Supercomputers: Speed within reach at affordable prices

This has been a milestone year in supercomputing. These high-performance machines, engineered to cope with the most mathematically intensive tasks, have hit new heights of power and performance, while becoming more widely available to businesses.

Twice yearly, in June and November, the 500 most powerful supercomputers in the world are ranked by Top500.org, a group of supercomputing experts from the Universities of Mannheim and Tennessee and the Lawrence Berkeley National Laboratory.

In June, IBM’s Roadrunner supercomputer, based at the Los Alamos National
Laboratory, became the first machine to achieve the long-coveted goal of breaking the “petaflop” barrier, processing more than one million billion calculations per second.

In November, Cray’s Jaguar supercomputer, installed at the Oak Ridge National Laboratory, also passed that milestone. In the most recent rankings, Roadrunner and Jaguar notched up performances of 1.105 petaflops and 1.059 petaflops, respectively.

These are giant leaps. “When I started out in the supercomputer industry 25 years ago, we were all talking about the megaflop barrier – millions of operations per second,” says Margaret “Peg” Williams, senior vice-president of engineering at Cray, an established supplier in the field. “These days, that’s just dull and we’ve broken the gigaflop and teraflop barriers along the way to today’s petaflop performance.”
Today, even the cheapest PC runs at gigaflop rates yet it is only 11 years since the US government spent about $33m to build ASCI Red, one of the first supercomputers to achieve one teraflop.

The rate of progress equates to a 1,000-fold increase in processing power roughly every 10 years, says Dave Turek, vice president of IBM’s Deep Computing group, and that in turn has led to “enhanced fidelity”, or more refined analysis, in terms of output.

“We’re now getting to the point where a supercomputer isn’t just used to model the behaviour of a single component in a jet engine, for example, but whole collections of components, working together in an integrated system.

“Or we can model different processes on a much smaller or larger scale, from the effect of a new drug on a disease at a molecular level, to the impact of human behaviour on climate change over hundreds of years. And we can perform these tasks far faster, too.”

This ability, he says, is largely down to parallel processing – breaking a large computing task into logical components and processing them simultaneously on a vast pool of multiple processors.

The more chips added to a machine, the quicker they can process large tasks, as long as the application is engineered to run in parallel, rather than sequentially.
Since parallel processing became available to commercial organisations in the early 1990s, the line has blurred between traditional supercomputing, geared to solving complex scientific problems, and high performance computing (HPC), which aims to deliver bulk processing power.
Indeed IDC, the consultancy, has redefined its supercomputer category to include HPC systems costing more than $500,000.

Sales of such systems, it estimates, grew by 24 per cent in 2007, to reach $3.2bn.
But as the technology takes on an increasingly important role in industry and government, as well as academic research, it is increasingly the low end of the market, for HPC systems costing less than $250,000, that is the real engine of growth.

The road to such growth began in the 1990s, when high-end computing systems started using standard components in place of proprietary technology.

“Twenty years ago, a supercomputer was built from exotic components that made it blindingly fast and blindingly expensive as well,” says John Barr, an analyst with the 451 Group, an industry research company.

Today’s supercomputers, he says, use commodity chipsets from Intel and AMD and interconnect technologies based on the widely used Ethernet and Infiniband standards. Prices have dropped and smaller organisations are now able to take advantage of the processing power.

Arguably the most striking example of this came in September with the announcement of a partnership between Cray and Microsoft, which has made rapid advances in the sector with its HPC Server operating system.

The two companies launched the Cray CX1, touted as ”the most affordable supercomputer Cray has ever offered”, with prices starting at $25,000.

Vince Mendillo, director of HPC for Microsoft, says the CX1 is aimed at organisations in industries such as life sciences and financial services that are exploring supercomputing and do not have the in-house expertise or resources to run a fully fledged supercomputer.

The partners also hope to lure scientists, and researchers with discretionary IT budgets, to use the system to run workloads locally, creating ”supercomputing on the desktop”, rather than on huge, centralised machines.

A similar model is offered by chip maker Nvidia, which in November announced a “Personal Supercomputer” design, that will enable other manufacturers to build high-performance computers that look like traditional workstations.

Already, a number of PC vendors, including Dell and Lenovo, have signed up to offer workstations based on this design, at prices starting below $10,000, according to Andy Keane, general manager of GPU computing at Nvidia.

These machines, he says, will enable researchers to conduct at least some of their research at their desks. “Now, researchers tend to write the code on a standard notebook and then take it to the supercomputer in order to deploy it,” he says.

Carrying out research at a workstation will reduce the time it takes to work on problems and crunch data, he adds. The growing availability of HPC capabilities, however, raises questions about environmental impact: they consume enormous amounts of energy.

Mr Turek of IBM points out: “Energy efficiency has become as important as raw performance for the modern supercomputer. Now, it’s all about how much computing you can get out of each megawatt consumed.”