The Inquirer-Home

Fujitsu delivers a 10 petaflops supercomputer

Some assembly required
Tue Sep 28 2010, 16:15

JAPANESE SUPERCOMPUTER MANUFACTURER Fujitsu has delivered its latest and greatest number cruncher to the Japanese Next Generation Supercomputing project.

The firm has shipped out 800 racks of servers containing 80,000 SPARC 64 VIIIfx processors. The chips were developed by Fujitsu, with Oracle licensing the Sparc architecture. Apparently the chips achieve 128Gflops each and are energy efficient, coming in at 2.2Gflops per Watt.

Those figures are pretty impressive when you consider that the top performing supercomputer on the Green 500 list tops out at 773.38Mflops per watt. However the chips are not the only component that consumes power and the interconnect should bring the overall energy efficiency figures down. Fujitsu is saying that it will deploy the world's first "six-dimensional mesh-torus topology" to interconnect the 80,000 Sparc chips.

At present both Fujitsu and its partner, Riken, have yet to bestow a name on this 80,000 chip Sparc monster. Most supercomputers now come with some edgy name like a go-faster stripe, such as Dawning Nebulae, Roadrunner or Jaguar, though with Fujitsu saying that the system won't reach full processing power until Autumn 2012, there's more than enough time for its Japanese government customer to come up with a suitably awesome moniker. Maybe it should hold a nationwide competition, with prizes.

If the system does hit its theoretical maximum computation power of 10Pflops (petaflops), that would comfortably put it in first place in the prestigious Top 500 supercomputers list. One petaflops is a measure of computing speed that represents 10^15 floating-point operations per second (FLOPS), in other words a quadrillion or million billion FLOPS.

However with two years left before Fujitsu can make its assault on the top spot, it is likely that the current title holder, the AMD Opteron shod Jaguar supercomputer rated at a peak theoretical speed of a little over 2.3Pflops, will increase its capacity or be surpassed by Dawning Nebulae, an Intel Xeon and Nvidia Tesla hybrid that is currently in second place.

The current vogue in supercomputers is to use GPGPUs such as Nvidia's Tesla or ATI's Firestream to deliver extremely high compute power and energy efficiency. Fujitsu's machine uses traditional CPUs and its 'next generation' machine could be the last that uses solely general purpose processors to try to break high performance computation records.

The next Top 500 list is expected to come out in November and will show whether a machine that includes Nvidia or ATI GPGPUs has managed to top the list for the first time. µ

 

Share this:

Comments
Yes traditional CPU

Each CPU has eight cores and each core is capable of four-way double-precision FMAD (eight FLOPs per clock). The CPUs are clocked at just over 2 GHz so that's just over 128 GFLOPS per CPU package.

posted by : Starglider, 15 April 2011 Complain about this comment
100 tereflop using AMD

Appro Deploys 324-node/100 Teraflops System to San Diego Supercomputing Center

http://www.hpcwire.com/offthewire/Appro-Deploys-324-node100-Teraflops-System-to-San-Diego-Supercomputing-Center-102351009.html

posted by : HPC Guru, 29 September 2010 Complain about this comment
Traditional CPU?

I dont see how a "traditional" CPU can acheive 128 GFlops/sec. Even if you assume its clocked at 4GHz, thats 32 Flops per clock.

posted by : Andromeda, 29 September 2010 Complain about this comment
aboutus
Advertisement
Subscribe to INQ newsletters
Advertisement
INQ Poll

Facebook starts selling shares

Will you buy Facebook shares?