3.1 tasks should not be sent to 400 series and above cards since it wastes resources since 3.1 tasks do not compute as efficiently on newer cards as 4.2 tasks do.

Likewise 4.2 tasks should not be sent to 200 series and below cards since they usually fail.

There is still merit in 3.1 tasks since a lot of people still have 200 series card which can do a lot of work but perhaps there needs to be a checkoff method for which (3.1 or 4.2) tasks a user will accept.

It's supposed to select the correct cuda for the cards automatically by detecting which video card driver is being used. At the moment, there seems to be a malfunction with the detection process (give it time).

The driver being used has nothing to do with the type of video card. Any nvidia card from 8600 to 690 can use 302 drivers. It also doesnt explain why we are getting both 3.1 and 4.2 tasks on the same machine.

How often are you seeing this happen on your machines?

From a pure BOINC perspective if you have multiple compatible apps, while you can say which app *should* be used, it will occasionally try the other just to make sure the given machine is being sent the best app versino for it. While that might not seem to make a lot of sense if you think about how big the world is and the huge number of variables that the project just can't test for this is actually a pretty good idea.

If the GPUGrid project was to take on any number of measures to always force a particular app version to a particular piece of hardware they would have to do quite a bit of work, all of which would make any future BOINC server upgrades difficult at best, or introduce a substantial amount of admin overhead to support.

I think for the minimal times I have seen this happen on my machines it's not that large of an issue.

---

Thanks - Steve

The thing its not a "particular set of hardware"

Its all nvidia cards 400 series and newer should get only 4.2 tasks.

You should have a FERMI/KEPLER flag. If it detects it...send only 4.2 tasks. If it doesnt send only 3.1 tasks. 4.2 tasks were being sent to my 275 card and routinely failing so I yanked it.

Its all nvidia cards 400 series and newer should get only 4.2 tasks.

It depends also on the driver used for the card. Older drivers don't have CUDA 4.2 support.

You should have a FERMI/KEPLER flag. If it detects it...send only 4.2 tasks. If it doesnt send only 3.1 tasks. 4.2 tasks were being sent to my 275 card and routinely failing so I yanked it.

My workaround for this problem can be used to solve your kind of problem too, if you swap the arguments of the copy command in the 4th step, and overwrite the CUDA 4.2 client with the CUDA 3.1 client. However it requires the host not to have any CUDA 4.2 capable GPU.