NASA, SGI Plan Petaflop Computer Breakthrough

The race to petaflop supercomputers is heating up, with the latest entry coming from NASA, Intel and SGI. The trio announced plans to build what will be a petaflop-capable supercomputer by next year, and up to 10 petaflops by 2012.

Supercomputers have been in a constant game of oneupsmanship and bragging rights. The definitive list of the fastest supercomputers, called Top500, is released twice a year, and for the last few years, IBM (NYSE: IBM) has dominated with its Blue Gene/L supercomputer at the Lawrence Livermore Laboratories.

With each update to the list, the number of teraflops goes up because no matter how much processing power you give these machines, there’s still more demand for them. These massive systems aren’t just doing one job at a time. They are rented out to other agencies or researchers who have a massive computing task they need done.

At any given time, a supercomputer likely has hundreds of computational tasks running on it. So there are never enough teraflops to go around. A teraflop is a trillion operations per second. As of last November, Blue Gene/L topped out at 600 teraflops and ran at a sustained rate of 478 teraflops. By contrast, a Core 2 Duo E6700 processor performs around 12-13 gigaflops, or billions of operations per second.

NASA’s Advanced Supercomputing (NAS) facility at NASA’s Ames Research Center, in Moffett Field, California, is a prime candidate for more computing resources. Ames had the 90 teraflop supercomputer dubbed Columbia, which once topped the Top 500 list. These days Columbia has a backlog to match rush hour traffic on the 101 freeway running just outside Moffett Field in Mountain View.

“The main thing that we’re really looking for is the ability to do a lot more of the work we’re doing. Columbia is absolutely packed, we have people waiting days to run a job,” Bill Thigpen, engineering branch chief at the NASA advanced supercomputing division told InternetNews.com.

“When we do that, that puts the work we’re doing at risk. Our scientists’ and engineers’ time is very valuable. So the ability to do significantly more work is valuable to us,” he said. He added that Ames is not getting rid of Columbia, it will continue to lend its CPU cycles to the mix.

SGI wins the bid

For the past year, Ames has been looking to upgrade its systems and went through a variety of bidders. It had previously purchased systems from SGI before but still went through the bidding process. In the end, SGI (NYSE: SGI) won out again, and will be the supplier for a new supercomputer named Pleiades.

“We have found them extremely supportive of our efforts and a very good vendor over the years,” said Thigpen. “It has worked out that they have come out on top in many situations.”

Thigpen said several elements played into SGI’s win, including the Integrated Computing Environment (ICE) built into the back plane of the computer, eliminating the need for lots of cabling, and the water cooling built into the cases. He also praised SGI systems for being easy to get up and running quickly, something the company prides itself on.

Intel (NASDAQ: INTC), which was involved in the bid, also had kind words about the ICE interconnects, which allow CPUs to communicate directly, rather than over high-speed Infiniband cables.

Page 2 of 2

“What SGI have done with ICE is bring the interconnect and switching into the system, which allows them to scale much more effectively than other vendors out there today,” said Richard Dracott, general manager for the high performance computing division at Intel.

“Certainly the ICE system’s architecture that they have developed has the ability to scale much more cost effectively and higher up than a traditional cluster based on Infiniband,” Dracott added.

The initial system will consist of 40 racks, each with 128 Xeon 5472 processors running at 3.0 GHz, plus seven support racks. The system will run SGI’s Linux implementation, which carries a lot of the enterprise technology from the old IRIX days.

The initial deployment will be a 245 teraflop system and gradually expanded to 1 petaflop by next year. Dracott said that the first systems will be the Xeons but there is the possibility of a change to an other Intel platform. “The next round of deployments could use Nehalem, but NASA hasn’t said what they are using,” he said.

Pleiades will be used for a variety of scientific calculations, ranging from climate modeling to shuttle simulations to nanotube simulations. “These machines are going to be able to take that kind of science and research out to a couple generations of improvement,” said Thigpen. He said the Pleiades system will allow for at least a 2.5-fold increase in work capacity, and perhaps more in some jobs.

NASA is upgrading its facility to handle Pleiades greater power requirements, but the first pieces will arrive this month. Thigpen expects to start putting it to work in June and be in full production by August.

News Around the Web