Nvidia GPU: OpenCL vs CUDA55
log in

Advanced search

Message boards : Number crunching : Nvidia GPU: OpenCL vs CUDA55

Author Message
SuperSluether
Send message
Joined: 17 Mar 14
Posts: 13
Credit: 43,691,916
RAC: 5
Message 19991 - Posted: 27 Oct 2014, 0:58:15 UTC
Last modified: 27 Oct 2014, 0:58:51 UTC

I'm receiving a mixture of OpenCL and CUDA55 tasks for my Nvidia GPU. I understand the difference between them, but why is there a separate OpenCL app for Nvidia when it can use CUDA55? Is this for compatability with older cards?

Edit: Does CUDA55 offer any advantages over OpenCL for this project?
____________

Profile FalconFly
Avatar
Send message
Joined: 25 Oct 09
Posts: 12
Credit: 207,961,802
RAC: 0
Message 19994 - Posted: 27 Oct 2014, 11:30:13 UTC - in response to Message 19991.
Last modified: 27 Oct 2014, 11:48:14 UTC

From what I've seen myself, CUDA should be preferred over OpenCL on NVidia cards any day (lower CPU load and likely better performance, resulting in an overall better energy efficiency).

I think the server (over time) will sort this out by itself and will mostly send your system Cuda tasks. But the quickest way is to manually disable NVidia OpenCL tasks in your preferences (unless you see any benefit from running the OpenCL application).

A much more important task would be to use Optimization switches, which will cause a drastically increased performance on your GTX760.

The configuration Files are normally located here :
C:\ProgramData\BOINC\projects\boinc.thesonntags.com_collatz\

mini_collatz_windows_x86_64_cuda.config
solo_collatz_windows_x86_64_cuda.config
large_collatz_windows_x86_64_cuda.config

By default, those config files are empty and Collatz will use very conservative default values, resulting in low GPU utilization and very long runtimes.

This should be just fine (near max. optimum) for your GTX760 :

verbose=1
items_per_kernel=21
kernels_per_reduction=9
threads=10
sleep=1


Place this into the existing empty config files and wait for a fresh Cuda workunit to begin at much higher speed.
(note that this will increase load on your GPU quite a bit, resulting in far higher temperatures and likely noise from the GPU fan - it's a good idea to monitor temps to see if your overall cooling allows for such high workloads over longer periods of time)

Since mini workunits will complete far below the credit bonus cutoff-line of 30 Minutes (i.e. any workunit running <30m gets about a 15% credit penalty), your card is best suited running only Solo & Large.
If your Solo Workunits finish <30 Minutes, it's best to run only Large WorkUnits for optimum crediting.

Mini/Solo/Large can be selected from your Project preferences as needed.
____________

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 20008 - Posted: 31 Oct 2014, 4:00:45 UTC

You can also disable opencl for nvidia via "run only the selected plan classes" item on the Collatz Conjecture preferences page. It can be accessed via your account page.

I've yet to find an nVidia GPU that didn't run better with the CUDA app than the OpenCL app, but since the code is the same as the OpenCL AMD, Intel, and OS X apps, I make it available for nVidia as well. On laptops, if one doesn't tweak the settings, the OpenCL runs cooler (and slower) than the CUDA app.


Post to thread

Message boards : Number crunching : Nvidia GPU: OpenCL vs CUDA55


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.