Core 18 and Maxwell

[Breach]

Good point. I guess the problem is that the Maxwell drivers aren't performing like they're supposed to, and when it's worked out then the points will go up.

Are other cards getting better numbers with core 18?

770/780/780 Ti-series? 680 is also a bit better. Dunno for AMD.

[VijayPande]

We are continuing to look into core18 PPD issues. The push or Core18 comes in part due to some critical issues in Core17 science-wise that are addressed in core18. If anyone is curious, the discussion of this is completely public at the OpenMM git hub page.

[Jesse_V]

We are continuing to look into core18 PPD issues. The push or Core18 comes in part due to some critical issues in Core17 science-wise that are addressed in core18. If anyone is curious, the discussion of this is completely public at the OpenMM git hub page.

https://github.com/pandegroup/openmm/issues/663 perhaps?

[VijayPande]

Yes, unfortunately, the issue here is in the Maxwell driver, which we don't have control over. We did a workaround to make Core18 work at all on Maxwell, but it does come at a performance hit. We're reported this to NVIDIA to correct this in a future driver release.

[Kjetil]

This is the second time you run core 18, first time slow, second time slow. And is this beta or not?

[bollix47]

AFAIK Core 18 is no longer restricted to beta but is on FAH (probably with the advanced setting) for Windows+Nvidia only and the projects are listed on the Project Summary with ZETA_DEV under the Code Column.

Also, see this announcement: Core18 Projects 10474-10476 to FAH

There is another much later post here that suggests the need for using the client-type advanced option.

[Kjetil]

Okay i now, but i am running 2 10471 and 73? And no info http://fah-web.stanford.edu/cgi-bin/fah ... ed?p=10471

[Joe_H]

As best as I follow the changes in assignment of Core_18 over the last month or so, it should be on Full FAH for Kepler and Fermi cards, and Advanced for Maxwell based GPU's. At one point Maxwell's not set to Beta or Advanced were getting Core_15 assignments, not sure if that is still the case.

[Joe_H]

Okay i now, but i am running 2 10471 and 73? And no info http://fah-web.stanford.edu/cgi-bin/fah ... ed?p=10471

The summary page may not have been properly setup, but information for the range from 10470-10476 is all available on the page for 10470.

[runpaint]

I'm running all Maxwells, no flags, all Windows, and I've been getting a lot of core15 and core18.

[JimF]

I'm running all Maxwells, no flags, all Windows, and I've been getting a lot of core15 and core18.

Same for me, for about a week now. I am actually a bit disappointed when I don't get an occassional Core_18, since the science is more advanced, but I will be still happier when the bug is fixed.
(I just hope that Nvidia does not reply that it isn't a bug but a feature.)

[runpaint]

lol, it's always a feature. If we're lucky, they'll add a new feature where we have to log in and maintain constant connection with the Nvidia server or the card won't work.

[Breach]

If this is in their OpenCL driver code we may have to wait a while...

[VijayPande]

We reported this to NVIDIA in March of 2014 BTW. We've pinged them again recently and "they're working on it".

[VijayPande]

Also, if people contact NVIDIA regarding this bug, that might get their attention.