SGI Prism XL Leverages CPU, GPU Computing

At the Supercomputing 2010 show, SGI unveils its Prism XL, which officials say uses accelerators to bring a petaflop of computing into a single cabinet.

SGI is unveiling a hybrid system aimed at the high-performance computing space that leverages the power of CPUs and accelerators such as graphics chips from the likes of Advanced Micro Devices and Nvidia.

Introduced at the Supercomputer 2010 show in New Orleans Nov. 15, the Prism XL takes advantage of SGI's new Stix architecture and the acelerators to enable the company to offer up to a petaflop (One quadrillion floating point operations per second) of computer power in a single cabinet, something company officials said would take 100 cabinets using traditional x86 processors.

Prism XL comes at a time when businesses and institutions looking for HPC capabilities are turning to GPUs to help them increase their compute power while reducing power and space costs. It also comes as SGI officials look to the technical computing space as a key growth driver for the company.

SGI officials first spoke of its "Project Mojo" at the Supercomputing show in June, promising the ability to create very dense, high-performing HPC systems. Prism XL will be available starting in December.

Bill Mannel, vice president of product marketing at SGI, ticked off a number of business sectors that he expects will embrace the high performance and extreme density of the Prism XL, including oil and gas, defense and intelligence, and bioinformatics.

"The majority of the industries ... are being challenged by the need to attain more compute capabilities without increasing-and in some cases decreasing-their power footprint," Mannell said in an interview with eWEEK. "And they do not want to outgrow their data center footprint."

Those demands make accelerators attractive, particularly to those businesses that run heavy loads of parallel computing jobs.

The Stix architecture uses what SGI calls "sticks" as the key part of the Prism XL. Each stick comprises two "slices" that include full PCI Express Gen 2 capabilities and a single socket motherboard powered by an AMD Opteron 4100. Each slice offers up to two 2.5-inch SATA drives and two 1.8-inch SSDs (solid-state disks), giving 4 terabytes of storage on each stick. The sticks also offer built-in fans and auto-sensing power supplies.

The sticks, each of which measures 5.75-by-3.34-by-37 inches, also offer the latest accelerator technology from AMD and Nvidia, via its Tesla technology, and Tilera, a company that offers 36- and 64-core chips, and is promising 100- and 200-core processors in the future. The long, rectangular sticks are situated on their side, and include up to four DIMM slots.

Prism XL will support CenOS 5.5. and Red Hat Enterprise Linux 5.5, while SGI's Management Center will be the management tool. Management Center also can be used to manage SGI's other offerings, including Altix UV, Altix ICE, and Rackable systems.

"They're all managed with the same GUI, the same software," Mannell said.

Prism XL will be a key ingrediant as SGI looks to gain greater traction in the technical computing space.

"This fits squarley in that vision," Mannell said. "This is a system that today is primiarly focused on technical computing."

In a recent interview with eWEEK, SGI CEO Mark Barranechea said that the first 12 months after Rackable Systems bought Silicon Valley icon SGI and assumed the name were strong, with the company reporting $525 million in revenues-more than the $500 million officials had initially anticipated. The financial results represented a solid return on the $42.5 million Rackable paid for the bankrupt SGI.

Now officials are looking for revenues to grow to between $550 million and $575 million in the second 12 months, and see the $9 billion technical computing space a as an important factor in that expected growth. Larger systems makers like IBM, Hewlett-Packard and Dell are optimizing their systems to run high-end business applications from the likes of SAP, Microsoft and Oracle, resulting in a technical computing space that Barranechea said is underserved.

"This is an explosive opportunity for us," he said.

The technical computing space is demanding not only systems with enough compute power and interconnect capabilities to run their programs, but also the ability to keep power and space costs down. The use of accelerators like GPUs is a key part of that, and both Barrenechea and Mannel said they expect the drive for more GPU-based computing will grow. Barrenechea said the use of GPUs in general-purpose computing is "in the second-inning of a nine-inning game," but that he is seeing enough demand from SGI customers to know that it won't slow down any time soon.

Mannel agreed.

"We do see accelerators as becoming a more and more significant part of the technical computing space," he said.

Both SGI officials also said Prism XL, combined with the company's Altix UV system, helps keep SGI moving down the path to exascale computing, where peformance will reach the exaflop scale, or one quintillion floating point operations per second.

"We think that over the next seven to 10 years, we'll go from petaflop computing to exafop computing," Barrenechea said.