Latest drivers (NVidia and ATI)

[Pepi]

Yes I have two crunch machines both with 64 bit Win 7
One have 9800 green and another is GT 240 and both computing via cuda 3 dlls without problem

[Richard Haselgrove]

Yes I have two crunch machines both with 64 bit Win 7
One have 9800 green and another is GT 240 and both computing via cuda 3 dlls without problem

What's your experience on the speed question?

Edit - and did you try it with only one of the files doubled-up? I can see you would need a named copy of cudart32_30_14.dll, because cufft32_30_14.dll depends on it: but there's no reverse dependency. Does it matter that cufft32_30_14.dll depends on itself ()

[MarkJ]

I wonder if it would be possible for one of the developers to compile it with the cuda 3 toolkit, so that way we get a true cuda 3 app to try.

[Pepi]

Yes I have two crunch machines both with 64 bit Win 7
One have 9800 green and another is GT 240 and both computing via cuda 3 dlls without problem

What's your experience on the speed question?

Edit - and did you try it with only one of the files doubled-up? I can see you would need a named copy of cudart32_30_14.dll, because cufft32_30_14.dll depends on it: but there's no reverse dependency. Does it matter that cufft32_30_14.dll depends on itself ()

There is no difference in speed. I try all combination but only one that I wrote few post above worked. Both dll  MUST  be renamed and also in original name state. And if you dont trust me: try yourself: the only thing you will get will be few "computation error" results 

[Richard Haselgrove]

That's funny - I came to the board to post that

cudart32_30_14.dll
cudart.dll
cudafft.dll

is running just fine on host 3751792.

But no increase in speed.

[perryjay]

Hmmm, interesting.

Ok, I took Richard's renamed files, copied them and renamed the copies back to the original 32_30_14.dll and put all four into my SAH folder. This time nothing errored out but one work unit that had started before I did this would not restart. I let it try for two or three minutes but the percent did not move. I suspended that one and the next in line started with no problem. For grins, I suspended that one and it went back to the first one. Again, after about two minutes it still had not started so I aborted it. It picked up on the second one where it had left off and is running fine so far.  As for speed, I seem to have sped up a few seconds with this new driver so I guess I shall stay with it.

[Sutaru Tsureku]

Ohh well, maybe a stupid question..

How should look the app_info.xml?
What's needed to mentioned?

This would be well?

Code: [Select]

.
.
<file_info>
<name>MB_6.08_CUDA_V12_VLARKill_FPLim2048.exe</name>
<executable/>
</file_info>
<file_info>
<name>cudart.dll</name>
<executable/>
</file_info>
<file_info>
<name>cudafft.dll</name>
<executable/>
</file_info>
<file_info>
<name>libfftw3f-3-1-1a_upx.dll</name>
<executable/>
</file_info>
.
.The same in the <file_ref> part?

Or I need to insert the other 2 file names, cudart32_30_14.dll and cudafft32_30_14.dll ?

Thanks!


BTW. Maybe we would see more speed up, if the current MB_6.08_CUDA_V12_app would be new compiled with the new CUDA V3.0 toolkit/or other prog?
I guess in near future it should be done for the new Fermi chips, or? But if we would see also speed up with the 'old' hardware?

[Pepi]

Your app.info is OK, you dont need to change anything. It will work fine.

[Sutaru Tsureku]

IMPORTANT!
The .dll files must be named:
cudart.dll
cudart32_30_14.dll
cufft.dll
cufft32_30_14.dll

The fft files without 'da'!

If you write cudafft.., you will trash your DLed WUs.


Hey guys..
I thought I'm stupid.. I'm crazy..
I done this and this, tested this and tested this..
CUDA V3.0 didn't worked.

Then I 'installed' the CUDA toolkit V3.0 and saw that the .dll file is named different.
Like the CUDA V2.x .dll's.

Next time I'll not believe all immediately what I read..

Now it work 197.13 and CUDA V3.0.14 .
Will look how the calculation times are..

[glennaxl]

You only need to duplicate the RT.
So it would be like this:
cudart32_30_14.dll
copy of cudart32_30_14.dll -> rename to cudart.dll
cufft32_30_14.dll -> rename to cufft.dll


Here's some good stuff. Some improvement (1%-2%) with 197.13 compared to my last test.

driver : 197.13
CPU: 100%
GPU: gtx 295 core 0

[attachment deleted by admin]

[Sutaru Tsureku]

But, ..maybe stupid question.. 
Why I need at least 3 .dll files of CUDA V3.0 instead of 2 like in past with CUDA V2.3?
If I don't count the libfftw3f-3-1-1a_upx.dll .

Why not useage only of cudart32_30_14.dll & cufft32_30_14.dll (original named) and change the entries in the app_info.xml?

[Richard Haselgrove]

Because the application (SETI MB CUDA) only knows the names "cudart.dll" and "cufft.dll".

But the new v3 cufft.dll expects to have "cudart32_30_14.dll" available.

[MarkJ]

Because the application (SETI MB CUDA) only knows the names "cudart.dll" and "cufft.dll".

But the new v3 cufft.dll expects to have "cudart32_30_14.dll" available.

If we can get it recompiled using the latest toolkit we wouldn't need to do this, although it would need testing. We might even be able to get a 64 bit version, although i'd doubt it would offer any speed advantages.

[Sutaru Tsureku]

[...]
Here's some good stuff. Some improvement (1%-2%) with 197.13 compared to my last test.

driver : 197.13
CPU: 100%
GPU: gtx 295 core 0

Which nVIDIA driver you tested with CUDA V2.3?
I guess you used only 197.13 with CUDA V2.3 and CUDA V3.0.14 in your test, right?

I have the feeling that 197.13 + CUDA V3.0.14 is ~ 1 - 2 % slower than 190.x + CUDA V2.3.
Have someone tested this driver/CUDA combinations?

[glennaxl]

Which nVIDIA driver you tested with CUDA V2.3?
I guess you used only 197.13 with CUDA V2.3 and CUDA V3.0.14 in your test, right?

My test was CUDA v2.3 vs CUDA v3.0.14 on MB_6.08_CUDA_V12_VLARKill_FPLim2048.exe