Japanese Supercomputer K Hits 10.51 Petaflops

coondoggie writes "The Japanese supercomputer ranked #1 on the Top 500 fastest supercomputers broke its own record this week by hitting 10 quadrillion calculations per second (10.51 petaflops), according to its operators, Fujitsu and Riken.
The supercomputer 'K' consists of 864 racks, comprising a total of 88,128 interconnected CPUs and has a theoretical calculation speed of 11.28 petaflops, the companies said."

Comments

Re:Skynet is bigger

[Plasmaphysiker]

consists of 864 racks, comprising a total of 88,128 interconnected CPUs

Where goes the border between a supercomputer and a cluster?

Communication time. Trying to run a massively parallelized plasma physics simulation on a mere cluster is essentially a waste of time. The scaling is terrible.

Re:Skynet is bigger

[jd]

Technically, a cluster can be a supercomputer if it is tightly-coupled, which basically means high bandwidth, low latency and as little overlap on the fabric as possible. (ie: 88,128 PCs linked via the Internet could be considered a grid but it would not be considered a supercomputer. The same number of PCs in a server room using a hundred or so switches, with each switch stuffed to the gills, would be considered a regular cluster. The same PCs in the same room using high-end switches linked as a Fat Tree, Butterfly or - ideally - a hypercube topology would be considered a supercomputer. The same PCs in the Cloud would be considered a torrential downpour.)

The problem is ultimately, as Plasmaphysiker says, communication time. From a technical standpoint you can just as easily say "a supercomputer is any computer that can mimic or better a vector processor's overall performance for the same compute power". Ok, maybe not as easily as it's longer to say, but it comes to the same thing.

K machine technology

[Required Snark]

The K supercomputer is build on SPARC technology.

The system is still under construction and is scheduled to enter full service in November 2012 with 864 cabinets. As of the November 2011 TOP500 list, it uses 68,544 2.0GHz 8-core SPARC64 VIIIfx processors packed in 672 cabinets, for a total of 548,352 cores, manufactured by Fujitsu with 45 nm CMOS process technology. Each cabinet contains 96 compute nodes in addition to 6 IO nodes. Each compute node contains a single processor and 16 GB of memory. Its water cooling system minimizes failure rate and power consumption.

The K uses a proprietary six-dimensional torus network interconnect called Tofu, and a Tofu-optimized Message Passing Interface based on the open-source Open MPI library. Users can create application programs adapted to either a one-, two-, or three-dimensional torus network.

http://en.wikipedia.org/wiki/K_computer

IBM has the Sequoia system coming on line in 2012 and it is also targeted at the 20 Petaflop range. It will be significantly more power efficient at 3000 Mflops/watt, three times lower then the K system

http://en.wikipedia.org/wiki/IBM_Sequoia