Hi Movieman,
Actually, there aren't all that many CUDA-capable GPUs out there. So any decent sized project is likely to last more than a year while gathering attention and publicity from a new group of users. And if we could get a project that ran on a console - wow!

We are definitely not discouraging project scientists from using these new technologies. But it is difficult to find somebody with a good science problem who also wants to take on the computer hassle. Most just want to adapt a current program that runs on Linux in their department to run on a big grid. For that matter, a small number of our projects use the same underlying program code. This happens when a project attracts attention among researchers who want to do the same thing on a different problem.

Lawrence

Lawrence and Jean;
Thanks for the info. Puts a new light on the subject.
I do however think you'd be surprised at how many CUDA enabled GPU's are out there.
I see just what our Xs FAH team has up and running and it's scary.
guys running Quads with 2-nVidia GTX 9800X2's or GTX280's in SLI.
Monster machines in FAH..
We also have quite a few WCG guys with those cards that although they run WCG on their cpu's rub FAH on the GPU at the same time.
Thanks again for the info.

Oh, I wish you would give a real and honest answer, but in all threads related to GPU use, you've made your point clear: you consider them all hype, no substance. A few months ago, it was debatable. Today, GPU in grid computing is used by F@H on a large scale and has been proved to be reliable, scalable, and effective. Lack of double precision arithmetic was one of your best arguments. Sorry for your sunshine, but that argument is gone with the new chips.

Did I pretend it was easy ?

Even without using GPUs, the need for some parallelism in the WCG applications is clear. In the current situation, when a quad core executes 4 work units at the same time, it uses 4 times the memory. With 6 and 8 cores CPUs just around the corner, the problem will only get worse. Using all the available cores to solve one work unit faster would bring the memory requirement back to a more reasonable level and make a better use of 3rd level CPU cache. Keeping a small, unobtrusive footprint is among the most important requirements for broad acceptance of grid processing.

Check the facts. NVidia is already sponsoring an Autodock adaptation for CUDA and reports a 12 to 20 times speedup :
http://www.nvidia.com/object/io_1209593316409.html
http://www.siliconinformatics.com/

Still believe it's impossible or not worth the trouble ? Just look at what F@H is doing !

When the 8087 appeared, some people quickly understood how important this was while others were doing the same kind of criticism, pretending that a floating point coprocessor was a waste of silicon and a faster CPU with software FP routines would be better...

Not quite correct. F@H ist still providing the GPU clients as beta version (I also crunch there). The team there is also still looking into some bugs and optimizations for the various GPU clients.
Yes, you are correct, the GPU clients work great - specifically the NVIDIA - and the team there is very excited, but please keep in mind that f@h does not use the BOINC client.


WCG is not only providing Autodock adaptation - and not everyone has a NVIDIA card (e.g. I use ATI). Anyways, the team at WCG are not the one who do the programming. You can address it to the scientists - how about that.


I'm sure a lot of people look what f@h is doing (not only WCG) - the opportunities are great ... no doubt. But you are jumping the gun. Just because one project (or maybe two) use the GPU (and just a BETA) doesn't mean other projects must use it immediatly as well. Please don't get me wrong, I also would like to see the WCG projects using the GPU, but give the people some time to look into it.
If you want to and even can help please contact the project teams.

Cheers,

Oh, this is priceless: "Multi-processing is a hard problem in computer science. It's been there for 30 years. It's not answered by software tools." That was Nvidia's Andy Keane, trying to downplay Larrabee last August. But a few months earlier, in the very press release that Eric-Montreal pointed out, it talks of "new user-friendly programming languages that will allow developers to exploit parallelism automatically"....

What a joke. Andy Keane is right, parallelism is hard. Even great programmers struggle with it, and the hash made by lesser programmers - well, I have seen examples of multithreaded code that would make you weep. Programming tools can only go part of the way. The rest requires a different way of thinking.

But, oh, the hypocrisy of Nvidia complaining about Intel's marketing!