Cray's New Supercomputer XC30 Delivers 66 TFlops/Cabinet
Cray unveiled its Cascade supercomputer architecture, one of the first to support Intel's upcoming Xeon Phi accelerators.
While the system has been officially launched only now, Cray has been selling these systems as early as 2010 and claims to have to have six Cascade supercomputers under contract valued at more than $100 million already.
Next to IBM's BlueGene/Q design, Cascade is the second major architecture that is claimed to scale to about 100 petaflops. The system, which carries the commercial name XC30, will run Intel's Xeon E5-2600 series processors and can, in theory, support more than 1 million cores, Cray said. Future versions of the XC architecture will also be available with Intel Xeon Phi coprocessors and Nvidia Tesla GPUs to enhance the system's floating point horsepower.
According to Cray, the XC30 can ship with up to 384 Xeon E5-2600 processors per cabinet, representing up to 3,072 cores and an initial peak performance of 66 TFlops per cabinet. There can be 32 to 128 GB of memory per node, and a memory bandwidth of up to 117 GB/s. A single fully equipped (liquid cooled) cabinet will weigh 3,450 lbs and require power supply of about 88 KW, Cray said.
Butttt... It's over 9000 X my budget.
The Titan supercomputer (a Cray XK7 system) at the Oak Ridge National Laboratory is currently the world's fastest supercomputer at 17.59 PFlops, consuming roughly 9 MW.
The IBM Sequoia (a Blue Gene/Q system) is the second fastest supercomputer at 16.32 PFlops, consuming 7.9 MW.
The XC30 delivers 66 TFlops at 88 KW per cabinet. To beat the current best, it would take 273 cabinets to push a little over 18 PFlops using 24 MW. Compared to Sequoia, you only get approximately 1.7 more PFlops while consuming an extra 16 MW!
For perspective, 16 MW is enough to power TWO more IBM Sequoias!
EDIT: I did misread the Titan power draw figure (it's 9 MW, not 20 MW), but I believe the math for the XC30 power draw scaled to 18 PFlops is correct, therefore not very energy efficient compared to alternatives like Blue Gene/Q.
Forget Crysis! Can it play Geometry Wars on 4K res?
Both Titan and Sequoia use GPU's to reach their numbers... if XC30 will get the Xeon Phi upgrade it's speed per cabinet will increase tremendously and beat the others easily.
GPUs however while having enormous floating point power, they can only process some kind of jobs, so not all supercomputer tasks can be run on them.
With 299,000 AMD x86 Cores, it's got 17.59 PFlops of processing power.
The new champion Titan contains almost 19,000 processing nodes and 710 terabytes of memory.
In addition to a 16-core AMD CPU, each node contains a GPU accelerator.
The Cray doesn't play games, and instead is for uses like tracking satellite and 'space junk' trajectories -- all of them, and an assortment of scientific number crunching and simulations.
"The first six XC30 machines will go to the Swiss National Supercomputing Centre (CSCS) in Lugano, the Pawsey Centre in Perth, Australia, the Finnish IT Center for Science (CSC), the US Department of Energy’s National Energy Research Scientific Computing Center (NERSC) in Berkeley, Japan’s Academic Center for Computing and Media Studies (ACCMS) at Kyoto Universit, and the University of Stuttgart’s High Performance Computing Center Stuttgart (HLRS) in Germany."
Yes... Titan does use Opteron processors, but that's not the primary source of its computing power. Roughly 90% of its performance, as measured by LINPACK, comes from its Tesla K20X coprocessors.
And I'm sure XC30 could be configured to outperform Titan, it's powered by inherently higher performing processors.
I'm not sure where you got the 20MW figure from, but the last time I checked Titan drew 8.21MW at its 17.59 PFLOPS peak.