to make its computations, yet is the second most-efficient system on the entire Top 500 list, consuming around 1KW of power for every 800 gigaflops of computation. By comparison, hosting and cloud company Rackspace's main UK datacentre consumes 3.3MW.
The power consumption is "not because of inefficiencies, it's simply driven by its size. It's a very large computer", said Hans Meuer, general chairman of the ISC.
In the new Top500 list, Tianhe-1A came second and US Oak Ridge National Lab's Jaguar came third. The rest of the top 10 was rounded out by four other US computers, alongside systems from China, Japan and France.
Sparc back, x86 relegated
K Computer uses 68,544 eight-core 45nm Sparc64 processors, for a total of 548,352 processing cores.
"The Sparc architecture breaks with a couple of trends which we have seen in the last couple of years in HPC and the Top500," Meuer said. "It is a traditional architecture in the sense it does not use any accelerators. That's the difference compared to the number two and number four [systems], which both use Nvidia accelerators."
The Sparc architecture breaks with a couple of trends which we have seen in the last couple of years in HPC and the Top500.
– Hans Meuer, ISC
Additionally, the Japanese machine uses an HPC six-dimensional mesh-torus network to pass information between processors. The network topology means that every computing node is connected to six others, and individual computing jobs can be assigned to groups of interconnected processors.
This mesh structure of networking is becoming more prevalent in the enterprise space, with dense server maker SeaMicro applying a torus network to its servers and many-core chip designer Tilera applying a mesh structure to its on-chip communication topology.
K Computer is still being developed and is expected to be completed in June 2012, at which point it will have around 15 percent more processors.
Based on developments such as K Computer, Meuer said the ISC expects a system to reach an exaflop by around 2019.
China is an outlier in the HPC field, Meuer noted, explaining the country "used to be virtually absent from the Top500 in the 1990s" but it now has two systems in the top 10. "China is the exception here," he said.