DirectX 11 vs. DirectX 12: Test results

Stevan Radovic
Stevan Radovic
Joined: 29 Jul 13
Posts: 9
Credit: 69915434
RAC: 30876
Topic 198511

So, I could not find any info on any advantages that DX12 might have over DX11 in regards to crunching data so I did a small test. Same machine, same drivers, same clocks and obviously the same app. The part above the line is Win 7 Pro SP1 & DX11, the part below the line is Win 10 Pro & DX12:

Average for (20 WUs): Binary Radio Pulsar Search (Parkes PMPS XT) v1.52 (BRP6-cuda32-nv301) windows_intelx86 - 5863.889 sec (CPU: 1158.082 sec)

----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Average for (20 WUs): Binary Radio Pulsar Search (Parkes PMPS XT) v1.52 (BRP6-cuda32-nv301) windows_intelx86 - 5115.885 sec (CPU: 1180.757 sec)

Mumak
Joined: 26 Feb 13
Posts: 325
Credit: 3525471157
RAC: 1500452

DirectX 11 vs. DirectX 12: Test results

DX cannot influence CUDA performance, it's a different interface not utilized for computing.
So the difference must be something else - OS, driver version with improved CUDA, or CPU availability to feed the tasks.

-----

Stevan Radovic
Stevan Radovic
Joined: 29 Jul 13
Posts: 9
Credit: 69915434
RAC: 30876

I did not know that. But, as

I did not know that. But, as I mentioned: same drivers, same machine with same clocks used in the exact same manner regarding CPU. However you want to classify the reason, the difference is there. I'm guessing that in the end its the OS... Or some optimization of the app that was done in the last 5-7 days that we are not aware of...

Bill592
Bill592
Joined: 25 Feb 05
Posts: 786
Credit: 70825065
RAC: 0

RE: Or some optimization of

Quote:
Or some optimization of the app that was done in the last 5-7 days that we are not aware of...

You would get much better performance using the Beta Cuda55
version of the app.

Bill

Stevan Radovic
Stevan Radovic
Joined: 29 Jul 13
Posts: 9
Credit: 69915434
RAC: 30876

Well, that's the version the

Well, that's the version the server sent. I have all of the apps approved for computing. I'll approve the test apps also and see if it'll send it in the next cycle.

archae86
archae86
Joined: 6 Dec 05
Posts: 3157
Credit: 7226494928
RAC: 1072585

RE: approve the test

Quote:
approve the test apps


Yes, that is the setting needed.

Sadly, here on the forums we nearly always say "beta", but the actual web page item people need to activate does not have that word (though it does appear in the label appearing in the application column of the tasks view of boincmgr once you actually get that type of work downloaded).

The CUDA55 benefit over non-beta varies with GPU family. The ones listed for your hosts are all GTX970, so Maxwell2, which we have generally seen to have a very considerable benefit. More work completed per hour. More work completed per watt-hour, and reduced CPU support requirement.

Stevan Radovic
Stevan Radovic
Joined: 29 Jul 13
Posts: 9
Credit: 69915434
RAC: 30876

Excellent. I'm currently

Excellent. I'm currently 'banned' from getting new tasks, I'm assuming its because I aborted a bunch of them because I was setting up the factor for the gpu (to 0.33) in order to fill up the usage of the GPU. Should be up and running when the server sends new tasks.

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5872
Credit: 117703935758
RAC: 35071364

RE: ... I'm assuming its

Quote:
... I'm assuming its because I aborted a bunch of them because I was setting up the factor for the gpu (to 0.33) ....


You do not need to abort anything in order to change the GPU utilization factor. You just need to download one new task and any existing tasks in your work cache will be crunched using the new factor.

The list of tasks for your host with the ID=12195870, shows a total of 856 tasks downloaded with 768 of those aborted and zero currently in progress. A lot of server resources would have been used to allocate all those tasks to your host, not to mention the wasted download bandwidth for all the large data files for the many different frequency ranges for the new gravitational wave run (O1AS20-100F). You must have huge work cache (days of work) preference settings to get so much allocated so quickly.

While you're in the 'sin bin' you should take the time to learn how to properly choose your preference settings. When starting, keep your work cache small so that if you change your mind about something or if something goes wrong you don't need to trash so much. You can choose which particular science runs (either for GPU or for CPU) you wish to support and exclude the others. If you don't understand something, please ask. Responses are usually quite fast.

E@H is a popular project and it's server resources are already under high pressure. You will notice this if you observe how long it can take at times to complete server requests. Please don't add to this unnecessarily.

Cheers,
Gary.

Stevan Radovic
Stevan Radovic
Joined: 29 Jul 13
Posts: 9
Credit: 69915434
RAC: 30876

Ok, understood. Apart from

Ok, understood. Apart from the Arecibo GPU, Parkes PMPS XT and their beta versions are there any other GPU apps at the moment?

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5872
Credit: 117703935758
RAC: 35071364

No.

No.

Cheers,
Gary.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.