News & Analysis
Comment
pmoyle
I think AMD would have been a better choice.
kinnar
Scientists and Engineers will like to have an overview of the architecture of ...
Intel wins sockets in 10PFLOPs super
Rick Merritt
9/22/2011 4:55 PM EDT
SAN JOSE, Calif. – Intel has landed the first major design win for its Many Integrated Core (MIC) processor in a supercomputer expected to hit 10 petaflops when completed in early 2013. The system is expected to surpass the world's current fastest computer that delivers eight PFLOPs using Fujitsu Sparc processors.
A supercomputing center at The University of Texas at Austin announced it will use a $27.5 grant under the Extreme Digital program of the National Science Foundation to help fund the effort. The so-called Stampede system ultimately will cost more than $50 million over four years.
The system will be based on thousands of dual-core Dell servers using the Sandy Bridge version of Intel's eight-core Xeon E5 processors. Stampede also will use thousands of the Intel MIC chips, known by the code name Knights Corner, which will provide 80 percent of the systems performance, Intel said.
The Texas center expects to upgrade Stampede with future generations of Intel MIC processors to drive peak performance to at least 15 petaflops.
Intel's MIC chips had their origin in an Intel effort to create the Larabbe graphics processors based on an array of x86 cores. Intel nixed its graphics plans but repurposed the project as a high-performance computing chip to rival Nvidia and AMD parts used in other supercomputers.
The Texas center said the use of the x86 instruction set both the Xeon and MIC chips will "simplify the task of porting and optimizing applications on Stampede."
The system is not built solely from Intel parts. Nvidia won 128 sockets in Stampede for a next-generation chip to handle remote visualization jobs.
The latest supercomputers are increasingly relying on a mix of many core graphics processors working in tandem with general purpose CPUs.
Altogether, Stampede will have 272 terabytes of total memory, and 14 petabytes of disk storage. It will be based on a InfiniBand FDR 56Gb/s network.
\The system will become part of a network of more than a dozen high-end computers created by the NSF program. "Many researchers will leverage Stampede not only for massive computational calculations, but for all of their scientific computing, including visualization, data analysis, and data-intensive computing," said Jay Boisseau, director of the supercomputer center, speaking in a press release.
\"My group is excited about the opportunities Stampede offers to greatly accelerate our work in quantifying uncertainties in computer models of dynamics of polar ice sheets, global seismic wave propagation, and whole-earth plate tectonics," said Omar Ghattas, a professor of geological sciences and mechanical engineering at UT Austin.


sharps_eng
9/22/2011 6:41 PM EDT
These huge computing systems obviously cost a lot, but do they really make money? Seems the number of sales will never pay back the R&D. I'd be interested if they are a kind of live research lab for technology that will trickle down into the marketplace where the real profits lie?
Sign in to Reply
AlunWang
9/23/2011 1:02 AM EDT
Right, I have no idea what is its usage?
Sign in to Reply
markhahn
9/23/2011 2:05 PM EDT
I don't think MIC is costing Intel much. yes, they have research groups that occasionally spin custom chips, but that by itself isn't a big expenditure. it's a little hard to tell, but Intel seems to have been pushing the same current rev of MIC for several years now, and it wasn't anything extreme when originally released.
most of the action will be software anyway.
Sign in to Reply
prabhakar_deosthali
9/24/2011 6:16 AM EDT
To my knowledge such huge high performing machines make their ways in the research laboratories where their main usage in simulating the atmospheric systems such as the world weather studies or modelling the phenomenon about the stars being formed and so on. For such simulations and modelling the more computing power the better. The benefits of such simulation cannot be really measured in the commercial terms but I am sure the govt funded R & D institutions are ready to pay for such supercomputers.
Sign in to Reply
KB3001
9/24/2011 12:42 PM EDT
Today's supercomputer is tomorrow's handheld device. Pushing the limit in supercomputing makes perfect sense in the long term...
Sign in to Reply
goafrit
9/24/2011 8:49 AM EDT
Not sure where this goes... At the end, price will decide. But Intel can make it work out.
Sign in to Reply
KB3001
9/24/2011 1:15 PM EDT
There is a race towards Exascale computing by ~2018. http://www.nccs.gov/wp-content/media/nccs_reports/Science%20Case%20_012808%20v3__final.pdf
Sign in to Reply
kinnar
9/25/2011 3:24 AM EDT
Scientists and Engineers will like to have an overview of the architecture of the supercomputer that uses thousands of dual core servers from dell and thousands of Xeon Processors from Intel.
Sign in to Reply
pmoyle
10/18/2011 1:30 PM EDT
I think AMD would have been a better choice.
Sign in to Reply