WARNING Do not upgrade to 375/376.xx drivers (for xx<48)

[snapshot]

I'm just hoping that Nvidia doesn't use the excuse that F@H's version of OpenCL is too old to support. The current installer is getting on for three years old and we know there are later versions of Gromacs available. It would be easy for Nvidia to blame Stanford for not keeping F@H up-to-date. I know it wouldn't be productive and the drivers have other problems too but....

Apologies to the mods if this is in the wrong place but it's something I think we need to consider.

[SombraGuerrero]

I'm pretty certain we've all been thinking about that in one form or another. Ultimately, anything that would be attempted on FAH's side to work around this would spur a conversation about whether rebuilding the whole infrastructure is worth not waiting around for NV. Just as 7.4.4 is beginning to age, there will also come a point where driver set 373.06 is no longer viable, at which point one or the other side will need to budge, lest we end up in an eternal stalemate.

[Joe_H]

I'm just hoping that Nvidia doesn't use the excuse that F@H's version of OpenCL is too old to support. The current installer is getting on for three years old and we know there are later versions of Gromacs available. It would be easy for Nvidia to blame Stanford for not keeping F@H up-to-date. I know it wouldn't be productive and the drivers have other problems too but....

Apologies to the mods if this is in the wrong place but it's something I think we need to consider.

Several misconceptions here. The GPU cores are based on OpenMM instead of Gromacs, and are not installed as part of the client. The current release of Core_21 was in beta test at this point last year. It is based on recent code, not 3 year old code.

Gromacs is used as the basis for the CPU folding cores, currently that is the older Core_A4 which has been out for several years, and Core_A7 based on a newer version of Gromacs which is in open testing.

Though it has no direct relation to this problem, currently there is a new version of the client out for public beta testing - 7.4.15. But the version in full release 7.4.4 is capable of downloading new cores and using them for WU processing.

[toTOW]

The funniest fact is that old OpenMM code is working (core 18) ... so NV couldn't blame OpenMM from being too old

[R.Giskard]

Hi people,

I just started using a new PC with GeForce 1060 3GB and the F@H GPU core fails to start with the log recording the errors mentioned above. I've checked my drivers but I can't find the version number that others are referring to. Most of the drivers have a version number XX.XX.13.7570. Is this the same as the 375.xx mentioned in the topic title?

[SombraGuerrero]

Strange, because those look like the Microsoft releases and for my 1070, the Microsoft drivers have been staying in the September/October range, which have been the ones that have been working. In any case though, the working driver set can be downloaded from Nvidia.
http://www.geforce.com/drivers/results/108326

Just FYI, Microsoft identifies this release as 21.21.13.7306 per my device manager.

[SombraGuerrero]

Also, I have edited my post on the Dev Nvidia forums to accurately reflect the information shared by Joe_H and toTOW. Not surprisingly, there has been no activity on my forum post since I originally wrote it...
I have now also filled out a very detailed report on the feedback form. It must have been down the other week when I tried it, or I was using an old link to get there.....Used the one found in the blog post.

[snapshot]

Several misconceptions here. The GPU cores are based on OpenMM instead of Gromacs, and are not installed as part of the client. The current release of Core_21 was in beta test at this point last year. It is based on recent code, not 3 year old code.

Gromacs is used as the basis for the CPU folding cores, currently that is the older Core_A4 which has been out for several years, and Core_A7 based on a newer version of Gromacs which is in open testing.

Though it has no direct relation to this problem, currently there is a new version of the client out for public beta testing - 7.4.15. But the version in full release 7.4.4 is capable of downloading new cores and using them for WU processing.

Thanks for the corrections. I'm delighted to be wrong. I'll be even more delighted when I hear the 'pop' of Nvidia removing its corporate head from wherever it's currently wedged.

[snapshot]

Hi people,

I just started using a new PC with GeForce 1060 3GB and the F@H GPU core fails to start with the log recording the errors mentioned above. I've checked my drivers but I can't find the version number that others are referring to. Most of the drivers have a version number XX.XX.13.7570. Is this the same as the 375.xx mentioned in the topic title?

Yes, it's 375.70. Don't ask me why the format changes as I don't know.....

[toTOW]

XX.XX.13.7570

Put the dot at the right place and you have your answer

[DutchForce]

NVidia released new drivers 376.19 WHQL today, but still no fix for FaH.
See NVidia forum post of Manuel Guzman:
Official 376.19 Game Ready WHQL Display Driver Feedback Thread (Released 12/5/16)

MammothJerk said:
Any update on the F@H WU problems?
----------------------

ManuelGuzman:
No fix in this driver for that issue. It is one of our top priority bugs though.

EDIT:
See also this post of Manuel Guzman:
Official 376.19 Game Ready WHQL Display Driver Feedback Thread (Released 12/5/16)

bb30994 said:
@Manuel: Do you have enough information to reproduce the F@H problems?

I do not see it listed as an Open Issue.
----------------------

ManuelGuzman:
Yes. We just have not yet narrowed down the root cause.

[yalexey]

Charmingly. They have favourite bugs. I do not think they will be easy to part with it.

[bruce]

Though I don't in any way agree with their emphais, you have to admit that they called this a "Game Ready" release and they have their hands full making it Game Ready.

Fixing regressions in OpenCL support probably doesn't come high on the faults found in the popular games.

Their documentation does say:

Support for OpenCL 1.2

Kepler, Maxwell, and Pascal GPUs are supported.

so they'll get around to fixing it eventually.

[SombraGuerrero]

That's the other thing is that ever since Nvidia started these "Game Ready" releases, my impression has been that they've adopted a paradigm more akin to extreme programming and/or continuous integration. I don't believe these releases are going through full regression testing because that is not the focus of the "game ready" branches.

[bruce]

Right, but if the only driver that will support a new GPU happens to be on a "game ready" branch, what's a person to do?