Latest drivers (NVidia and ATI)

[RottenMutt]

do you have any idea what is causing the optimized app to thrash.  currently i'm running the v.2 installer i believe.  what is the opt cuda MB build?
what is the speed difference between the stock seti cuda app and the optimized app, would four stock threads be faster then two optimized apps; what about the extra power for running four threads?

[Raistmer]

opt app reduces CPU load time (importand for fast GPU cards - less idle time) + has some tweaking of threshold where one or another pulcefinding kernel used.
+ few more little modification positively affecting speed.

Why it can result in driver crash - maybe because of changed threshold.

x4 build will lock each instance on separate core. It could help with GPU feeding in multi-GPU environment.

Another possibility of crash - AFAIK last available build from installer was compiled against CUDA 2.2 SDK, not 2.3 one. Usually it doesn't matters but maybe there is some subtle sync issue that shows up only on heavely loaded systems... Maybe it's worth to do build refresh versus SDK 2.3

[Sutaru Tsureku]

Maybe it's possible that someone DL the CUDA toolkit and publish only the CUDA V3.0 .dll's?

'developer.nvidia.com/object/cuda_3_0'

Maybe officially - someone of the opt. crew and here in the DL area?

Thanks!


BTW. nVIDIA driver 197.13 is already available.

[Richard Haselgrove]

Why?

[Sutaru Tsureku]

Why?

For to make an 'easy test' of CUDA V3.0 and the newest nVIDIA driver..

Why to install a whole toolkit if I need only the two CUDA .dll's?

[Richard Haselgrove]

I would have though it better to go the other way round - testers download the DLLs quietly and privately, make the tests, publish the results.

If there is significant improvement (as I found with v2.3), then publish the DLLs for general download and write instructions on how to install them and what preconditions apply.

But I think v3.0 was tested months ago, and found to have no benefit for current hardware/applications, so there was no need to waste time on the follow-up.

[Sutaru Tsureku]

How to get out two .dll's out of one DL file?
I need to install the whole toolkit.
I don't know what will be installed and where.

It was CUDA V3.0 BETA which was tested in past, right?

Now it's CUDA V3.0 (final). Nothing different?

IIRC, some people saw speed up with CUDA V3.0 BETA. Very small, but little bit.

So why not to publish CUDA V3.0 (final) .dll's and all people could make a test of their hardware?

Not needed to make an instruction, only like the CUDA V2.3 .dll's in the DL area.
So how I see, for CUDA V2.3 there isn't an instruction.
http://lunatics.kwsn.net/index.php?module=Downloads;sa=dlview;id=208

But, why an instruction, only for to copy two files in an folder?


EDIT: BTW. Now is an other/newer nVIDIA driver available. Maybe this time it match better with the CUDA V3.0.

[Richard Haselgrove]

OK, here you are. These are the 32-bit v3.0.14 DLLs from the public toolkit, renamed and ready to drop into a project folder (didn't take long to download and unpack them, did it?)

You can test and report back. If your test report is convincing enough, I might even update my drivers to the newer, slower version!  

[attachment deleted by admin]

[perryjay]

Morning Richard,
I tried your 3.0 cudart and cufft files and updated to the latest 197.13 driver for my 9500GT. It immediately errored out 6 or 7 work units with exit code 1073741515.  I stopped BM and put 2.3 stuff back in and it is running fine.  I was running 195.62 and 197.13 seems to be running faster but it is only on the first work unit since I went back to 2.3. I will keep an eye on it and report back later after I've run a few.

[Richard Haselgrove]

Google (s)he say: install Dot Net.

Not sure I believe it.

[perryjay]

I saw that too but I think I've already got dot net. I'll leave it like this for awhile and see how it does.

[Sutaru Tsureku]

OK, here you are. These are the 32-bit v3.0.14 DLLs from the public toolkit, renamed and ready to drop into a project folder (didn't take long to download and unpack them, did it?)

You can test and report back. If your test report is convincing enough, I might even update my drivers to the newer, slower version!  

Thanks a lot!

Will make the test soon.
Tomorrow I'll search eggs with my family in the garden. 
Maybe Monday or Tuesday I'll install and after a few results I'll report how it was.


EDIT: BTW. I have not fast DSL, so I needed more time for DL.
After DL of the CUDA toolkit and execution of the .exe, what will happen?
Only an extract of the file in the folder, or also an installation of something?

[Sutaru Tsureku]

I made my test before I gone out to the garden..

Hey, I was surprised! 
It's look like 197.13 is fast like 190.x (with CUDA V2.3).
I made a test with 190.38, 197.13 and 190.62 .

Maybe someone could confirm my result with some bench WUs?

BUT, if I copy the two CUDA V3.0.14 .dll files (with installed 197.13) in the project folder I get only errors.
Immediately, the BOINC Manager isn't longer useful. I needed to stop boinc.exe in TaskManager.

Example out of the stdoutdae..

Code: [Select]

Starting xxxxxxxxxxxxxxxx.26_1
Starting task xxxxxxxxxxxxxxxx.26_1 using setiathome_enhanced version 608
.. ~ 2 sec. after..
Computation for task xxxxxxxxxxxxxxxx.26_1 finished
Output file xxxxxxxxxxxxxxxx.26_1_0 for task xxxxxxxxxxxxxxxx.26_1 absent
After restart and UL/report I see:

Code: [Select]

Exit status -1073741515 (0xffffffffc0000135)

<core_client_version>6.10.18</core_client_version>
<![CDATA[
<message>
 - exit code -1073741515 (0xc0000135)
</message>
]]>
So CUDA V3.0.14 don't work on my manufacturer OCed GTX260-216.
Why it don't work?
The MB_6.08_CUDA_V12_app don't fit together with CUDA V3.0 ?


Also I saw with 190.38 my GPU had ~ 850 MB free, after app init in <stderr_txt>.
The 197.13 driver let free ~ 860 MB. So ~ 10 MB less GPU RAM used from the CUDA app.

GPU-Z V0.4.0 say with 197.13, that 246 MB GPU RAM is used.
IIRC, before with 190.38 it was 252 MB.

Hmm.. what is better?

[Pepi]

Solution is simple
rename cufft32_30_14.dll and cudart32_30_14.dll  to cudart.dll and cudafft.dll and replace it
then copy again but without rename
so in the end of process you will get
cudart32_30_14.dll
cudart.dll
cufft32_30_14.dll
cufft.dll

[Sutaru Tsureku]

No, this wasn't the prob. 
Richard was so kind and renamed the .dll's for easy usage.
Only extract of his attached file, copy to the project folder.

On your hardware the CUDA V3.0.14 work?
Which system?


EDIT: BTW. I have WinXP 32bit, SP3 - (if this is important)