Standardize, Don’t Commoditize

In the past few years, the Supercomputing show has grown more noteworthy. The 2005 Supercomputing Conference (more commonly referred to as SC05) got under way in Seattle, Wash. Monday with all the usual players and some not so expected ones. The most apparent thing about the show this year is the breadth of announcements taking place both at the show and in its limelight as well as the players making them..

From the din of announcements coming out of this week’s Supercomputing Conference, the theme ‘standardize, but don’t commoditize’ seemed to echo from every corner.

As usual, the 26th edition of the Top500 list was released with much fanfare at the start of the show. The list, compiled by the Top500 group, a project started in 1993 to track trends in high-performance computing and released in June and November each year, saw much shuffling. IBM hit something of a trifecta with its machines taking the top-three spots on the list.

IBM’s BlueGene/L supercomputer retained its chokehold on the top spot, largely due to the record speed of 280.6 trillion calculations per second BlueGene/L, which was developed by IBM and the National Nuclear Security Administration (NNSA) for safe nuclear weapons testing, and has more than doubled its speed twice in the past year, up from 135.3 teraflops and 70.7 teraflops a year ago.

BlueGene/L will likely remain in the No. 1 spot for at least the next few lists because no other supercomputer has yet cracked the 100 teraflop barrier, Top500 researchers said in a statement.

IBM’s Watson Blue Gene system again held the No. 2 spot at 91.29 teraflops. Big Blue’s ASCI Purple system, based on IBM’s p575 server, reached 63.4 teraflops to take the third slot, bumping the 51.87 teraflop Columbia system built by SGI to No. 4. IBM machines inhabited roughly 44 percent of the Top500 list.

Subsequent slots on the Top 10 list changed much since the June list. Four of the top 10 systems from the June 2005 Top 500 list have been displaced by new systems, and the last 221 systems from June 2005 are now too small to be included, demonstrates just how fast high-performance computing innovation is ramping up.

Two new systems at DOE’s Sandia’s National Laboratories secured the No. 5 and 6 spots: The Dell PowerEdge 1850 Thunderbird system edged out Cray’s Red Storm system, 38.27 teraflops to 31.69 teraflops.

The NEC-built Earth Simulator, which at 35.86 teraflops had held the No. 1 position for five straight Top500 lists before being replaced by BlueGene/L last November, has slipped to No. 7.

Rounding out the Top 10 were IBM’s MareNostrum JS20 Cluster, Astron eServer Blue Gene Solution, and Cray’s XT3 system at DOE’s Oak Ridge National Laboratory.

Standardization, not commoditization seemed to be the rallying cry, both in the performance numbers and in the announcements.

For starters, x86 Intel chips dominated the landscape. Two-thirds of the systems on the Top 500 list systems using the No. 1 chipmaker’s silicon, including 81 with Intel’s new EM64T-based processors. There were four times as many Xeon deployments as its next closest competitor, Jason Waxman, a director in Intel’s Server Platform Group, told ServerWatch.

Itanium found a home in a number of the servers on the list, and is poised to penetrate further in 2006. Waxman said 9,000 Itanium processors have been ordered for supercomputing systems for next year.

Other processors that made the list include IBM’s Power processors inhabit 73 systems, but AMD’s Opteron chips are gaining ground, with 55 systems using them compared to only 25 systems six months ago.

Notably missing was Sun’s SPARC processor — a slippage that UltraSparc T1 with CoolThreads (formerly referred to as Niagara), which Sun revealed details of on Monday, will presumably remedy.

The new chips have the same instruction set as previous UltraSparc chips, but a different design. They are designed to address energy and spatial issues, which are now paramount to many enterprises. According to Director of Outbound Marketing Fadi Azhari, 80 percent of data centers are constrained by heat and power issues. By addressing this issues, Sun believes it will save customers millions of dollars.

The multithreaded chip contains eight cores, with four threads each, for a total of 32 threads, which is akin to 32 systems on a chip. The increased integration on the chip, Azhari said “removes the bottleneck between the data and the unit.”

The other notable aspect of UltraSPARC is its strong security. It has seven to 10 times the RSA encryption levels of chips on the market, and it processes “fully encrypted messages at the same rate as straight text,” Azhari said.

The chip is currently in volume ramp. Systems are expected to be announced by year end. Azhari describes the chip as being in a “cross-over category” whose sweet spot will be found in the Web tier space, where the chips will compete directly with Xeon.

Nowhere is the trend of standardization more evident than in those vendors once synonymous with supercomputing. List veterans SGI and Cray, both had announcements Monday. SGI unveiled a new Altix 4000 flagship system, which begins the transition of the vendor’s high-end line to a blade form factor.

The Altix 4700 can hold up to 512 processors and has up to 100 TB of memory address space. The Altix 4700 integrates SUSE and Red Hat with SGI’s FPGA-based SGI RASC technologies and is compatible at the socket level with Itanium 2 processors and multicore Itanium processors.

It is scheduled to begin shipping in latter part of 1Q06.

Cray’s announcement wasn’t nearly as dramatic. The vendor agreed to continue using Opteron processors for its microprocessor-based supercomputer products through the end of the decade. Cray and AMD also agreed to collaborate on Cray’s mid-2006 proposal for Phase 3 of the federal government’s DARPA HPCS program.

Cray currently sells two Opteron-based products: the XT3 and the XD1. Cray CTO Steve Scott told ServerWatch that 90 XD1 systems have been sold since it hit the market last year. Cray’s other system, the X1E supercomputer, is based on Cray custom-built vector processors. This latest collaboration will further embed AMD technology in Cray’s next-generation supercomputer products, which are designed to increase performance for HPC applications by drawing on both AMD Opteron and custom, specialized processors.

Standardization appears to pervade the list, and even the stalwarts seem to have caught on to this and are adjusting their road maps accordingly. With supercomputing power increasingly accessible to nearly all organizations of all sizes and types, it may be time to, in the words of Sun’s Jonathan Schwartz, start thinking of computing power as the commodity but not the computer.