The bragging rights for building the world’s fastest supercomputer pass to Japan and Fujitsu’s K-supercomputer. For most people this is a fleeting tidbit of technology news, but it is one kind of milestone marking the increasing power of computers. For the computer industries in the countries involved, it is a rather big deal. In this case, the manufacturer is a private company not a government agency or academic organization, which is somewhat unusual. Fujitsu may not directly profit from the K-supercomputer itself, but in the commercial world bragging rights can be mighty influential. Just ask IBM.
The ‘numbers,’ always important in supercomputer contests, are impressive: The old champ, China’s Tianhe 1a, managed 2.507 petaflops. [SciTechStory: Tianhe 1a: China and the world’s fastest supercomputer] A petaflop is a thousand trillion floating point calculations per second. The new champ can do 8.2 petaflops. This is a smashing win. To achieve this, Fujitsu used 68,544 SPARC64 VIIIfx CPUs with a total of 548,352 chip cores – double the nearest competitor. The performance is roughly equivalent to a million desktop computers connected into one system. Only, of course, a million desktop computers would never perform like this. Truth is, the secret to a super-fast supercomputer is software, specifically hyper-specialized forms of network software. That’s where Fujitsu has taken the lead. Developing this kind of software is a monumental intellectual and practical task where even the most minute inefficiency in the programming can cost gigaflops.
The race continues, of course. Three U.S. computers (Oak Ridge National Laboratory, Cray, and IBM) have already been announced as 2012 entries that are expected to achieve about 20 petaflops and China is not sleeping.