AMD GPU works for GPUGrid

[JugNut]

GPUgrid will have work for AMD GPU's.     Another medical GPU app for AMD users. 

The new app is in testing in the beta queue, so if you want to try it out before it's full release you'll have to edit your preferences at the web site.  See here for more info...  http://www.gpugrid.net/forum_thread.php?id=3995

There has been some early troubles with the app so it's probably best to check the thread regularly for info if your interested.

Good luck

[kashi]

Thanks for the info.

Gave it a go, eventually got it to download a task after fiddling about with project preferences, gives error on my HD 7790 as at version 9.10.

Hopefully they'll get it working soon, medical app other than Folding for AMD GPU would be good.

[Sean]

It's now working on my 5770, 60-70% gpu utilisation    :drool: 

[tazzduke]

Greetings All

Am also running a few workunits on an AMD HD5850 quad core 8400, just a couple of observations

Workunit  Version is up to 9.15

Need to allocate all your cpu cores to the one work unit, so bit of a bummer if you want to crunch other projects on the same machine (I am running a few till morning then back to WCG)

Also only will down load workunits depening on how many cores your machine is, ie 4 cores 4 workunits, 8 cores 8 workunits at a time.

Still in testing, but ironing out the bugs, the credit has been adjusted a couple of times as well.

Please read the forum thread in relation to this for updates.

Regards

[kashi]

It's supposedly a multi threaded application but it doesn't go any faster with more than 2 CPU cores with no other project running. It doesn't even slow down that much using a single CPU core. Of course it slows down a bit with contention losses when you are running a CPU project at the same time as well.  Terribly inefficient waste of resources, lucky it's only a test.

You don't need to run it on all cores if you don't want to. CPU cores can be limited with an app_config.xml file (example below). If you want it to run all the time you may need to use an app_config.xml file with the other project CPU application you are running so that there are enough free CPU cores. For example, I am running Universe on 5 CPU cores (<max_concurrent>5</max_concurrent>) and GPUGrid on 2 CPU cores. 

Credit is all over the place, last 2 tasks I did ran for 30 minutes (7 CPUs) and 36 minutes (2 CPUs) and granted 11.47 and 11.57 credits, yet I have seen some get over 50,000 credits for the same more recent longer tasks. One of the earlier short tasks granted 1,809 credits and ran 5 minutes. I'm happy to help test a BOINC AMD GPU application but if GPUGrid uses CreditNew then I won't be supporting it if/when they ever get it running properly.

<app_config>
<app>
      <name>amdmd</name>
      <max_concurrent>1</max_concurrent>
</app>
    <app_version>
       <app_name>amdmd</app_name>
       <plan_class>mt</plan_class>
       <avg_ncpus>2</avg_ncpus>
       <cmdline>--nthreads 2</cmdline>
   </app_version>
  </app_config>

[JugNut]

Unfortunately I haven't been able to get any running at all.  At the moment they only give work to to people with single AMD GPU's which rules me out.  All my rigs are mixed rigs(AMD & Nvidia in the one PC) & the way it's looks at the moment I wonder if they will ever support mixed rigs like these? The only other PC I have has 2 x Nvidia's in it.  I'm not to fussed about it ATM as there's still plenty of POEM WU for my AMD's to crunch on & besides my Nvidia's are crunching more than their fair share of GPUgrid awayway.

[kashi]

Yes it's good that GPUGrid is starting to develop applications for AMD GPUs, however this test one is not really a true GPU application but a modified multi threaded CPU application with some GPU assist.

In the way it works it's currently similar to the horror GPU application that WCG developed a while ago. Efficiency and performance depends on CPU as much as GPU, making the use of a powerful GPU a waste of resources. It's a shame that they are starting so far behind the eight ball and can't get some help from the developers at Folding@Home. The Folding AMD GPU app has been developed over many years, is very efficient, and uses almost no CPU resources.

At least they're having a go, not very promising so far but hopefully one day they'll get away from this silly MT CPU concept and be able to develop a real AMD GPU application. 
Oh, and one that pays somewhat more than 550 credits per day for the use of a GPU and 7 CPU cores.

Yeh, looks like you're certainly going to town on GPUGrid on your nVidia cards, so no worries there, haha. An excellent contribution.

[Sean]

It's supposedly a multi threaded application but it doesn't go any faster with more than 2 CPU cores with no other project running.

Out of interest did the CPU utilisation % change between using 7 and 2 cores?

[kashi]

Yes CPU % for mdrun-915.exe was about 87% when configured to use 7 CPU cores and 25% with 2 CPU cores configured. This is on an 8 "core" CPU, so each CPU core is 12.5%.

CPU contention slowed down the GPUGrid application a fair bit (25-50%), when other project CPU tasks were run. In other words it behaved like a CPU application rather than a true GPU application in that respect. The WCG Help Conquer Cancer GPU application which also did a lot of processing on the CPU behaved in a similar undesirable inefficient manner.

[Sean]

Cool thanks for the reply, that's kind of odd but interesting.  


Edit: Message 39973 in the main GPUGrid thread that JugNut linked to has a bit about using the Stderr file to set the CPU allocation, I haven't tried it yet though.

[kashi]

Yes that's right, you can refer to Stderr output of tasks to check to see if load is balanced between CPU and GPU. When I was using 3 or more CPU cores it always said "The GPU has >20% more load than the CPU. This imbalance causes performance loss, consider using a shorter cut-off and a finer PME grid." When I was using 1 or 2 CPU cores that message was not there in Stderr output.

I had already confirmed Werkstatt's experience mentioned in Message 39984 that 2 CPU cores was all that was needed after testing the run time with various numbers of CPU cores allocated with no other CPU projects running. Seeing as I am running a relatively low spec AMD HD 7790 I think it's unlikely that more than 2 CPU cores would be needed for "balance".