[solved] nVidia 295.51 beta driver problems - Failed tasks - Cannot create a symbolic link in a registry key that already has subkeys or values. (0x3fc) - exit code 1020 (0x3fc)
You might also consider chiming in on the nVidia Feedback thread, for the 295.73 drivers, located here: http://forums.nvidia.com/index.php?showtopic=223426 ... and providing as much feedback as possible. Once mine start failing, I'll post there too, and hopefully they hear us, as I roll back to 290.53 (the last ones that worked for me).
It's sad that we haven't received responses -- either they are completely swamped with bugs, or they just don't care. Either way, it saddening and frustrating.
I posted a report to that thread on friday, (as well as my Bug report that i've already mentioned) no response as of yet from eithier,
Thanks for pushing the issue with NVIDIA. Given how long it took them to fix a driver they once put out that was shutting down the fans on their GPU's, I'm not overly optimistic as to how long they'll take to fix this problem. Unfortunately, dropping back to an earlier driver version isn't an option for some of us. But again, thanks for pushing the issue.
I thought I'd chime in with some more information.
If you want to use 295.73 WHQL or 295.51 Beta without CUDA failures:
The workaround for the bug is that you must set the Windows Power Options to "Turn off the display: Never". You may still use a screen saver, and you may still physically turn the monitor off, but you must not let the software power the monitor down... according to my testing.
Also, for anyone trying to reproduce the problem, I have found that the problem occurs when Windows powers off the monitor first, and then BOINC tries to start or resume a CUDA task while the monitor is off. This means that, if you try to reproduce it using tasks that are already running before Windows powers down the monitor, those tasks will not fail. But any tasks that try to start or resume, while the monitor is off, will fail... according to my testing.
Finally, the best news yet, I have been contacted privately by an nVidia employee, who was having trouble recreating the problem. I assisted him, and he can now repro on demand now (it's easiest to repro with Einstein@Home, and he didn't know that the monitor has to power down before BOINC begins CUDA processing), and he will be presenting information to the developers.
I am now going to run 295.73 WHQL with a "Blank" screensaver and "Turn off the display: Never", and try to remember to physically turn the monitor off if I get up for an extended period of time.
Finally, the best news yet, I have been contacted privately by an nVidia employee, who was having trouble recreating the problem. I assisted him, and he can now repro on demand now (it's easiest to repro with Einstein@Home, and he didn't know that the monitor has to power down before BOINC begins CUDA processing), and he will be presenting information to the developers.
Can you get me in touch with that guy? I just noticed this problem and it's causing us quite some trouble, since we're really loosing potential computing power (Pulsar discoveries anyone?!). It might very well be that our CUDA 3.2 app simply doesn't get along with the latest 4.1 drivers and an upgrade (on our end) might equally solve it or serve as a workaround until NVIDIA fixed the root cause.
And by the way, sorry that it took us so long to become aware of this issue, Bernd and I are just swamped with other E@H topics like OpenCL (AMD GPUs). Time to look after CUDA again...
I see Nvidia released a new driver version today (296.10) but I didn't see anything in the release notes about the cuda issue. Hopefully now they're aware of what's causing the issue they'll be able to address it in the not too distant future.
I see Nvidia released a new driver version today (296.10) but I didn't see anything in the release notes about the cuda issue. Hopefully now they're aware of what's causing the issue they'll be able to address it in the not too distant future.
I have passed the request to the nVidia employee, to see if I can share his email address to you. I will let you know if/when he responds. I wanted to get his approval first, mainly for 2 reasons -- I don't want to cross him, and also I know they're swamped with other issues too.
Hopefully you can understand, and hopefully he responds positively.
Will let you know,
Jacob
I have sent you a Private Message that has his contact information, along with my prior communications.
Also, I have confirmed that the 296.10 WHQL drivers still do not work correctly unless you use "Turn off the display: Never".
If there's anything you need in terms of testing a potential interim solution, please don't hesitate to ask me.
FYI, we won't ship any CUDA tasks for systems using a driver newer than 290.53 for the time being. This decision was made to avoid unnecessary computation errors, data set downloads, daily quota reduction and volunteer frustration.
We're in contact with NVIDIA and we'll try to help them fix the situation as fast as possible. If there's a workaround without user-action (not like "Turn off the display: Never") we'll implement it. Until that time, please understand that the majority of our volunteers might not follow this forum/thread and we need to take care of this problem without requiring them to work around it themselves.
RE: Thanks for the response
)
I posted a report to that thread on friday, (as well as my Bug report that i've already mentioned) no response as of yet from eithier,
Claggy
Thanks for pushing the issue
)
Thanks for pushing the issue with NVIDIA. Given how long it took them to fix a driver they once put out that was shutting down the fans on their GPU's, I'm not overly optimistic as to how long they'll take to fix this problem. Unfortunately, dropping back to an earlier driver version isn't an option for some of us. But again, thanks for pushing the issue.
I thought I'd chime in with
)
I thought I'd chime in with some more information.
If you want to use 295.73 WHQL or 295.51 Beta without CUDA failures:
The workaround for the bug is that you must set the Windows Power Options to "Turn off the display: Never". You may still use a screen saver, and you may still physically turn the monitor off, but you must not let the software power the monitor down... according to my testing.
Also, for anyone trying to reproduce the problem, I have found that the problem occurs when Windows powers off the monitor first, and then BOINC tries to start or resume a CUDA task while the monitor is off. This means that, if you try to reproduce it using tasks that are already running before Windows powers down the monitor, those tasks will not fail. But any tasks that try to start or resume, while the monitor is off, will fail... according to my testing.
Finally, the best news yet, I have been contacted privately by an nVidia employee, who was having trouble recreating the problem. I assisted him, and he can now repro on demand now (it's easiest to repro with Einstein@Home, and he didn't know that the monitor has to power down before BOINC begins CUDA processing), and he will be presenting information to the developers.
I am now going to run 295.73 WHQL with a "Blank" screensaver and "Turn off the display: Never", and try to remember to physically turn the monitor off if I get up for an extended period of time.
Regards,
Jacob Klein
Hi
)
Hi Jacob,
Can you get me in touch with that guy? I just noticed this problem and it's causing us quite some trouble, since we're really loosing potential computing power (Pulsar discoveries anyone?!). It might very well be that our CUDA 3.2 app simply doesn't get along with the latest 4.1 drivers and an upgrade (on our end) might equally solve it or serve as a workaround until NVIDIA fixed the root cause.
And by the way, sorry that it took us so long to become aware of this issue, Bernd and I are just swamped with other E@H topics like OpenCL (AMD GPUs). Time to look after CUDA again...
Thanks,
Oliver
Einstein@Home Project
I see Nvidia released a new
)
I see Nvidia released a new driver version today (296.10) but I didn't see anything in the release notes about the cuda issue. Hopefully now they're aware of what's causing the issue they'll be able to address it in the not too distant future.
RE: I see Nvidia released a
)
It's been tested elsewhere already. Not fixed.
Oliver, I have passed the
)
Oliver,
I have passed the request to the nVidia employee, to see if I can share his email address to you. I will let you know if/when he responds. I wanted to get his approval first, mainly for 2 reasons -- I don't want to cross him, and also I know they're swamped with other issues too.
Hopefully you can understand, and hopefully he responds positively.
Will let you know,
Jacob
Oliver, I have sent you a
)
Oliver,
I have sent you a Private Message that has his contact information, along with my prior communications.
Also, I have confirmed that the 296.10 WHQL drivers still do not work correctly unless you use "Turn off the display: Never".
If there's anything you need in terms of testing a potential interim solution, please don't hesitate to ask me.
Regards,
Jacob
Thanks Jacob, well
)
Thanks Jacob, well done.
Oliver
Einstein@Home Project
FYI, we won't ship any CUDA
)
FYI, we won't ship any CUDA tasks for systems using a driver newer than 290.53 for the time being. This decision was made to avoid unnecessary computation errors, data set downloads, daily quota reduction and volunteer frustration.
We're in contact with NVIDIA and we'll try to help them fix the situation as fast as possible. If there's a workaround without user-action (not like "Turn off the display: Never") we'll implement it. Until that time, please understand that the majority of our volunteers might not follow this forum/thread and we need to take care of this problem without requiring them to work around it themselves.
We'll keep you posted.
Thanks,
Oliver
Einstein@Home Project