Theoretical GPU2 Client Loss Question

[compdewd]

Hello friends! I am currently running a GPU2 FahCore_11 Client under FAH GPU Tracker and I was wondering what would happen if no one ran these GPU2 cores? Would these work units for these cores end up being sent to the GPU3 core 15 clients until they were dried up? (which I hear was predicted to happen in September of last year? I assume this hasn't happened yet because of the loss of the number of GPU2 clients. It would also be nice to know if that is the reason that GPU2 WUs haven't dried up ). Just curious with a "what if" situation

[Zagen30]

I think you may be misunderstanding what you heard, or you didn't get the clearest explanation. Deprecation of pre-5000 series AMD GPUs was originally predicted to occur no sooner than last September, and over the last few months work for the AMD version of Core 11 has almost completely dried up. Nvidia cards have a separate Core 11, and I have yet to hear official news of its deprecation, nor do I think it's all that close to happening. AFAIK there's no shortage of Nvidia Core 11 work.

I'll answer the rest of your questions anyway, since knowledge is a good thing:

Work units can only be run on the core for which it they were designed. When the client first starts, it contacts the appropriate work server, which sends back a WU (assuming one is available). At that point, the client checks to see if it has the right core, including whether or not the core is a recent enough version to run that WU; if the core is absent or too old, the proper one is downloaded. Both v6 and v7 can handle all the Nvidia cores, so people upgrading to v7 won't change anything regarding what work they can do. Those with AMD GPUs capable of running the OpenCL core (which v6 doens't support) who are migrating to v7 are likely reducing the number of AMD GPUs running AMD Core 11 work, but since there's so little of that work left anyway I don't believe it's an issue.

In the case of AMD GPU2 work, the WUs will run out first, which was what was actually being predicted for last September. That Core 11 (as opposed to the Nvidia Core 11) will probably be removed from the servers once there's absolutely no more need for it, but until then it'll still be available, and it'll remain on people's computers indefinitely. I've never heard of the client deleting a core unless it's replacing it with a newer version.

[bruce]

Continuing on the "knowledge is a good thing" let me add that it is technically possible to convert individual WUs from one FahCore to another one but it's not likely to ever happen. It's a lot of extra (unnecessary) work and can introduce questions about the scientific validity of the entire project since it would no longer be officially reproducible using the stated methods (i.e.- somebody might have to prove that switching cores introduces precisely zero changes to the results or put together a justification that they're "essentially equivalent." Either of those fact are probably true but it's easier not to have to prove it and to write a justification. It's much simpler just to let a project "finish" using the originally stated method and avoid all the extra scientific justification.

[compdewd]

Okay, I confused the old ATI news with NVIDIA too. Thank you for replying! I am unsure if my "what if" question was really answered , but that's okay. A real question that has arisen for me is what is the difference between the ATI and NVIDIA Core 11's?

[bruce]

I'm reasonably certain that the GROMACS code in both versions of Core_11 are identical (hence the name GPU2) but they use different calls to the system routines.

The ATI version makes calls to the Brook interface; the NVidia core makes calls to the CUDA API. The Brook interface is no longer supported by ATI/AMD, having moved all of their support to the OpenCL API. NVidia still fully supports CUDA as well as OpenCL.

I'm also reasonably certain that Core_15/Core_16 contain a different version of GROMACS from the version used by the Core_11's (hence the name GPU3) but they interface with CUDA/OpenCL, respectively.

[compdewd]

Thanks for the info!

[Zagen30]

My long-winded explanation may not have conveyed what I wanted to say properly. Let me try again:

" what would happen if no one ran these GPU2 cores?": That can't happen, since what work you get determines what core is used, and you don't have a choice in the matter. If the work is available on the server, the necessary core will be there too, and the client will get the core if you don't have it on your computer. All of the GPU cores except Core 16 are supported by the v6 GPU client, and every GPU core is supported by v7, so you won't get a situation where the older work is incompatible with the latest client (incompatible with newer cores, yes, but the client version has no bearing on that, since all it really does is handle uploads/downloads and starting/stopping the cores).

[compdewd]

Ahh I understand what you were saying now. The reason I thought v7 wasn't compatible with GPU2 was because when I tried my 8400GS on v7, it wouldn't download the necessary core or any work units but FAH GPU Tracker worked because it used v6(?) and -forcegpu flag which v7 apparently doesn't include. But now I remember the real reason my GPU didn't work on v7 was I believe because GPUs.txt said it was unsupported; perhaps you recall viewtopic.php?f=83&t=21783?