Pande Group Member
Joined: Mon Dec 03, 2007 4:20 am
ihaque on Sat Mar 21, 2009 12:06 am
I've given the results from these drivers (or at least what I think are results from these drivers, based on posts in this thread) a look and I find them inconclusive. The new drivers appear to be messing with the timing of various processing steps in the core. While the simulation trajectories are not grossly or obviously incorrect, I haven't done enough detailed analysis to see if they are subtly wrong. Nevertheless, it's not a simple case of "everything got faster", which is enough to make these drivers suspect in my eyes. There's no changelog attached to the leaked README, so I have no way of knowing what's been changed internally in the driver.
There's been speculation in the thread about these drivers overriding some "throttling" going on in the core. There is no "throttling" in the core; as has been explained in other C14/P590x threads, core 14 runs additional debugging and verification tests (relative to core 11) to make sure the results coming back are reliably produced.
Bottom line: I don't fully trust the results coming from these drivers, and I do not encourage their use, because of strange timing inconsistencies relative to previous driver revs. Further, because these are unofficial, uncertified, leaked drivers, it's entirely possible (even likely) that they have not been fully tested and verified for correct operation.
by ihaque on Sat Mar 21, 2009 7:43 pm
Please understand that my conservatism in this matter is borne of necessity. It's not good science to make a change that dramatically affects performance without first validating that the behavior is correct - especially when there are some clues in the returned data that things are not exactly right. The fact that this 185.20 driver seems never to have been released (and a later one from the 185 line - 185.26 - was) suggests to me that NVIDIA realized something was wrong with this one. I'll try to ask our contacts, but the response won't be instantaneous.
I'm doing my best to test this, but good validation takes time. I dropped the earlier note to inform you of my early findings, which is that there are inconsistencies that make me suspicious of what's going on. Thanks to turnoz for offering to help me do some testing of these drivers.
I'm not out to be a points gremlin - if the results produced by these drivers really turn out to be legit, then great! We like faster science too. From our end, though, we just have to be sure that the science is right before worrying whether it's fast