Ok guys. The Code for the CUDA app is out. Who is able to help Travis. Look into the Code Application Discussion Thread

Hopefully we can get it compiling properly so we can open up the milkyway_gpu project :) That should really help with the server load.

Dave P. and I will be working on that in the next week or so, and hopefully get work available for the GPU applications using the newer code. Once work is flowing smoothly there, we'll stop awarding credits to GPU applications here.

The milkyway_gpu workunits should be around 100-300 times more work (or even more!) than the current milkyway@home workunits, which will really help the server load.

---

Also, sorry for my lack of posts recently. I've been locked in a cave with no internet access and only a CUDA manual and GPU. They threw me scraps of food occasionally.

---

Looking to Cluster Physic's changes to the CUDA client modified so that ATI HD 38xx and 48xx GPUs can run under the anonymous platform for MW@H_GPU coming online soon.

Travis

I presume you will turn off the MW@H ATI client work acceptance when the third party ATI software is ready for MW@H_GPU, and not before?

---

Go away, I was asleep

Looking to Cluster Physic's changes to the CUDA client modified for ATI HD GPUs.

Travis

I presume you will turn off the MW@H ATI client work acceptance when the third party ATI software is ready for MW@H_GPU, and not before?

That's the plan. Since the Milkyway_GPU@Home code basically calls the same evaluate kernel multiple times, it should be very simple to change over the ATI client.

Considering it will probably take us a week or two to get work flowing on the milkyway_gpu@home project, there should be more than enough time for that :)

---

Thanks Travis

It makes it easier for me to confirm my intended purchase of another ATI card, and then look long term for a few nVidia beasties for the Milkyway_GPU work. I am currently sporting an old HD3850 on an AGP slot.

---

Go away, I was asleep

Looking to Cluster Physic's changes to the CUDA client modified for ATI HD GPUs.

Travis

I presume you will turn off the MW@H ATI client work acceptance when the third party ATI software is ready for MW@H_GPU, and not before?

That's the plan. Since the Milkyway_GPU@Home code basically calls the same evaluate kernel multiple times, it should be very simple to change over the ATI client.

Considering it will probably take us a week or two to get work flowing on the milkyway_gpu@home project, there should be more than enough time for that :)

Could you also, in addition to confirming that the code / project should not be considered rock-solid stable at this point, take some time to reiterate that once the GPU apps are functioning properly over there, GPU apps will not get credit here? There seems to be a contingent of people who are using GPUs that want to be able to use both projects with their GPU. I don't have any issue with them using a CPU here, but allowing GPUs to come over here if there is any credit disparity between here and there or work availability issue there will just bring this project back to a crawl.

Thanks!

Brian

Travis, until the GPU project has an available ATI GPU application, I'd suggest that you NOT stop awarding credits to GPU applications over here otherwise, you'd be stranding folks -- not a good thing.

Hopefully we can get it compiling properly so we can open up the milkyway_gpu project :) That should really help with the server load.

Dave P. and I will be working on that in the next week or so, and hopefully get work available for the GPU applications using the newer code. Once work is flowing smoothly there, we'll stop awarding credits to GPU applications here.

The milkyway_gpu workunits should be around 100-300 times more work (or even more!) than the current milkyway@home workunits, which will really help the server load.

---

Could you also, in addition to confirming that the code / project should not be considered rock-solid stable at this point, take some time to reiterate that once the GPU apps are functioning properly over there, GPU apps will not get credit here? There seems to be a contingent of people who are using GPUs that want to be able to use both projects with their GPU. I don't have any issue with them using a CPU here, but allowing GPUs to come over here if there is any credit disparity between here and there or work availability issue there will just bring this project back to a crawl.

Thanks!

Brian

Yeah, the CUDA code is not rock-solid stable by any means. The whole point in making milkyway_gpu@home is so that we can give the GPUs MUCH longer workunits and greatly reduce the server strain so that we can keep work smoothly flowing to all our users.

Allowing GPUs to continue crunching regular MW@Home would just be self defeating in that regard. The GPU apps should get the same credit here or there, so there will really be no benefit to for a GPU app to continue running regular MW@com. That and the fact that once MW_gpu@home is up and running smoothly we'll cut off the credits to GPU apps here at regular MW@home :P

If people with GPUs want, they should still be able to run regular MW@home on their CPU, and MW_gpu@home on their GPUs for double the crunching fun. :)

---

Travis, until the GPU project has an available ATI GPU application, I'd suggest that you NOT stop awarding credits to GPU applications over here otherwise, you'd be stranding folks -- not a good thing.

We have no plans on doing that at all. I have a feeling that it won't even be an issue, given the state of the ATI GPU code and Cluster Physik's skills. I'm half expecting that there'll be a working version of the ATI GPU application for milkyway_gpu@home before we get the CUDA version all ironed out :P

---

Could you also, in addition to confirming that the code / project should not be considered rock-solid stable at this point, take some time to reiterate that once the GPU apps are functioning properly over there, GPU apps will not get credit here? There seems to be a contingent of people who are using GPUs that want to be able to use both projects with their GPU. I don't have any issue with them using a CPU here, but allowing GPUs to come over here if there is any credit disparity between here and there or work availability issue there will just bring this project back to a crawl.

Thanks!

Brian

Yeah, the CUDA code is not rock-solid stable by any means. The whole point in making milkyway_gpu@home is so that we can give the GPUs MUCH longer workunits and greatly reduce the server strain so that we can keep work smoothly flowing to all our users.

Allowing GPUs to continue crunching regular MW@Home would just be self defeating in that regard. The GPU apps should get the same credit here or there, so there will really be no benefit to for a GPU app to continue running regular MW@com. That and the fact that once MW_gpu@home is up and running smoothly we'll cut off the credits to GPU apps here at regular MW@home :P

If people with GPUs want, they should still be able to run regular MW@home on their CPU, and MW_gpu@home on their GPUs for double the crunching fun. :)

I don't know whether it is misinterpretation of the "cutting credits off" wording that you're using, or that the people that are seeming to want both projects available to them with a GPU are concerned about work availability...

Like you said though, allowing GPUs to continue here partially (or totally depending on the amount of people who stay) defeats the purpose...

Interpretation of the "cutting credits off" wording that you're using, or that the people that are seeming to want both projects available to them with a GPU are concerned about work availability...

Well to put it simply, if the GPU applications stay at milkyway_gpu@home, and the CPU applications stay at milkyway@home, there shouldn't be any work availability issues for anyone.

If for whatever reason GPU users still try to crunch milkyway@home after milkyway_gpu@home is up and running smoothly, there will be work availability issues.

Like you said though, allowing GPUs to continue here partially (or totally depending on the amount of people who stay) defeats the purpose...

Yes, it would. Since credits will be the same here and at milkyway_gpu for the same amount of work, there's no reason for a GPU to continue crunching milkyway@home once the milkyway_gpu@home is up and running. If anything they'll be getting more credit letting their GPUs crunch milkyway_gpu and their CPUs crunch milkyway@home. Not to mention that will drastically reduce their bandwidth usage :D

---

Hey, Travis?

Thanks. Seriously.

-jim

So once MW_GPU is up and running good, my CPU's should be able to get a steady supply supply of work at MW_CPU and I can finally try to reach 1 million credits?

And I will not need to babysit the CPU's?

Sounds good to me. But I think that is still a little bit off though...

---

Never surrender and never give up. In the darkest hour there is always hope.

Will there be any changes to the current CPU app?

---

Reading what Travis has posted here, it would seem that the "Official" GPU app. will be CUDA, with continuing support for the existing ATI app. and that the CPU apps. will carry on as now, with the only difference being that there should always be work for the original CPU project, as the WUs for GPU will be generated and handled on their own under the MW@Home_GPU banner.

Please correct me if I'm wrong! :)

---

Seejay **Proud Member and Founder of BOINC Team Allprojectstats.com**

Reading what Travis has posted here, it would seem that the "Official" GPU app. will be CUDA, with continuing support for the existing ATI app. and that the CPU apps. will carry on as now, with the only difference being that there should always be work for the original CPU project, as the WUs for GPU will be generated and handled on their own under the MW@Home_GPU banner.

Please correct me if I'm wrong! :)

As I read his intent it will be that once the apps are up and running, both ATI GPU and CUDA GPU will be supported on the GPU project side only. CPU hosts will be supplied from this project side. What has been killing the server here is the fact that the 30 min to hour tasks can be run off in seconds on a GPU. Which is fine for the participant doing that, but it kills the server. So, to rein that in the project is being split.

Unlike SaH where they feed all from one pot.

In a way this is nice because you can attach to both projects, set different Resource Shares and then proceed. For example, I may want to not share as much GPU of MW with GPU Grid ... so I could set a lower share on MW GPU and yet still have a high CPU side share ...

This is only one of the flaws in the current mind model of Dr. Anderson on the addition of GPU processing to BOINC...

On one hand I did not want to split my MW credit, on the other, in a way it makes it nice because (theoretically) you can separate who earned 1M CS with CPU processing and who did it with GPU processing ... of course in the long run we all be dead ... :)

Reading what Travis has posted here, it would seem that the "Official" GPU app. will be CUDA, with continuing support for the existing ATI app. and that the CPU apps. will carry on as now, with the only difference being that there should always be work for the original CPU project, as the WUs for GPU will be generated and handled on their own under the MW@Home_GPU banner.

Please correct me if I'm wrong! :)

That's pretty much it. ATI and CUDA on milkyway_gpu, and regular CPU's here. If/when OpenCL takes off (which should support both ATI and Nvidia cards) we'll most likely use that.

---

Reading what Travis has posted here, it would seem that the "Official" GPU app. will be CUDA, with continuing support for the existing ATI app. and that the CPU apps. will carry on as now, with the only difference being that there should always be work for the original CPU project, as the WUs for GPU will be generated and handled on their own under the MW@Home_GPU banner.

Please correct me if I'm wrong! :)

That's pretty much it. ATI and CUDA on milkyway_gpu, and regular CPU's here. If/when OpenCL takes off (which should support both ATI and Nvidia cards) we'll most likely use that.

And hopefully the Larrabee :)

I did see a Cluster Physik post that you've changed the double precision requirement so that only single precision is needed? Would just like to confirm that - going to a theoretical max of 2.4TFlops on a 4870X2 (now 240GBP and falling) is a big leap.

---

And hopefully the Larrabee :)

I did see a Cluster Physik post that you've changed the double precision requirement so that only single precision is needed? Would just like to confirm that - going to a theoretical max of 2.4TFlops on a 4870X2 (now 240GBP and falling) is a big leap.

We'll be using both single and double precision results on milkyway_gpu. The double precision is still needed for the last stages of optimization but until then the single precision is good enough. So we'd really like cards which support double precision to use double precision (and credits will be awarded appropriately for that).

We'll be using the single precision results more for "exploration" in the search space, and the double precision results for "exploitation" of interesting search areas.

---