DCU applies ‘cluster power’ to chip processing

26 Feb 2004

Anyone who has visited Dublin City University (DCU) on the capital’s northside will be aware of the incredible expansion of the campus in recent years. One of the newest buildings is the €7m National Centre for Plasma Science and Technology, which is part of DCU’s Physical Sciences Department and funded under the Government’s Programme for Research in Third-Level Institutes scheme.

The centre’s focal activity is basic research into plasma, the luminous-ionised gas used in fluorescent lights and the latest must-have TVs. It is also a key ingredient used in the production of silicon microprocessors. One of the sponsors of DCU’s plasma research is Intel, which it is believed uses some kind of plasma processing device in about two thirds of its semiconductor process steps.

“The reason why there’s so much work going on in chip manufacturing is because the added value is so big,” explains Professor Miles Turner (pictured), director of the centre, which houses 25 plasma researchers. “You don’t have to tweak up the efficiency of Intel’s plant in Leixlip too much in order to justify your investment. They’re taking a silicon wafer worth about US$5 and when it gets to the end of the production line it’s got something like US$100,000 worth of chips on it,” he adds.

Turner’s plasma research team got a major boost last year when Science Foundation Ireland awarded it 5m to fund Plasma Formation Applications and Controls, a project being undertaken jointly with the Process Control Group at NUI Maynooth.

“Chip manufacturers are finding it harder and harder to get where they want to go using the existing technology,” notes Turner. “By bringing together what we know about plasma simulation with Maynooth’s expertise in process control, we’re hoping to add a lot of value. What we are really aiming for is more consistency in the [manufacturing] process.”

In simple terms, Turner hopes that the research will lead to semiconductor manufacturers being able to minimise wastage and the amount of test wafer production needed. This would in turn reduce production costs and help drive down processor prices.

To support plasma research and a range of other mathematical and scientific projects, DCU recently took delivery of a high-performance cluster from Fujitsu-Siemens. Costing approximately €200k the machine consists of 32 1.8GHz AMD processors running in parallel.

“High-performance computing allows you to link multiple processors/systems together and operate as one entity. It means academics can share processing power across campus to solve problems,” says James Finnerty, technical director of Fujitsu-Siemens Ireland, who added that the company had recently won a contract to deliver a cluster to a second unspecified Irish university.

High-performance or parallel computing is playing an increasingly important role in a range of scientific disciplines. Its massive computing power means that bigger and more complex calculations can be conducted much faster. Turner estimates that for some calculations the cluster will get results 100 times faster than if an individual large processor were used. He also notes that as processor technology starts to plateau the performance gap between specialised single processors and the ‘mass market’ Intel or AMD equivalent is narrowing all the time. This means that significant jumps in computing power can only be achieved by linking together or ‘clustering’ a group of processors.

On its own, however, a large computing cluster will not deliver the promised higher computing speeds. To harness its full power, there must also be a fast network in place that can instantly share data between different processors. At DCU the infrastructure consists of a 100MB network linking various departments and sites. Critically important too is ‘network latency’ – the time it takes to establish a network connection. To optimise the existing network, the Fujitsu-Siemens cluster uses Myrinet communications and switching technology to move data around a network as efficiently as possible.

“Usually with parallel computing the thing that kills you is the communication time between the different processors. The network speed is the limiting factor for a lot of computations,” remarks Turner.

No more than you need to see the power station that supplies electricity to your home, it is not necessary to be physically close to the cluster to benefit from it. In fact, Turner admits to never having actually seen the machine, which is located in the machine room of the neighbouring library building. For Turner, it is not necessary to see in order to believe; he is already a committed disciple of high-performance computing and believes it has a major contribution to make to plasma research.

By Brian Skelly