CUDA for Milkyway@Home

@Travis:
I've posted the result file for the stripe 20 test WU in the other thread.

The app for ATI is almost ready, only some "convenience features" are still missing: (correct) checkpointing, release of the CPU during the GPU computations (CPU load is really low as the likelihood is now done on the GPU too), higher responsivity of the GUI/OS by adaptive limiting of the execution domain size (only important for slower GPUs).

The exactness of the single precision computation is astonishingly high. I've experimented with some tricks for the likehood stuff. Not only the usual Kahan summation in the likelihood kernel but also first order corrections for the exp and log functions extending them to more than 24bit precision. But frankly, it does not help at all, the resulting changes of the fitness are just at the noise level (10th decimal place). So I decided to take all that stuff out again (even not doing a Kahan summation within the convolve loop of the likelihood computation [but I do that in the integral part]) and restrict the Kahan summation to the GPU reduction (there it is required, as one won't get 9 decimal places precision without it). Maybe one has to do the more advanced corrections together in the integral and the likelihood parts, because the precision is limited by the worse of the two (or it does not matter at all), but actually I think the current state is already satisfying if you don't want to replace the double precision version completely ;)

Are there any news about the CUDA App for MW (running on Windows) ?

I would like to compare the "Power" of the CUDA-Cards running MW unit with Double Prec. like the current ATI-Cards.

Hi all, i have a video card based on nvidia 8600GT
After successfully compiled milkyway Cuda and installing, when MilkywayGPU elaborate a WU it's crash with "floating point exception". Why?
Anyone compile and successfully execute milkyway CUDA?

Thank's at all!

Vincenzo

is there ANY NEW NEWS on a CUDA app for MW?????

its been talked about alot but havent really seen

any updates on its progress.... is there or will

there ever be a CUDA app for MW ????

BTW AQUA's MT app kicks ass!!!!!!!!! no need for

GPU apps with the multi thread app,, and its a whole

lot faster then GPU.

---

Anthony Waters Posted this 18 Jul 2009 1:55:17 UTC

Hello,

This is Anthony, I'm working on the CUDA application with Travis.

There are a few things that need to be resolved in order to get the CUDA app released, namely

1. Checking the accuracy of the double precision math
2. Integration with BOINC

Anthony Waters Posted this 24 Jul 2009 21:58:58 UTC

The GPU app will go up as a beta application on the regular Milkyway@Home site when it is ready.

The final likelihood is accurate to about 12 decimal points, while we would like it higher, 12 is good enough. Therefore the next step is setting up the server, since it requires a major upgrade it will take a couple of weeks and will most likely be released with 0.19.

---

Travis Posted 22 Jul 2009 5:02:47 UTC

We actually have a working CUDA double precision app running right now, our new researcher Anthony is working on it. I think he's going to be releasing it as a beta application in the next week or so -- the only issue is that we're having an issue with the likelihood calculation that we want to resolve, when that happens we'll have to update all the applications.

Here is a new release of the GPU App source (v0.08)

(Binaries for Windows and Linux included)

http://milkyway.cs.rpi.edu/milkyway/download/code_release/mw_gpu_v0.08.tar
http://milkyway.cs.rpi.edu/milkyway/download/code_release/mw_gpu_v0.08.zip

please read the readme
http://milkyway.cs.rpi.edu/milkyway/download/code_release/readme

This will be the final release version that goes on BOINC pending any bugs and the new likelihood fixes. Also this release is more of a "does this work on other people's machines besides mine" so don't expect it to run with BOINC

(Binaries for Windows and Linux included)

http://milkyway.cs.rpi.edu/milkyway/download/code_release/mw_gpu_v0.08.tar
http://milkyway.cs.rpi.edu/milkyway/download/code_release/mw_gpu_v0.08.zip

please read the readme
http://milkyway.cs.rpi.edu/milkyway/download/code_release/readme

from the readme:

it takes around 90 seconds to execute on a GTX 285

By the way, the test WUs supplied in that package have the size of the (now quite rare) ~18 credits (stripes 79, 82, 86) or ~27 credit (stripes 20, 21) WUs.

Cluster Physik,

I do hope that you'll be able to work your magic and update the ATI application to be compatible with this. Is there any way to get it working better with the latest Catalyst / CAL versions? As always, your work is very much appreciated!

Cluster Physik,

I do hope that you'll be able to work your magic and update the ATI application to be compatible with this. Is there any way to get it working better with the latest Catalyst / CAL versions? As always, your work is very much appreciated!

In the moment it is compatible. The project has still to figure out what they want to change ;)

And I really don't know what is up with the new Catalyst drivers. In some configurations under Vista it runs quite well, but WinXP is a no go with anything newer than 9.2. If I find the reason, I will of course update the apps.

Hi development, i try to run milkyway cuda gpu on my 8600GT, but is not supported (cuda 1.1) you have drop the support under 1.3?

Thank's

Vincenzo

CUDA 1.3 is required for Double Precision support, single precision was tried, however, it only gave accuracy to 6/7 decimal places which is not enough.

So it will not work on my 8800 GTS as well?

---

ALL GLORY TO THE HYPNOTOAD!
Do You Dare?

CUDA 1.3 is required for Double Precision support, single precision was tried, however, it only gave accuracy to 6/7 decimal places which is not enough.

then your volunteers just took a nose-dive... i only have one card 1.3 compatible which is a gtx 260.... my 2 other cards are 9400 gt which are 1.1.....

CUDA 1.3 is required for Double Precision support, single precision was tried, however, it only gave accuracy to 6/7 decimal places which is not enough.

then your volunteers just took a nose-dive... i only have one card 1.3 compatible which is a gtx 260.... my 2 other cards are 9400 gt which are 1.1.....

It is the same restriction as for ATI GPUs.

Yep, excludes a lot more people though due to Nvidia's reluctance to move to double precision. Are there still plans to make a single precision GPU sister project?

So this will only work on GTX cards? I have 9800GT cards -- sounds like I'm not in the lucky subset at the moment.

CUDA 1.3 is required for Double Precision support, single precision was tried, however, it only gave accuracy to 6/7 decimal places which is not enough.

then your volunteers just took a nose-dive... i only have one card 1.3 compatible which is a gtx 260.... my 2 other cards are 9400 gt which are 1.1.....

It is the same restriction as for ATI GPUs.

---

gpugrid has currently
total active
Users 5,174 2,391
Hosts 8,322 3,575
source: Boincstats

and this is a mixed bag of gpu... most have above 50 cores...

ps3 is in this figure but a small percentage...8% i think... looking at boincstats...

and to add things over, AQUA has a gpu app out.

and seti lets any gpu cuda capable card run seti MB.


so, your pretty much, asking the people from other projects with high-end cards to come over play.. not many of those... i myself have 1 "1.3" card.

and lets mention the fact that, there are other projects with cuda app/gpu in the works as some of you know...

i have a ati 2600 xt card setting in a box...

---

I recommend Secunia PSI: http://secunia.com/vulnerability_scanning/personal/

Sorry zpm. You must be at least this tall to ride the attraction.

what, do i have to be as tall as Shaq. i'm 6'11" as is....


i'm just pointing out the fact that the volunteer base of gpu is very low, gpu computing with boinc is still in it's first year...