Nvidia talks up GPU 'supercomputers'
Date : 02 24 2008 Category : TechnologyShaun Nichols in California, vnunet.com, Monday 25 February 2008 at 00:00:00
Company touts computing power of graphics chips
Nvidia is claiming that its graphics chips are not just for gaming and watching movies. At a press event in San Francisco, the company highlighted other ways in which users were harnessing the power of its graphics processing unit (GPU) chips. Chief scientist David Kirk explained how scientists have been taking up a practice called general purpose computing on graphical processing units (GPGPU) to perform huge computing tasks on GPU chips. The advantage of the GPU, explained Kirk is its ability to perform multiple calculations. Where a quad-core CPU is only able to handle up to eight processing threads at once, a GPU can handle as many 12,000 threads. That ability to multi-thread makes the GPU extremely well suited for computation-intensive tasks, such as model,ing physics or crunching large amounts of statistical data for models and simulations, Kirk said. Initially, however, users were forced to perform GPGPU operations on top of graphics API's in order to tap into the chips, a process Kirk likened to "using a screwdriver to open a can". To allow scientists to tap into the GPU's power, the company introduced the CUDA development tool in June of 2007. Since then, more than 50,000 users have downloaded the software. "GPUs are extremely powerful processors, they have hundreds of times more processing power than CPU's do," said Kirk. "A lot of real world scientific and consumer applications get hundreds of times better performance." Do not, he cautioned, expect the GPU to replace the CPU in your computer any time soon, though. "It doesn't replace the CPU. It doesn't do a lot of things the CPU does," said Kirk. "GPUs have a very hard time dealing with irregular data where there is a lot of decision making and not a lot of arithmetic." Instead, Kirk sees that the GPU chip taking over the large parallel computing tasks traditionally relegated to large supercomputer systems. Nothing the growing popularity of cluster systems, Kirk predicted that by 2010 three of the five fastest supercomputers will use GPU chips. "If you think about it, this is a massively parallel supercomputer on your desktop," he said. "It's truly the democratisation of supercomputing."

