Hi!

I've recently joined. I've been a not so frequent GPU cruncher for another project, but since their servers are quite random distributing work for GPUs i've moved to GPUGRID.

I would like to allocate only the GPU for this project, leaving the CPU free to crunch on CPU specifics projects.

I'm aware it is always needed a small % to feed the GPU. What i would like was no WUs being specifically sent to the CPU.

On other projects there is the option, on the preferences page, to "disable" the CPU. I've been to that preference page here on GPUGRID, but even after unticking the USE CPU box and applying the settings, after refreshing the page it still continues enabled.

Is it project behavior or some malfunction? Can't i permanently disable the CPU?

They don't have any CPU-app here.
But, especially on Linux, the app uses a considerably large amount of CPU (if they restore the old one again, the actual one isn't capable of running on Linux at all).

---

Gruesse vom Saenger

For questions about Boinc look in the BOINC-Wiki

Thanks for your answer!

It is a fact there is no CPU app on the site. However BOINC is the application that i am using and it can use both CPU and GPU.

Since there is an option at the account preferences for CPU and or GPU use my doubt was motivated by that.

GPUGrid does not have a CPU only application. GPUGrid only runs GPU tasks on some high end NVidia cards. The CUDA app does however use some CPU time, but this depends on the GPU you are running the tasks on and the operating system you are using. Until recently Linux used a full CPU core to facilitate the GPU (the cards that work here can have anywhere from 48 to 512 shaders/cuda cores). Different cuda applications are being tested at the minute, and some recent tasks used less CPU on Linux suffered badly. So more testing is in progress.

Well as it seems i won't be also able to help this project. At least in a relatively normal way. I've got a 260GTX (65nm) and after several WU's crashing the video driver, i've been searching around and found that apparently these cards suffer from some kind of CUDA/application bug that may render many WU's aborted due to computation error...

Information is quite scarce and i'm still trying to find more details. However if anyone is enlightened enough to shed some light...?

I suppose you have already updated your nvidia drivers...
The most recent drivers from nvidia seem to be causing some problems...
you might want to consider reverting to some previous version

---

If I un-check the "Use Central Processing Unit (CPU)" box in the project settings, what it will result in?
And what does this setting mean "Maximum CPU % for graphics"?
I understand GPUgrid is not supposed to involve any CPU usage when its WU is running? I really would not like that because I want all my CPUs working with other purely CPU projects...
Then, what does it mean, when a WU is running there is this comment next to its name "(NVidia + 0,7 CPU)?
I have a GTX680 and Windows 7-64
Please advise how to minimize the CPU usage.
Thank you.

If I un-check the "Use Central Processing Unit (CPU)" box in the project settings, what it will result in?

Nothing will change.

And what does this setting mean "Maximum CPU % for graphics"?

This setting is for those CPU projects, which can show their progress in some form of graphics (for example a screen saver).
It has nothing to do with the GPUGrid client.

I understand GPUgrid is not supposed to involve any CPU usage when its WU is running?

GPUGrid does involve more or less CPU usage (depending on the workunit), since it is doing only the single precision calculations on the GPU, and therefore the double precision calculations are running on the CPU (since the reduced DP capabilities of the consumer GPUs).

I really would not like that because I want all my CPUs working with other purely CPU projects...

The GPU is still "only" a coprocessor at the moment. It can do parallel calculations at incredible speed, but since it's not an independent device, it needs a CPU which feeds the GPU with fresh work, and stores the results on the hard drive through an OS. So every GPU task needs some CPU time, even those which need less DP calculatins.

Then, what does it mean, when a WU is running there is this comment next to its name "(NVidia + 0,7 CPU)?

This is a quite misleading information, based on the speed ratio of the CPU and the GPU measured by the BOINC manager.

I have a GTX680 and Windows 7-64
Please advise how to minimize the CPU usage.

The GTX680 is a very fast GPU with many cores. For maximizing it's performance, the GPUGrid client always will use a full CPU thread to feed that monster. There are some workunits, which doesn't use a full CPU thread by mistake, and their performance is very poor (doubled running time). Believe me, you wouldn't like that.
BTW, Your CPU has 4 hyperhtreaded cores, so it can handle 8 threads at the same time, but it still has only 4 FPUs (Floating Point Units, this unit does the calculations). So if you run 8 CPU tasks on it at the same time, those can saturate the 4 FPU, and you probably won't earn more credits per day than if you run only 7 (6,5 or 4) tasks on your CPU at the same time (depending on the project, and the actual workunits).
After all, it's quite worth to "sacrifice" a thread on a hyperthreaded quadcore CPU to feed a 1536-core GPU.

Thank you for the detailed reply.
My situation was that I had 2 projects running on my Boinc client: GPUgrid and Rosetta. At some point, Rosetta WUs started to fail, they finished working all right, went off the the server and got "Client Error" there. So my daily quota dropped to minimal. I suspected the reason could be the 2 projects ran simultaneously and GPUgrid somehow interfered and spoilt Rosetta WUs. However, after some experimenting and discussing the problem on my team forum, I got convinced it was not the case. I am glad I can continue supporting GPUgrid.

Thank you for the detailed reply.
My situation was that I had 2 projects running on my Boinc client: GPUgrid and Rosetta. At some point, Rosetta WUs started to fail, they finished working all right, went off the the server and got "Client Error" there. So my daily quota dropped to minimal. I suspected the reason could be the 2 projects ran simultaneously and GPUgrid somehow interfered and spoilt Rosetta WUs. However, after some experimenting and discussing the problem on my team forum, I got convinced it was not the case. I am glad I can continue supporting GPUgrid.

You and your team were correct...Rosetta is having some MAJOR error problems and have been for awhile now. Their problem seems to be that it works for SOME people but not ALL people so they are too interested in figuring out what the problem really is. They seem to have enough people crunching that it does work for that they are fairly happy. The one person that MAY work on it says he is REALLY busy with other stuff and maybe in "a couple of weeks" he will look into it!

I left Rostta due to this. It isn't a GPUGRID bug. I couldn't even run Rosetta exclusively!

BOINC makes it difficult to download enough CPU workunits if you have a GPU but no GPU workunits - it insists on getting a GPU workunit before it will even try to get CPU workunits, even if you've told it not to run anything on the GPU.

Rosetta@Home tried recompiling their application to run on a GPU, and found that it is so serial in nature that it often runs slower on a GPU than on a CPU. Therefore, they gave up on offering any GPU workunits any time soon.

One way to get close to running on the GPU only is to subscribe only to BOINC projects that offer only GPU workunits.

BOINC makes it difficult to download enough CPU workunits if you have a GPU but no GPU workunits - it insists on getting a GPU workunit before it will even try to get CPU workunits, even if you've told it not to run anything on the GPU.

Rosetta@Home tried recompiling their application to run on a GPU, and found that it is so serial in nature that it often runs slower on a GPU than on a CPU. Therefore, they gave up on offering any GPU workunits any time soon.

One way to get close to running on the GPU only is to subscribe only to BOINC projects that offer only GPU workunits.

Also MOST projects that have both offer a selection on the webpage to get cpu AND gpu units OR one or the other.

Also Alexey I crunch both cpu units AND gpu units on most of my pc's, I just NEVER do it for the SAME project on the SAME pc! Right now I am conecntrating on GpuGrid for my Nvidia gpu's and Moo and Collatz for my AMD gpu's. MilkyWay and DistRTgen are not actively beign crunched right now and their RAC is declining.I am concentrating on Cosmology and a little bit on Eon and Malaria for my cpu's. I have also stopped at WCG for now as I have all the badges at the max of 2 years already, LHC is also declining as they have a problem keeping the workunits available so I have no cpu's there right now either. All of my pc's seem to be getting work just fine as long as the projects have work available. I will let my signature come thru this time so you can see it visually.

---