App 9.18 seems to take way longer to get the same WUs done than 9.15, is there an explanation behind this? Upwards of 20-30% longer

My WUs are running for around 27h on my GTX 660 Ti cards. I cannot use my 650 Ti cards as the run time is much higher. I guess my days at GPUGrid may be drawing to a close as these cards can't keep up.

---

John

If I compare the crunching times of 918-80 with what they were before with 847-65 / 848-65, I, too, notice a considerable increase.
This is true for my GTX750ti on Windows10 and the GTX970 on Windows 10 - they now need about 30% longer :-(

Since at the same time I had to update the drivers to 381.65, it's hard to say whether the raise in crunching time is mainly due to the new acemd version, or mainly due to the new driver, or both.
What could also be the case is that, for what reason ever, the WDDM overhead now has a much more negative effect.

Hence, the main question seems to be whether anything can be done in order to get the former speed back.

What concerns the change from 847-65 / 848-65 to 849-65 for Windows XP, no WDDM overhead, (one GTX750ti and two GTX980ti), the crunching times seem to be only slightly higher than before.

Would be interesting to hear how the situation is with Linux.

Would be interesting to hear how the situation is with Linux.

Linux has stayed on the 9.14 application for the past few months, at least for my GTX 960/970 and I think the Pascal cards too. It was originally the usual 15% or so speed improvement over Windows, though that may have now increased.

But I don't see the page listing the applications at the moment. Either I am not looking in the right place, or maybe they are updating it.

But I don't see the page listing the applications at the moment. Either I am not looking in the right place, or maybe they are updating it.

Jim, are you talking about this:
http://gpugrid.net/forum_thread.php?id=4551&nowrap=true#46981

Didn't they have a separate page listing all the apps for Window, Linux, etc, along with the dates? I thought so, and I saw that the Linux 9.14 version came out last year. Or else I am just imagining it; maybe I got it right? Who knows.

With the lack of Pascal Windows XP support, and now with the dramatic slowdown of WDDM modern windows OS's. It looks like we must change our crunching rigs to linux as it is still running the 9.14 app, with no WDDM and no 9.18 slowdown.

... It looks like we must change our crunching rigs to linux as it is still running the 9.14 app, with no WDDM and no 9.18 slowdown.

but who can tell how long Linux will be running the 9.14 app - before changing to 9.18 or any newer (and also slower) app maybe in the near future?

but who can tell how long Linux will be running the 9.14 app - before changing to 9.18 or any newer (and also slower) app maybe in the near future?

Clearly it's a bug of 9.18 so they definitely would not hinder themselves any further until 9.19 comes out hopefully fixing the problem. Scientists are some of the most impatient people, so I'm sure they want this fixed faster than any of us

But I don't see the page listing the applications at the moment. Either I am not looking in the right place, or maybe they are updating it.

You can find them here. This page is not linked from the GPUGrid pages, so it's no wonder that you didn't see it.

You can find them here.

although this information is somewhat misleading what concerns 64bit Windows. Because if the 64bit Windows is a Windows XP, it runs with 8.49 (cuda65)

Microsoft Windows (98 or later) running on an Intel x86-compatible CPU 8.49 (cuda65) 17 Apr 2017 | 19:06:31 UTC
Microsoft Windows (98 or later) running on an Intel x86-compatible CPU 9.18 (cuda80) 16 Apr 2017 | 0:40:05 UTC
Linux running on an AMD x86_64 or Intel EM64T CPU 9.14 (cuda80) 1 Nov 2016 | 21:27:32 UTC
64bit Windows 9.18 (cuda80)

If I compare the crunching times of 918-80 with what they were before with 847-65 / 848-65, I, too, notice a considerable increase.
This is true for my GTX750ti on Windows10 and the GTX970 on Windows 10 - they now need about 30% longer :-(

Since at the same time I had to update the drivers to 381.65, it's hard to say whether the raise in crunching time is mainly due to the new acemd version, or mainly due to the new driver, or both.
What could also be the case is that, for what reason ever, the WDDM overhead now has a much more negative effect.

Hence, the main question seems to be whether anything can be done in order to get the former speed back.

What concerns the change from 847-65 / 848-65 to 849-65 for Windows XP, no WDDM overhead, (one GTX750ti and two GTX980ti), the crunching times seem to be only slightly higher than before.

Would be interesting to hear how the situation is with Linux.

In view of the above-said, plus my experience from the last three weeks, I strongly suggest that in the preferences for download of tasks, beside "short runs" and "long runs", a third category like "extra long runs" or similar, should be introduced.

A typical candidate for such an "extra long runs" would be "ADRIA_FOLDGREED90_crystal_ss_contacts_100_ubiquitin..."
which on my GTX750Ti in the Windows10 machine takes close to 60 hours (on Windows XP with software 849-65 and an older driver, crunching time is considerably shorter).

Beside the extremely, if not rediculously long crunching time, crunching these sort of tasks with the GTX750Ti on Windows10 is rather unstable. Even the slightest overclocking can lead to immediate halts - the only thing one then can do (as soon as one finds out, which might take a while - for example over night) is to suspend and restart the task. So it's been the case here that the total time for such a task was 3 days! Forget about it! A GTX750ti on Windows10 is definitely overtaxed with such a task.

Hence, crunchers with smaller/older cards would definitely wish to exclude tasks like the "_100_ubiquitin..." from download.
Be it coinsidence or not: what I have watched all time long was that my two GTX750Ti tend to receive such long tasks almost all the time, whereas my GTX980Ti's get smaller tasks (like the "Adria") as well. No idea why, it just seems worth to be mentioned (maybe someone knows more about this phenomen).

Anyway, at the bottom line: while it's clear that we crunchers should not be allowed to choose a specific type of task (this clearly remains in the priority determination of the GPUGRID people), a third category "extra long runs" would help quite a lot.
As everyone can imagine: total running time of nearly 3 days is rather frustrating, besides that it does not make a whole lot of sense :-(

I think you have it backwards. The "long" category is for the latest cards, which the GTX 750 Ti is not. It is wonderful for efficiency (I have several), but it is not a powerhouse. So use it on "shorts", where it can do its thing.

I know, there are not enough shorts to keep it busy. That is a consequence of the science they need to do. Rule #1 for me is that we are here to help them. They are not here to provide us something to do with our cards.

PS: An extra category for "extra longs" is a good idea, but it won't save the GTX 750 Ti.

PS: An extra category for "extra longs" is a good idea, but it won't save the GTX 750 Ti.

But it would help to assign to the GTX750ti those "longruns" which are not that much of a problem for it :-) (and there are still many of such task around)

PS: An extra category for "extra longs" is a good idea, but it won't save the GTX 750 Ti.

But it would help to assign to the GTX750ti those "longruns" which are not that much of a problem for it :-) (and there are still many of such task around)

I guess you are saying that they could then exclude those extra long runs. So it could be a good idea with that in mind.

PS - I just pulled a GTX 750 Ti out and replaced it with a GTX 1060 on my Win7 64-bit machine. Too bad the app is not optimized for it now, but I will try it later. And anything to extend the life of the GTX 750 Ti is a good idea if you don't insist on the 50% bonus, which is not that important anyway.

I guess you are saying that they could then exclude those extra long runs. So it could be a good idea with that in mind.

yes, this is exactly what I am saying :-) Would make a lot of sense.

Of course, like in many other areas, things are progressing. So at GPUGRID, too.
When some 15 months ago, I started adding two GTX980ti to my crunching hardware (which had been very modest until then), no "longruns" took longer than 6-7 hours.
And with the GTX750ti, I could finish almost any "longruns" within the 24-hours-bonus-period.

But things have changed, obviously.
Nowadays, for some "longruns" it takes my GTX980ti's almost double time than a year ago.
And the GTX750ti's are brought to the verge of their capibilities with those extreme "longruns" like, as already mentioned, the "ADRIA_FOLDGREED90_crystal_ss_contacts_100_ubiquitin" - for example.
Besides that, as also said above, the latest crunching software increases the crunching time markedly, up to 30% (any idea at all why this is so?)

Hence, as said before, the introduction of 2 types of "longruns" would be very useful. I would guess that quite a number of crunchers are still using the GTX750ti, which definitely is a perfect card for mid-range tasks.

PS - I just pulled a GTX 750 Ti out and replaced it with a GTX 1060 on my Win7 64-bit machine. Too bad the app is not optimized for it now, but I will try it later.

I also was playing with the same idea. However, I am unsure now, particularly after you are saying "the app is not optimized for it now" - what exactly do you mean?

Another setback would be that, at least from what I read somewhere here in the Forum, the Pascals dont's work with the NVIDIA Inspector, which is too bad.

I also was playing with the same idea. However, I am unsure now, particularly after you are saying "the app is not optimized for it now" - what exactly do you mean?

I just mean that 9.18 is not as fast as 9.15, at least from the reports, though I never tried the GTX 1060 on 9.15 so I don't know directly.

But here it is:
http://www.gpugrid.net/results.php?hostid=177999

That is still efficient, but not as much as it could be.

If I compare the crunching times of 918-80 with what they were before with 847-65 / 848-65, I, too, notice a considerable increase.
This is true for my GTX750ti on Windows10 and the GTX970 on Windows 10 - they now need about 30% longer :-(

Since at the same time I had to update the drivers to 381.65, it's hard to say whether the raise in crunching time is mainly due to the new acemd version, or mainly due to the new driver, or both.
What could also be the case is that, for what reason ever, the WDDM overhead now has a much more negative effect.

Hence, the main question seems to be whether anything can be done in order to get the former speed back.

What concerns the change from 847-65 / 848-65 to 849-65 for Windows XP, no WDDM overhead, (one GTX750ti and two GTX980ti), the crunching times seem to be only slightly higher than before.

Would be interesting to hear how the situation is with Linux.

In view of the above-said, plus my experience from the last three weeks, I strongly suggest that in the preferences for download of tasks, beside "short runs" and "long runs", a third category like "extra long runs" or similar, should be introduced.

A typical candidate for such an "extra long runs" would be "ADRIA_FOLDGREED90_crystal_ss_contacts_100_ubiquitin..."
which on my GTX750Ti in the Windows10 machine takes close to 60 hours (on Windows XP with software 849-65 and an older driver, crunching time is considerably shorter).

Beside the extremely, if not rediculously long crunching time, crunching these sort of tasks with the GTX750Ti on Windows10 is rather unstable. Even the slightest overclocking can lead to immediate halts - the only thing one then can do (as soon as one finds out, which might take a while - for example over night) is to suspend and restart the task. So it's been the case here that the total time for such a task was 3 days! Forget about it! A GTX750ti on Windows10 is definitely overtaxed with such a task.

Hence, crunchers with smaller/older cards would definitely wish to exclude tasks like the "_100_ubiquitin..." from download.
Be it coinsidence or not: what I have watched all time long was that my two GTX750Ti tend to receive such long tasks almost all the time, whereas my GTX980Ti's get smaller tasks (like the "Adria") as well. No idea why, it just seems worth to be mentioned (maybe someone knows more about this phenomen).

Anyway, at the bottom line: while it's clear that we crunchers should not be allowed to choose a specific type of task (this clearly remains in the priority determination of the GPUGRID people), a third category "extra long runs" would help quite a lot.
As everyone can imagine: total running time of nearly 3 days is rather frustrating, besides that it does not make a whole lot of sense :-(

any thoughts from GPUGRID side on my suggestion above ?
Good idea - bad idea ?
Feasible - not feasible ?

...
any thoughts from GPUGRID side on my suggestion above ?
Good idea - bad idea ?
Feasible - not feasible ?

still no reaction whatsoever from the GPUGRID team - why so?