Question on "The GPU Advances: ATI's Stream Processing & Folding@Home"

[imported_SoF]

When the GPU beta client is released, how soon will it be available for SETI@home?

 

[BlackMountainCow]

I doubt that it'll ever be released for SETI, as the the algorithms used to calculate proteins (massive floating point ops, just like a GPU) and searching SETI WUs are to my knowledge way different. That is not to say that SETI will never use GPUs. But it took F@H over three years to develop that GPU client and I doubt that they'll give away that advantage too easily. I bet they want people to give their CPUs to whatever DC project they want, but "all you GPUs are belon to us"!

F@H doesn't use BOINC yet either, so that would be another hurdle.

 

[imported_SoF]

Thanks for the fast reply.

My hopes were a bit high because many articles here had been discussing about the untapped power of GPUs. I only came across the Sept 30th article today after a long absence from reading the latest from Anandtech. So on seeing this naturally my hopes were raised.

At least the first step has been taken. Hopefully they will also overcome the Nvidia GeForce hurdle soon.

 

[BobDaMenkey]

Originally posted by: SoF
Thanks for the fast reply.

My hopes were a bit high because many articles here had been discussing about the untapped power of GPUs. I only came across the Sept 30th article today after a long absence from reading the latest from Anandtech. So on seeing this naturally my hopes were raised.

At least the first step has been taken. Hopefully they will also overcome the Nvidia GeForce hurdle soon.

It's not so much of a hurdle, as it is an efficient useage of time. Right now the ATi chips are a good 4 - 8x faster than equivelent nVidia GPU's for the task at hand. They might eventually release an nVidia client, but for the researchers it's a much more practical solution to work only on the ATi client right now.

Yes it's disapointing to hear that, but hopefully when they get the ATi client sorted out fully, they'll start working on both the lower end ATi cards and nVidia GPUs as well.

 

[Ken g6]

Well, on one hand, the algorithms for SETI could actually be very good for a GPU. The main algorithm is the FFT, the same algorithm used by various large-prime-finding projects.

The problem is that graphics cards use single-precision math, while most projects require double-precision. You probably don't care if a pixel rendered in a game is off by 1/1000th of a pixel; but if the math in SETI is off by the same amount, it causes big differences in results.

There are probably ways around this (I know there are for integer math), but they might be so slow that the card would go from 40x faster than the CPU to something like 40x slower. (I'd at least expect no higher performance than the CPU itself.) We may have to wait a generation or two (or maybe ten) before GPUs can do SETI effectively.

 

[GLeeM]

Originally posted by: BlackMountainCow
... I doubt that they'll give away that advantage too easily. I bet they want people to give their CPUs to whatever DC project they want, but "all you GPUs are belon to us"!

Not so:

In this thread - Great News from Toms Hardware for F@H I quoted the following Pande quote from somewhere:

"Will also make our software available for others so they can push their work faster with GPU."

 

[BlackMountainCow]

@GLeeM

Ok, I didn't know that. That's pretty generous of Pande, though.

 

[bobsmith1492]

So, is there a F@H thing available yet for the GPU? I might have to sign up for it if it is... I want to put my computer to work. Right now I have the UD agent going on one core, but that leaves one free along with my x1900GT for F@H...

EDIT: Nevermind... I read the forum.

 

[Zoomer]

Don't hold your breath for nvidia gpu support. Read anandtech's articles about them both and you'll find that nvidia skimped on precision AND control flow ops.

Read: Slow. ati designed their chips with this gpgpu in mind. nVidia didn't, or at least, they didn't do anything that made it in.