With ATI's new 4800 series, there is going to be a lot of unused processing power out there. The 4870 can do 1.2Tflops and the 4870x2 will be significantly more - hopefully close to double. With 4 of those in crossfire config, that might be as much as 8 Tflops. For comparison, a stock q6600 is less than 10 Gflops - so basically a difference of nearly 3 orders of magnitude. Somebody has to want to tap into that.

Hello twilyth, hello Sekerob,



If my understanding of Sekerob's post (Jul 13, 2008 9:08:05 AM) is any indication of the "prospects of GPU programming for DC purposes", it seems to me that the idea of tapping into a GPU's power is equated to the idea of tapping the power of the Sun to potentially supply the entire power needs of all nations for centuries and beyond. The power is doubtless there, and everyone wants it. Could we use it? In the case of the GPU, well, you be the judge.

Not that the idea of a GPU crunching DC projects is not desirable from a simply desirability point of view. The sticking point, it seems to me, is that many people dismiss the idea straightaway, equating desirability to the means to realize the object of the desirability. To these people, if there is no current means to an end, then the desirabilty of that end is thrown out the window. For example, if there is no means or technology to go to the moon, then one shouldn't want to go to the moon, possibly labeling (with uncomplimentary words) those who harbor the idea.

On the same vein, if there is no way that is currently available for a GPU to be used to crunch projects, then, the thinking goes, one shouldn't want to put effort into investigating approaches to find some means to have a GPU do the work of a CPU (in crunching projects). Thankfully there is work (at Folding@Home and possibly elsewhere) involving GPU programming to crunch projects. They have a tough and difficult work ahead of them before GPU crunchng would be as common as, shall we say, going to the moon!

The programming of a GPU program is the aspect that interests me. Technically, BOINC 6 can run GPU programs. (It reserves the whole GPU for BOINC.) So the remaining practical difficulties are for us to check, support and maintain a GPU program. Then there is the practical aspect - just how many members will run it on their GPU. This last point is very important. Folding@Home can expect to get everybody who wants to run distributed computer programs on their GPU. Once there are multiple choices, this happy state of affairs is over.

I have not read much comment from GPU programmers. Therefore, I have to guess. My guess is that it is crucial to start with a program that is organized around data-parallel semantics. This is the crucial step. After that, it is just ( ) grinding out the code. Just off the top of my head, it sounds as though the original program (pre-GPU) should be written in APL ( http://en.wikipedia.org/wiki/APL_(programming_language) ) or written to use the BLAS (Basic Linear Algebra Subroutine) library ( http://en.wikipedia.org/wiki/Basic_Linear_Algebra_Subprograms ). This would organize the program so that it would be (relatively) easy to use SIMD instructions (SSE extensions) or code for the PS3 or any GPU.

Do current programming classes stress the importance of using the BLAS library? For a long time, parallel programming was an abstract computer science idea that was only implemented on special hardware. Now that it is starting to show up on standard real-world computers, have teachers adapted to this change?

WCG just boards the programs given to us to run on BOINC. The project scientists will have to actually design and write the application program. Unless they write it from the start for parallel computing, it is likely to be a multi-year task to adapt it to a GPU.

Does anybody have some additional ideas about how to structure a program to run on parallel hardware?

Lawrence

Not a good sign if there is a multi-core with 1 GPU in the system or even multiple GPU's and it says:

"All of these are in principle possible, but not tested yet.
You should be able to use 4GPUs cores as long as you have a quad core CPU. If it does not work, we will fix it. The only known issue you have to deactivate SLI support. CUDA works only without SLI."

So you have this power system which, you want to crunch on multiple cards and then one is told to disable a key feature why the hardware was procured in the first place. Who's doing that?

And then it says you need Linux to run it.... very mainstream and tinker-free crunching for the masses.

Prediction when first seen on a main grid using windows? Not venturing out on that limb. My accelerator of NVidia is not even recognized by current BOINC 6. developers version. I'll see it when i see it.