GCN Insider | Trends & technologies that affect the way government does IT
LOOKS ARE DECEIVING: AMD's Stream Processor is a PCI Express-based add-in board that looks like a graphics card, but comes with higher memory densities than any consumer graphics card.
Can you use your computer's graphics card to speed applications? It might seem a far-fetched idea, but that was the buzz at the SC06
supercomputing show last month in Tampa, Fla.
At the show, Nvidia Corp.
of Santa Clara, Calif., introduced a new chip architecture, called Cuda
, that lets the company's graphics processing units take on additional computing tasks.
Not to be outdone, Advanced Micro Devices Inc.
of Sunnyvale, Calif., announced it would deliver what it called a processor based on a GPU that would be used solely for parallel computations. The AMD Stream Processor
is a PCI Express add-in board based on work from longtime Nvidia rival ATI Technologies
, which AMD acquired earlier this year.
On the software side, PeakStream Inc.
of Redwood City, Calif., demonstrated its PeakStream Platform
, a run-time and development environment that lets developers write programs to utilize GPUs.
'Graphics cards are very powerful floating-point engines, but until now, they've only been used for graphics,' said Michael Mullany, a vice president of marketing for PeakStream. Nvidia general manager Andy Keane claimed certain computations, such as medical imaging and signal processing, can run 100 times faster on a GPU. While a regular CPU can execute only one thread at a time, a GPU can have 128 or more execution units. So, simply assign the GPU the fine-grained, data-intensive parallel processing, and leave control and data management to the CPU.
Not everyone is sold on the idea, though. 'There are some issues to be worked out,' said Donald Becker, chief scientist for cluster software provider Scyld Software
, a subsidiary of Penguin Computing Inc.
of San Francisco. Becker noted GPUs can't run programs in protected mode, so running more than one GPU program on a computer can lead to chaos.
Others noted GPUs don't have double-precision floating-point accuracy, a shortcoming that will inaccurately truncate numbers longer than 32 bits. Louisiana State University computer scientist Thomas Sterling also was critical of reusing GPUs.
During his talk at SC06, he said he considered GPUs to be special-purpose accelerators and mocked the idea of 'general-purpose special-purpose units.