NVIDIA's CUDA turns GPUs into high-powered CPUs

NVIDIA's been dancing around the general-purpose processor market for a while now -- we've heard reports that the company is developing an x86 chip, and it bought PortalPlayer last year for $357 million. Well, at this year's Microprocessor Forum the company took another small step by announcing that the final release of CUDA, its framework for utilizing high-end NVIDIA GPUs as CPUs, which will be available to developers in the second half of the year. While the idea of using a GPU as a secondary high-performance processor isn't a new one -- Folding@Home already runs on NVIDIA and ATI chips, and the Peakstream system already leverages GPUs -- CUDA should make it easier for developers to tap into high-performance graphics devices whenever they're available, without having to specifically tailor their apps to do so. CUDA, which stands for "compute unifed device architecture," currently only supports the GeForce 8800 and 8600 and Quadro FX 4600 and 5600, so it's of limited appeal right now, but here's hoping the next gen of NVIDIA chips supports CUDA from the get-go -- the Engadget Folding@Home team is looking for a few new recruits.

[Pete @ May 25th 2007 8:46AM]

Last I checked there wasn't a current Folding @ Home client that worked on NVIDIA GPUs. They used to have one, but now it seems that they are only supporting ATI.

[RAdZer0 @ May 25th 2007 10:30AM]

Nope, same thing here i cant find a nvidia client.

[Richard Lai @ May 25th 2007 11:28AM]

A friend of mine who works for QinetiQ had already been writting codes for using GPUs as CPUs. Don't exactly know which cards they used though.

[psyc0de @ May 25th 2007 10:15PM]

I've been using CUDA on my 8800 GTS 320mb to do a computer architecture research project. My unoptimised code could do fast fourier transforms up to 16x faster than an Intel E6600 for large data sets.

NVIDIA have said that CUDA will be supported on all future GPUs (however coderss will have to maintain backwards compatibility with the G80, as it is not capable of handling double precision floating point numbers).

[PS3-FOLDING@HOME @ Jun 18th 2007 12:50AM]

@psyc0de Yes I understand CUDA does a really good job of turning a GPU into a CPU. Well sort of, but it's still limited as to what tasks it can perform. But the scuttlebutt is that Nvidia is going to join the convergence zone. Where GPU's are becoming more CPU like and CPU are becoming more GPU like. More like the Cell BE!

"In fact, according to Folding@home's Vijay Pande, the project's GPU code was originally developed for NVIDIA GPUs" He also says they are working on a program right now to bring them online again.

What if Nvidia was to pair up with Intel? A sort of AMD/ATI deal. Make a multicore chip with both GPU and CPU cores on it! Anybody think that's feasible?