CUDA vs nVidia OpenCL User Poll
log in

Advanced search

Message boards : News : CUDA vs nVidia OpenCL User Poll

Previous · 1 · 2 · 3 · 4
Author Message
Profile pons66
Send message
Joined: 8 Sep 12
Posts: 5
Credit: 690,559,617
RAC: 0
Message 18627 - Posted: 17 Feb 2014, 19:14:00 UTC

I'd prefer CUDA!

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2528
Credit: 740,580,099
RAC: 0
Message 18628 - Posted: 17 Feb 2014, 19:16:27 UTC

You should be getting Windows and Linux CUDA WUs now.

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2528
Credit: 740,580,099
RAC: 0
Message 18629 - Posted: 17 Feb 2014, 20:08:28 UTC - in response to Message 18605.

Source of the below info : PrimeGrid Preferences :

CUDA is usually faster on pre-Kepler GPUs (GTX 5xx and older).
OpenCL is usually faster on Kepler GPUs (GTX 6xx and newer). Use whichever is faster.


Nothing is impossible. It isn't build into BOINC which means lots of custom code.


=> Possible to choose between OpenCl and CUDA + adapt the "block size"
This can be used to vary the kernel size within GeneferCUDA

If possible, this would be the ideal solution.


You already can! That what the "items_per_kernel" config setting is. In fact, you get even more control since you can also control the local work group (a.k.a. threads), kernels per reduction (e.g. summary and verification) and sleep time in milliseconds between checks whether the kernel has completed.


Every cruncher will be able to try CUDA and OpenCL for nVidia and use whichever is faster / optimal.

If not possible to implement, than solution A.

Again, Thank You Very Much, and please accept my appologies for this "second vote".

Philippe


The CUDA app uses the same config settings as the OpenCL app and after optimizing for the GPU, they seem to be pretty close in speed. That wasn't the case with older OpenCL versions. But, as you said, people will be able to choose. It seems that most want CUDA. But those that want to use the OpenCL version will be able to via either the app_config or app_info XML files. So, the choice will be manual for the time being. If I can come up with a generic way to choose the plan class in preferences and select it via the scheduler, I'll give that a try so the choice won't require installing anything manually.

alantiagofelipe
Send message
Joined: 11 Aug 09
Posts: 1
Credit: 318,017
RAC: 0
Message 18711 - Posted: 24 Feb 2014, 1:35:57 UTC - in response to Message 18629.

I think there should be an adaptive way that would adjust automatically to the best performance found in the hardware in question. As described in Alternative "C", which I believe should be the default.

Either way, the user should have the possibility to change the default for each host, between the above-mentioned alternative ("C"), the Cuda ("A") and OpenCL ("B").

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2528
Credit: 740,580,099
RAC: 0
Message 18723 - Posted: 26 Feb 2014, 20:27:16 UTC - in response to Message 18711.

I think there should be an adaptive way that would adjust automatically to the best performance found in the hardware in question. As described in Alternative "C", which I believe should be the default.

Either way, the user should have the possibility to change the default for each host, between the above-mentioned alternative ("C"), the Cuda ("A") and OpenCL ("B").


There already is. Kind of. Choose both applications in your preferences and the BOINC client is supposed to figure out which is faster. For those who don't think the BOINC client is really smart enough to do that, they can choose just one - either CUDA or OpenCL for nVidia. Regardless of which is chosen, the apps each have a Collatz config file (and in the same format) which you can edit.

As far as the application optimizing itself, that really can't happen because not everyone agrees on the definition of "good video response" or "temperatures not too high". The best I can do is have those with the slowest GPUs report back on the config settings that work on their GPUs which will allow me to adjust the settings in the next version of the default applications.

Perette Barella
Send message
Joined: 21 Oct 13
Posts: 1
Credit: 557,207
RAC: 0
Message 18732 - Posted: 27 Feb 2014, 13:07:48 UTC - in response to Message 18723.

Option A. I'm happy to let Collatz have spare GPU time, but my processor cores are busy on other things.

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2528
Credit: 740,580,099
RAC: 0
Message 18734 - Posted: 27 Feb 2014, 16:02:31 UTC

This thread is now locked because it is no longer relevant. You can choose which apps to run, (e.g. option a, b, or c) via the project preferences web page.

Previous · 1 · 2 · 3 · 4

Message boards : News : CUDA vs nVidia OpenCL User Poll


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.