Can't get two (different) GPUs to work

[bruce]

I have an AMD GPU and a Fermi GPU folding successfully in a V7 setup. Saying that it categorically does not work in V7 is simply not true. (I have not tried a G8x/G9x GPU that gets GPU2 assignments for GPU2 and runs FahCore_11 so I can't vouch for that configuration, but Core_15 and Core_16 seem to work very nicely together and I have great hopes for Core_17 whenever it's mature enough to be released.)

First: drivers for both GPUs must be installed so that the OS recognizes both GPUs. (In my case, the screen-configuration menu looks just like the setup on my laptop which gives me a choice of one monitor or the other, even though there's only one monitor plugged in) ... so the first question is can you make that much work successfully with a monitor plugged into each GPU or does something else happen? Obviously so far this has nothing to do with V6 or V7 yet because you can test that much without installing FAH.

Second: what does V7 reports in the info section. Does it say you have two GPUs and identify them correctly. If not, that's potentially a whitelist problem but there might be other possibilities at this point.

Third: V7 will create two GPU slots and assign default values for slot number, gpu-index, opencl-index and cuda-index. Sometimes it gets them right and sometimes it does not. If that's where you run into trouble, there's a procedure somewhere that gives suggestions for possible changes to the index values but I'm pretty confident that if you get this far, it can be made to work.

[art_l_j_PlanetAMD64]

I have an AMD GPU and a Fermi GPU folding successfully in a V7 setup. Saying that it categorically does not work in V7 is simply not true. (I have not tried a G8x/G9x GPU that gets GPU2 assignments for GPU2 and runs FahCore_11 so I can't vouch for that configuration, but Core_15 and Core_16 seem to work very nicely together and I have great hopes for Core_17 whenever it's mature enough to be released.)

Once again, you're putting words in my mouth that I did not say. I referred specifically to two NVidia cards, using Core 0x11 for the "older" card, and Core 0x15 for the "newer" card. I most emphatically did not "extend" that statement to any other configuration.

First: drivers for both GPUs must be installed so that the OS recognizes both GPUs. (In my case, the screen-configuration menu looks just like the setup on my laptop which gives me a choice of one monitor or the other, even though there's only one monitor plugged in) ... so the first question is can you make that much work successfully with a monitor plugged into each GPU or does something else happen? Obviously so far this has nothing to do with V6 or V7 yet because you can test that much without installing FAH.

There is only one driver for the NVidia cards, which applies to both the "older" card, and the "newer" card. In my case, that is the v306.97 driver. The OS definitely recognized both cards individually, as did the NVidia Control Panel (which you can access by right-clicking on an empty area of the desktop). FAHClient and FAHControl both recognized each card individually as well, and both cards started folding independently of each other, in separate processes. It was only about 2 hours later, that the GPU_MEMTEST_ERROR occurred.

Second: what does V7 reports in the info section. Does it say you have two GPUs and identify them correctly. If not, that's potentially a whitelist problem but there might be other possibilities at this point.

Yes, each GPU was correctly identified by V7, and they each started folding independently of the other in separate processes. It was only about 2 hours later, that the GPU_MEMTEST_ERROR occurred.

Third: V7 will create two GPU slots and assign default values for slot number, gpu-index, opencl-index and cuda-index. Sometimes it gets them right and sometimes it does not. If that's where you run into trouble, there's a procedure somewhere that gives suggestions for possible changes to the index values but I'm pretty confident that if you get this far, it can be made to work.

No, this is not what happened. Please re-read my post, as it is clearly described how the pre-existing slots were separated by 'cloning' the FAHClient directory, and 'removing' the "other" GPU from each directory's config.xml file. There was no problem with the slot number, gpu-index, opencl-index and cuda-index for either card. Both cards started folding correctly, independently of each other, in separate processes. It was only about 2 hours later, that the GPU_MEMTEST_ERROR occurred.

[bruce]

Please don't be so hostile. We're talking about two different things.

Third: V7 will create two GPU slots and assign default values for slot number, gpu-index, opencl-index and cuda-index. Sometimes it gets them right and sometimes it does not. If that's where you run into trouble, there's a procedure somewhere that gives suggestions for possible changes to the index values but I'm pretty confident that if you get this far, it can be made to work.

No, this is not what happened. Please re-read my post, as it is clearly described how the pre-existing slots were separated by 'cloning' the FAHClient directory, and 'removing' the "other" GPU from each directory's config.xml file. There was no problem with the slot number, gpu-index, opencl-index and cuda-index for either card. Both cards started folding correctly, independently of each other, in separate processes. It was only about 2 hours later, that the GPU_MEMTEST_ERROR occurred.

Excuse me ... it did happen exactly as I said: V7 DID create two GPU slots. You're the one who decided you needed to override V7's default behaviour by cloning the client and creating two separate installations and then deleting the duplicate slots. At one time I successfully configured a client that ran both a Fermi and a Keppler GPU. Two slots were created and they both worked. They used the same drivers but the drivers were smart enough to recognize that the devices had different capabilities.

The objective of my post was to explain how my V7 installation works successfully with two (different) GPUs -- answering the original question, not to challenge the configuration that you created. As far as my installation went, there was no need for me to configure V7 in a different way than it was designed to work. Your installation is different than mine in a number of ways.

If there's a bug in the client that can be fixed, you have done nothing to document it. If there's a bug in the drivers, we don't know what that is. You've done what appears to be an effective workaround for an unknown problem, but we need to put some effort into both improving the client by identifying the problems AND into figuring out how to work around existing problems.

[bruce]

By the way, Fermi (GTX 400+) and later are supported by FahCore_15 and pre-Fermi are supported by FahCore_11. If V7 mixes them up by assigning the wrong cuda-indexes, WUs of the wrong architecture will be assigned. Memory errors are a common symptom of that problem. You could test that possibility by simply interchanging the values of cuda-index and dumping the WUs in both slots. It would be interesting to know if that might have fixed the original problem without having to clone the client. Did anybody try that?

At some point, GPU2 (FahCore_11) will be deprecated as has previously been announced, eliminating this particular combination.

[PantherX]

FYI, there are two different versions of FahCore_15. One version woks for Femi and higher and the other version works for Pre-Fermi GPUs. However, it has been a very long time since I have seen a FahCore_15 WU on my Pre-Fermi GPU so am unsure what will happen in the future.

[NathanH]

one last quick question: is the G80 / 8xxx series gpu going to be completely unsupported any time soon? i was tossing around the idea of getting a cheap sli board to dedicate to folding 100% with the 3 8xxx cards. bad idea?

[PantherX]

Only the current projects will work. No new projects will be released. There plug hasn't been pulled yet officially but it has been stated that Pre-Fermi GPUs don't have the required hardware so will not be supported by the New FahCore_17. Whenever the WUs run out (or when PG stops assigning them), that's when these GPUs will be discontinued.

If it's really cheap (cheaper than a Fermi GPU), then it might worth a shot if you can afford the additional electricity. If you can't, might be better off saving for an upgrade.