Rewriting the code should apply to all projects (since inefficient code couses us to waste the money on electricity bill instead of upgrading our machines). I just see that more projects are about to use CUDA (AQUA@home recently) and theire giving better deadlines for WU (my card is unable to meet the deadline for GPUGrid). So there will be a lot of competition between projects for the computing power. And the GPUGrid acts like it doesn't care. Right now the only project capable of using ATI cards is MilkyWay (F@H is not a boinc project) and You can see how much they gained. I just don't understand the choices of GPUGrid (and as i do not understand, it is hard for me to accept). It seems that GPUGrid is for the people with fast/new nvidia cards. So the owners of the slower nvidia cards and owners of ATI cards are not the target for GPUGrid...

On the below graph you can see the RAC of the GPUGrid



You cna see the point when SETI and AQUA launched their CUDA apps :D

I really would like to perform some work for GPUGrid as i think this project is more important than searching for the ETI or simulating the quantum computer. But unfortunately the project is beyond my limits (i am now finishing the first WU for AQUA CUDA and it is mainly because my card is good enough to meet the deadline, which is a month or so, not a week like it does in GPUGrid).

Hi,
don't get us wrong, we do care, but unfortunately our complex application is not portable on ATI with their current software support.

gdf

So maybe You should make a petition to ATI guys :D i would gladly sign it :D

Hi,
don't get us wrong, we do care, but unfortunately our complex application is not portable on ATI with their current software support.

gdf

Give Gipsel aka Cluster Physik, the creater of the MW_ATI-apl a try.
Send him the Code, he can than have a look if it is possible to port the code with Brooke/CAL to ATI cards.
Im shure he will handle the code secretly, so no 3rd person will see the code.
Just give him a try.

Gipsel helped them :D http://boinc.thesonntags.com/collatz/ and since you guys don't need double precission (like MW does) maybe You can benefit from ATI cards. Since there are alreday 5 projects using CUDA and only two using ATI. It will be more compettion for you on the CUDA power :D You really should consider creating ATI app...

any sign of the petition to the ATI guys??

Is anything moving in that topic? Do You guys even try to create ATI app??

There is no sign of an OpenCL for GPUs yet on ATI.

gdf

Do any of you asking for an ATI version want to contribute enough of a donation to allow hiring a computer programmer already specializing in whatever computer languages are already available for writing ATI card programs? Or would you prefer to wait for ATI/AMD to offer suitable compilers to allow using more familiar computer languages to write programs for ATI cards?

http://boinc.berkeley.edu/trac/milestone/6.10

I see no sign that includes any compilers for producing programs to run on ATI cards, only some support for running such programs once they are ready.

http://boinc.berkeley.edu/trac/milestone/6.10

6.10.x adds CAL/Brook+ support so that BOINC can use ATI cards. It doesn't add OpenCL support which is the longer term aim.

OpenCL is supposedly supported by ATI and Nvidia so the code is supposed to be compatible. My understanding is they will need to compile a version for each brand of card, but the source code doesn't need to change. The means that GPUgrid could develop a single version instead of having a CUDA version and a CAL/Brook+ version.

Unfortunately OpenCL isn't mature and the development is well behind what CUDA offers. CAL/Brook+ is also well behind CUDA in development. There is much postulating between the vendors about who's is better and who is supporting the OpenCL standard.

---

BOINC blog

We will support OpenCL. The day that ATI provides a working OpenCL implementation, we will be there.
gdf

Great news, I have 3 opencl capable ati cards waiting :D

The only project which need to be pushed towards gpu work is CPDN ;), those guys simply to lazy to rewrite their crappy old fortran code ;)

You might tell CPDN that Nvidia has announced that during the next few weeks, they will start providing a Fortran compiler for at least some of their chips, which should make the task of converting it to a GPU version much easier.

However, Nvidia does not seem to have gone to much effort to make it clear whether this compiler will work for the chips they're already selling, only the new 300 series of their chips they also plan to start making available in the next few weeks.

They also plan to start making some other compilers available at the same time, which may be of more interest to projects already using a few other computer languages, including C and C++. Same unclearness about whether those will work with the chips they've already sold as well.

One problem to expect with any project that currently has high memory requirements, though: Don't expect them to be able to get a full speedup by using all the GPU cores; only as many as the graphics card has memory for, at least until they do a significant rewrite to let many of the remaining GPU cores do something useful with much less memory.

For similar reasons, don't expect very many BOINC projects to start making both Nvidia GPU versions and ATI GPU versions available until some time after both Nvidia and ATI start making enough compilers available that the same computer language can be used to write both versions.

Also, expect some BOINC projects with algorithms with enough things that need to be done one after the other to start offering GPU versions that essentially run multiples of the workunits now offered, each on a small cluster of the available GPU cores, with similar restrictions on whether there's enough GPU memory to run very many of these CPU clusters at once.

I think that this is a misconception probably given by the marketing people.
My opinion of what they mean by giving a C++ or fortran compiler is that they will support CUDA development in C++ (transparently) and fortran. Now this is only possible by using C within fortran or C++.

So good, but nothing really exceptionally different from before.

gfd

The only project which need to be pushed towards gpu work is CPDN ;), those guys simply to lazy to rewrite their crappy old fortran code ;)

I took a look at the CPDN project and found the following:

1. Their workunits are so large that I doubt that many Nvidia graphics cards would have enough memory to run them on more than one GPU core at a time. So don't expect a GPU version to be much faster than a CPU version, if they ever decide to have both.

2. Their applications have gone over a year without any changes. Do you expect that to continue after a major update, such as adding GPU use or rewriting the program in a different computer language?

3. Their description of their algorithm sounds like enough things have to happen AFTER other things that trying to split the calculations onto different GPU cores running at the same time wouldn't speed up the workunit. It might, however, allow running more than one workunit at once on the same GPU, but only if there is enough graphics memory.

4. Many of their workunits take months to run. How long would you expect it to take to test a new version that won't run significantly faster?

However if you participate there, you might give them this link and ask them to check whether Nvidia's new FORTRAN compiler for their chips at least makes the conversion to GPU use easy enough that they're willing to try it anyway, although without trying to rewrite their applications in any other computer language:

http://www.gpugrid.net/forum_thread.php?id=1406

Note, however, that Nvidia does NOT make it clear whether their new compilers will work for any of their chips other than the GT300 that they haven't quite started selling yet.

When ATI starts offering a similar selection of compilers, and they work well enough, I'd expect to see more BOINC projects start offering both Nvidia and ATI versions of their applications, even those that now offer only CPU versions; but rather few before then.

http://developer.amd.com/GPU/ATISTREAMSDKBETAPROGRAM/Pages/default.aspx

is it the news we were waiting for?

Almost.
It still lacks atomic operations which are required by ACEMD.
GDF

maybe you could send in a request as a research organization. Maybe that'll speed up the process :)