20 petaflops: New supercomputer for Oak Ridge facility to regain speed lead over the Chinese
(PhysOrg.com) -- The Oak Ridge National Laboratory (ORNL) campus in Oak Ridge Tennessee will soon play host once again to the fastest computer in the world (barring any new sudden announcements by the Chinese). The computer, dubbed "Titan" has been commissioned by the U.S. Department of Energy, and is expected to achieve 20,000 trillion calculations (20 petaflops) per second.
It was only last October that Chinas National University of Defense team unveiled the Tianhe-1A, a machine capable of computing at 2.5 petaflops.
The Titan, built by Cray Computer, will become part of a collection of some of the fastest computers in the world at the ORNL facility, joining NOAAs Gaea, the NSFs Kraken and the DOEs current workhorse, the Jaguar, though new space will have to be found, as the current structure has no room. Plans are in the works for an entirely new facility to be built over the next year, which should fit in well with the delivery date for the first stage of the Titan expected to be by the end of this year, with the second stage slated for sometime next year.
The Titan architecture will rely on use of XT3, 4 and 5 processor boxes, but will use a "Gemini" XE interconnect, and it will be configured in a 3D torus topology, rather than as an array.
Supercomputers achieve their ability to process enormous amounts of data in very short amounts of time by in essence, hooking together a lot of processing boxes and then using a device to connect them all together, this is why the Gemini XE interconnect is so important; its actually one of only two new pieces of proprietary hardware that will be added to create the new machine; the other is the Graphics Display Unit (GDU) co-processor (likely provided by Nvidia) that will help to perform calculations more quickly. This is also why computer scientists are so easily able to choose ahead of time just how fast a new computer will be; the more processor boxes you add, the faster the end result, so long as you have an interconnect that can handle them. The Titan will also use what is being described as globally addressable memory, which means data wont have to slow down as it passes through I/O channels.
The Titan is expected to be used by the DOE to calculate complex energy systems and will cost the government somewhere in the neighborhood of $100 million dollars.
More information: computing.ornl.gov/SC10/docume … Booth_Talk_Bland.pdf
© 2010 PhysOrg.com