IBM speeds past HP to supercomputing supremacy

IBM has regained dominance on a list of the 500 fastest supercomputers and has also landed two unusual prototypes in the top 10.

Of the systems on the latest Top500 list, Big Blue built 224 and HP built 140, giving IBM back the lead it lost in 2001. Two new systems, ranked number four and number eight, are prototypes of Blue Gene/L, a system that uses vastly less space and power than its competitors.

IBM calls the prototype systems evidence that its years-long research effort is bearing fruit. But HP, which in 2003 had the most sales in the broader high-performance technical computing market, counters that it had the same number of new systems on the list - 108 - as IBM.

The list, released twice a year at supercomputing conferences, is based on a mathematical speed test called Linpack. The top system - NEC's long-dominant, 5,120-processor Earth Simulator - can perform 35.8 trillions of calculations per second, or 35.8 teraflops.

Linpack measures one aspect of supercomputer performance, but list organiser Jack Dongarra of the University of Tennessee is working on a broader suite of tests so a government supercomputing project can gain a better-rounded view. He expects the Top500 to stick with the Linpack yardstick, though.

As with November's Top500 list, there are plenty of new entrants.

The 1-teraflop threshold, crossed by 12 systems three years ago, now has been attained by 242 systems.

And the bottom-ranked system on this list was the 242nd six months ago - record turnover, according to the list organizers at the University of Mannheim, the University of Tennessee and Lawrence Berkeley National Laboratory.

As expected, the number two position is held by California Digital's Thunder, a system at Lawrence Livermore National Laboratory that boasts 4,096 of Intel's Itanium 2 processors and a speed of 19.9 teraflops.

Also new in the top 10 is an IBM system at the European Centre for Medium-Range Weather Forecasts that features 2,112 Power4 + processors. It ranks sixth with a speed of 9 teraflops. Fujitsu's Super Combined Cluster at the Institute of Physical and Chemical Research reached 8.7 teraflops. There's also a Chinese system, the Dawning 4000A, which sports 2,560 Advanced Micro Devices Opteron processors and can manage a speed of 8 teraflops.

Missing during this go-around is Apple Computer's System X at the Virginia Polytechnic Institute and State University, which soared to the number three position in November then vanished for an upgrade to newer hardware.

AMD's Opteron, which first arrived on the list six months ago, has made substantial inroads with 30 systems on the list.

There also were 30 of IBM's BladeCenters - chassis that holds many slim "blade" servers. One of them - a 168-processor system at the Joint Supercomputer Center in Moscow - uses IBM's JS20 blades. The JS20 uses the same PowerPC 970 processors that were in the System X.

The Blue Gene/L machines - three of which have been sold so far - are an unusual design based on a variation of IBM's Power family of processors. Each chip has dual processors, and each processor has built-in circuitry to communicate with five separate networks. That means the systems don't need a separate, expensive switch to link all the processors, and that doubling the number of processors comes close to doubling the overall performance, said Bill Pulleyblank, a director of IBM's Deep Computing group.

IBM built two prototypes, the number four ranked system, with 8,192 500MHz processors and a speed of 13.8 teraflops, and the number eight ranked machine, with 4,096 700MHz processors and a speed of 8.7 teraflops. The 700MHz chips will be used in a Blue Gene/L, the construction of which will start this fall at Lawrence Livermore National Laboratory.

There's a "good chance" the Livermore system will top the Top500 this November, said Dave Turek, leader of the Deep Computing team. He scoffs at governmental fretting that the US is losing dominance to the Japanese because of Earth Simulator.

"The trajectory of it is already clear. Those parties who might be wringing their hands about the supremacy of the Japanese ought to be wringing their hands about something else," Turek said.

One of the reasons is compactness. The top-ranked Blue Gene/L is about the size of eight refrigerators. The Earth Simulator, though three times faster, occupies a room 71 yards long and 55 yards wide.

IBM eventually plans to make a product out of Blue Gene/L - but not yet. "When does Blue Gene make the transition from our characterisation as a research project to a more mainstream product? We're not ready to make that declaration yet," Turek said. "Pretty soon we'll be able to declare our intent here."

IBM also is working on another design, Blue Gene/C. Pulleyblank acknowledges that its chips have many more processors on each slice of silicon but otherwise remains mum.

Supercomputers are at the high end of the much larger high-performance technical computing market, which is growing faster than the overall economy.

Market researcher IDC expects the market to expand from $6.1bn in 2004 to $7.6bn in 2008. Unit shipments will grow even faster, from about 112,000 in 2004 to about 176,000 in 2008, according to the forecast.

It's common for falling computer prices to mean that shipment growth outpaces revenue growth. But in the high-performance technical computing market, the trend is reinforced by the idea of computing clusters - groups of low-end systems linked with high-speed networks.

Clusters, often using Linux running on computers with Intel or AMD processors, don't run some jobs as efficiently as more monolithic designs. But they work well enough at a low enough price that major customers are snapping them up. For example, of the seven Boeing supercomputers on this Top500, six are clusters using Xeon processors and one is a Cray X1.

And clusters let Dell, a company better known for its reliance on Intel and Microsoft research than its own engineering abilities, build the number five system called Tungsten at the National Center for Supercomputing Applications.

Clusters, which also can be assembled from higher-end parts, accounted for 281 of the 500 systems, for the first time crossing the halfway mark.

But some jobs aren't amenable to cluster approach - one being decryption of coded communications. And there still is a large market for modest machines, such as those with four or eight processors churning away on jobs such as structural and stress analysis, said Bruce Toal, HP's marketing director of high-performance technical computing.

Big Blue is getting bigger in the market, nearing HP's top rank. From 2002 to 2003, IBM's share of the market grew from 28.2 per cent to 30.2 percent with sales of $1.62bn. HP, with $1.79bn in revenue for 2003, shrank slightly from 33.6 per cent to 33.5 per cent.

Next in line in 2003 were Sun, with $981m in revenue, Dell with $387m, Silicon Graphics with $224m, Cray with $157m and NEC with $67.5m, according to IDC.

HP likes to direct attention to the overall market share statistics. But it's clear the company still has ambitions for the Top500.

"It is a two-horse race," Toal said, conceding the loss to IBM on the current list, "but stay tuned for next time."

Stephen Shankland writes for CNET News.com