Why a fallback mode.

[efmer (fred)]

Why do the GPU application still have a fallback mode.
In theory this is a nice feature, but the speed difference between CPU and GPU is way to high.
Whenever this happens on my machine, all CPU fallback tasks end up in timeout errors.
The only good thing is that it doesn't trash all work at a higher rate.

[Jason G]

Why do the GPU application still have a fallback mode.
In theory this is a nice feature, but the speed difference between CPU and GPU is way to high.
Whenever this happens on my machine, all CPU fallback tasks end up in timeout errors.
The only good thing is that it doesn't trash all work at a higher rate.

? no CPU fallback in x41g Fred.

[efmer (fred)]

? no CPU fallback in x41g Fred.

Did I say anything?

Another question do the new drivers 285.62 with X41g use up more GPU memory.
I use to run my 590 with <COUNT>0.25, I set this back to 0.5.

[Jason G]

Did I say anything?

Another question do the new drivers 285.62 with X41g use up more GPU memory.
I use to run my 590 with <COUNT>0.25, I set this back to 0.5.

Shouldn't do, there aren't any new allocations, but we have noticed variability with different drivers.  So far as we've been able to figure most of that is to do with OS & CUFFT libraries, as the actual memory the apps explicitly allocate is pretty small ( ~90meg or so, listed in stderr).  Apparently with effort the app can still 'just' run on a 256meg card, though stock won't, and IIRC some drivers won't leave enough VRAM at all.

Jason

[efmer (fred)]

After removing the new driver and installing the old one. The problem was still there.
It took me a while to see that the SLI on the 590 was activated by the new driver.
That was something X41g didn't like at all. So after that 0.25 worked again without some of the tasks dropping out.
The new driver did enable the adaptive power....

[Jason G]

The SLI thing seems to be an issue that comes & goes with various drivers.  I had thought nVidia had left things that Cuda sees through Sli & ignores it, so a bit weird.  I suppose that monster card is relatively new & so they might still be messing around with some basics in the drivers & Cuda runtime. 

Probably Cuda 4.1 release will understand that card a bit better. There is a prototype (not for release) 4.1rc1 build of x41g in the dev area that appears to perform 'interestingly' on the 480 & 560ti here.  If you have time to give that a run, that & newer drivers might be happier on that setup.  nVidia apparently also want dev feedback on the new compiler performance in particular, so any weird issues like that I could pass on. I've already passed on some positive & negative reports, but for 'us' I think 4.1 release will be a worthy migration (when they get to final release stage).

Jason

[MarkJ]

Probably Cuda 4.1 release will understand that card a bit better. There is a prototype (not for release) 4.1rc1 build of x41g in the dev area that appears to perform 'interestingly' on the 480 & 560ti here.  If you have time to give that a run, that & newer drivers might be happier on that setup.  nVidia apparently also want dev feedback on the new compiler performance in particular, so any weird issues like that I could pass on. I've already passed on some positive & negative reports, but for 'us' I think 4.1 release will be a worthy migration (when they get to final release stage).

Jason

I have a 560Ti in one machine. Would you like me to give it a run?

I would probably have to update drivers 1st though. Any "recommended" version to try or just the latest whql? Machine is a Q6600 with 4Gb RAM running Win 7 x64.

[Jason G]

I have a 560Ti in one machine. Would you like me to give it a run?

I would probably have to update drivers 1st though. Any "recommended" version to try or just the latest whql? Machine is a Q6600 with 4Gb RAM running Win 7 x64.

Might as well hold at x41g v3.2 build for now If that runs stably for you. Definitely please report anything that seems wrong.  Fred's queries, and the queries about '-12's are appreciated as reminders & checks of the shifting priorities as some problems are knocked off the list. 
 4.1rc1 is already replaced by Cuda 4.1rc2  since that post .  Some issues, not related to 560ti's but Pre-Fermis,
have been reported to nVidia, and probably once they address those I'll try get as wide a pre-release testing going as possible (and allowed).  Until then the logic in the 3.2 build should serve to prove the foundations. 4.1 release probably won't be long, whether or not they fully addressed all the issues yet.

Jason