datasheets.com EBN.com EDN.com EETimes.com Embedded.com PlanetAnalog.com TechOnline.com  
Events
UBM Tech
UBM Tech

News & Analysis

Comment


KB3001

9/12/2010 4:51 PM EDT

High speed and low latency interconnect, low power processors and memory, and ...

More...



rick.merritt

9/9/2010 12:43 PM EDT

What are you doing to pave a road to exascale systems?

More...

SGI ASIC is first step to exascale system

Rick Merritt

9/9/2010 7:44 AM EDT

SAN JOSE, Calif. – SGI plans to deliver by the end of the year a system upgraded with a new ASIC that marks its first step on the road to an exascale-class supercomputer. The move is just one of many for a computer industry beginning to turn its attention toward cracking the exascale barrier by the end of the decade.

At least three U.S. government programs are now pursuing exascale systems, some aimed to consume as little as 20 megawatts. The initiatives could spawn new systems architectures at a time when a handful of systems have cracked the petaflops barrier.

SGI announced in June its Altix UV system that supports up to 256 Intel Nehalem-EX processors and up to 16 terabytes of main memory all housed in four cabinets. It uses 32 four-port node controller ASICs designed by SGI.

By December, SGI will ship versions of the system using a 16-port router ASIC to allow users to connect 128 of the four-cabinet nodes into a loosely coupled system supporting eight petabytes of total aggregate memory. The design is the first implementation of what SGI calls a global memory architecture that could scale to support an exascale-class system by 2018.

SGI's chief technology officer Eng Lim Goh described global memory as a middle ground between today's shared memory systems and clusters. The approach is the starting point of a partnership SGI announced in mid-August with Intel and others in an exascale program sponsored by the Defense Advanced Research Projects Agency (DARPA).

The partnership under the so-called Ubiquitous High Performance Computing program is believed to include an unnamed memory vendor. IBM and Cray are now in a late stage of a separate DARPA competition to create systems that can sustain petaflops-class performance but are significantly easier to program than today's massively parallel systems.

IBM described its approach to the DARPA High Productivity Computing System project in a recent paper at the Hot Chips conference.

Both IBM and Cray are developing new parallel programming tools for the HPCS project that could supplant today's Message Passing Interface software widely used in today's supercomputers based on clusters. SGI's Goh downplayed the move to any new programming environments and said SGI will deliver MPI libraries for its global memory systems.

"It took a long time for people to switch from Fortran to C, and we still have people doing high-performance computing in Fortran," said Goh. "I expect many people will still be using MPI at the end of this decade--even if something beyond MPI emerges," he said.

SGI is one of the early pioneers of high-end computing but fell on hard times in the last decade. It changed architectures and filed for bankruptcy twice and was acquired a year ago by Rackable Systems, a growing provider of x86-based servers.

In its latest financial results, the combined company, operating under the SGI name, has grown revenues but has yet to have or forecast profits.





rick.merritt

9/9/2010 12:43 PM EDT

What are you doing to pave a road to exascale systems?

Sign in to Reply



KB3001

9/12/2010 4:51 PM EDT

High speed and low latency interconnect, low power processors and memory, and fault-tolerant software and hardware are all key ingredients in exascale systems. That said, we need to develop enough software applications to take advantage of exascale systems, which is far from being the case. I think the latter point is not being thought of seriously.

Sign in to Reply



Please sign in to post comment

Navigate to related information

Datasheets.com Parts Search

185 million searchable parts
(please enter a part number or hit search to begin)