Japan's 10 Petaflop SPARC Super Starts Shipping for 2012 Deployment

Print Friendly, PDF & Email

This week Fujitsu began shipping computing units for a 10-Petaflop “K” Supercomputer based on SPARC 64 VIIIfx processors. Jointly developed with RIKEN, an independent research institution funded by the Japanese government, the system is being delivered to RIKEN’s Kobe-based computational science research facility and is expected to begin operations in autumn 2012.

Rendering of the Japans 10 Petaflop Super, to be deployed 3Q2012

Rendering of Japan's 10 Petaflop "K" Supercomputer, to be deployed 3Q2012

The K supercomputer will comprise more than 800 computer racks housing a total of 80,000 SPARC 64 VIIIfx processors developed by Fujitsu. With a peak performance of 128 gigaflops, the SPARC 64 VIIIfx produces 2.2 gigaflops per watt, a reduction of power consumption by 2/3 compared to previous generations of the chip.

To provide high bisection bandwidth and fault tolerance, the interconnect for the 640,000-core system will be the “world’s first six-dimensional mesh-torus topology” developed by Fujitsu. Water cooling will reportedly enable high-density packaging along with improved component life and reduced failure rates.

This is going to be one massive installation. If any of our readers know how many Megawatts this baby is going to burn, please let us know in the comments.

Trackbacks

  1. […] Japan’s 10 Petaflop SPARC Super Starts Shipping for 2012 Deployment via insideHPC by Rich Brueckner on 9/29/10 […]

  2. […] Japan’s 10 Petaflop SPARC Super Starts Shipping for 2012 Deployment (insidehpc.com) […]

Comments

  1. Regarding power, simple arithmetic gives 4.5 MW for the processors alone. A supercomputer rule-of-thumb is that the cores are approximately half of the total power consumption, with a great deal of variability for how much DRAM is attached to each node, air versus water cooling, and how many spinning disks are connected. It is not unreasonable to assume 10-15 MW for the whole installation.

  2. The QCDOC supercomputer had a 6-dimensional mesh interconnect back in 2003. (http://phys.columbia.edu/~cqft/qcdoc/qcdoc.htm).

  3. If we take the thumb rule that would be around 1 GF/W for the total system. Sounds good however 10 PF in 2012 probably won’t be TOP5, what do you think?