GPU projects aren't problem free. You can have a four year old CPU and gain fairly good amount of work done. But a three year old GPU can be too old. So don't look for the future projects but those that are currently available. Then browse their web forums and read what kind of gear people are using and will you be able to accept those runtimes... Here's an example what I mean: to my 8800gts one small gpugrid wu took 11 days to finish. I just find it uninteresting to update my GPU every one or two years just to be able to finish wus in <6 hours.

Yes, a GPU can be considerably faster and 250 times is not unheard of. Now for a reality check.

This cannot be done in the real world. It can be accomplished under ideal situations with a single or a few models of GPU.

Distributed computing is designed to take advantage of the widest possible range of hardware. As has been previously mentioned, CPUs are compatible for many years. GPUs require highly modified programs and few are compatible.

I'm not going to make up numbers about the potential, but I can assure you there is a problem of diminishing returns when attempting to write an application for GPU crunching. The more cards you attempt to include the more difficult the task becomes.

Is it worthwhile to spend considerable resources for a handful of crunchers when there is already a backlog of resource demands?

That is not correct, it HAS been done, read it agin [url]http://www.biomedcentral.com/1756-0500/4/97[/url]

Or is it a false statement?

Hi Simplex0,
We are both right. The hardware / software barriers to GPU computing will have been surmounted when corporations start specifying GPUs for their computers. After all, they pay their professional programmers a great deal. Both AMD and Intel are aware of this and have development programs for GPU programming.

So there is great promise, but there are still problems. Almost by definition, devices that are sold mainly to enthusiasts have either limited utility or serious problems.

Lawrence

I'm not trying to win arguments here, but to quote the article:

"The test system that we used is a modest one and is exceeded by many home computers used for gaming or for high definition video."

I don't know why, but I have the feeling that this would be the next GPU based WCG project. You know what, I'll just post the whole paragraph.

"GPUs with the power of small supercomputers are becoming ubiquitous in consumer computing devices. The test system that we used is a modest one and is exceeded by many home computers used for gaming or for high definition video. These devices are being made accessible to scientific applications through community grids which link millions of volunteer nodes together. Although GPU-Q-J was developed for clustering large ensemble sets on our local servers, the method will form part of a new GPU-aware protein folding client that is in development. Such GPU-aware clients have already made an impact in projects such as Folding@home [20]. Effective GPU adaptations of routines for commonly used calculations such as the optimal superposition/RMSD are important if we are to fully utilize the enormous power being made available through the generosity of participants in projects such as NRW."