Graphics on speed

A specially adapted computer graphics card claims to offer ‘desktop supercomputing’ capabilities to scientists and engineers. It comes after several years of determined researchers cannibalising the power of graphics chips for their own purposes. The Tesla chip, from US graphics hardware specialist Nvidia, will have a processing power of around a teraflop – which the company says is 250 times faster than a conventional desktop computer. “You will see a big effect, by putting teraflops of computing power into the hands of scientists,” says Nvidia’s chief scientist David Kirk. The chip is enabled with Nvidia’s Cuda parallel computing architecture – and can be programmed in a number of languages. It is effectively a graphics chip adapted for non-graphics applications. “We realised that the massively parallel structure of a graphics card could be used for other things – so we developed Cuda to write general purpose applications,” says Kirk. The company says this will “democratise” supercomputing, making it accessible to more people – and not controlled by the need to “buy supercomputer time”. This official unveiling of the chip will now allow business and academic IT departments – and commercial PC assemblers – to build their own ‘Tesla supercomputers’. But many researchers, frustrated at the lack of computing power for their own applications, have been building their own computers for years using these kinds of graphics chips. Joost Batenburg, of the University of Antwerp in Belgium, leads the Astra research group – which is developing new algoriths for a 3D imaging process called tomography. The technique is commonly used in medical imaging – and is “computationally intensive”. He and his team created a PC based on eight Nvidia graphics processing units (GPUs) – at a cost of around $6000 (£4000). This ‘unofficial’ PC was actually faster than the university’s own cluster-based supercomputer. “It was 350 times faster than a single Intel CPU core,” he says. While his area of research, tomography, is aimed mainly at medicine, he says it could also have use in reverse engineering and related areas. “There is a diamond company using this to scan diamonds and work out how best to cut them,” he says. “This could increase their profits by 3%.” German university spin-out FluiDyna is also ready to take advantage of cheaper supercomputing. It offers CFD services to a variety of companies – using open source CFD codes and its own proprietary CFD software. The company has been using GPUs – rather than standard CPUs – for just a few weeks, and has noticed the difference. “We ran 15 times faster than before,” says Eugen Riegel, a PhD student on placement at the company. “ Nvidia’s other case studies show project accelerations of anything from 20 to 150 times – such as a cell phone RF simulation, which went from eight hours to 13 minutes. “It’s like entering a time machine and going 15 years into the future,” says Kirk. Pointers An adapted graphics card claims to offer ‘desktop supercomputing’ to scientists and engineers CFD will be a key engineering application, due to its need for high processing power A single chip has a processing power of around a teraflop