Folding Forum

I was reading AnandTech's review of Nvidia's Maxwell cards yesterday (the GTX 750 and 750 Ti) and noticed this little part in the review:

AnandTech wrote:The release 334 drivers include a fairly impressive tune up of NVIDIA’s OpenCL stack, so we’re finding that OpenCL performance is significantly improved in some of our benchmarks, which helps to close the gap with AMD.

Naturally, I had to see if this was the case for my GTX 680 and FAH/FAHBench. Unfortunately, no.

Explicit SP
314.22: 35.4306 ns/day
327.23: 34.8926 ns/day
331.82: 15.5557 ns/day
332.21: 15.6549 ns/day
334.89: 15.1724 ns/day

Implicit SP:
314.22: 137.861 ns/day
327.23: 139.067 ns/day
331.82: 144.298 ns/day
332.21: 144.576 ns/day
334.89: 146.297 ns/day

Back to 327.23 for me...

I have just updated to this release yesterday and so far, I can't fault it.
It has stopped the screen going black for about 3 seconds every so often so,
until it does something very nasty, I will keep an eye on it.

I suspect that "the screen going black for about 3 seconds every so often" was a GPU-Reset and there's a pretty good chance that it caused the active WU to be dumped. I could be wrong, of course, but while those can be attributed to a Driver bug (which was hopefully fixed) but they can also be due to too much overclocking and/or the use of specific screensavers or specific software like Remote Desktop. I really hope you're reporting a successful fix of a driver bug.

I recently updated my 3 folding rigs to 334.89 WHQL.
My PPD has dropped SHARPLY - but it's too early blame this version of the driver.
It could be a rash of poorly designed WUs that got into the AS hopper (some of which take over 24~72 hours to complete on a GTX-770 or GTX-690)!

Nvidia pays a lot more attention to the gaming community than it does to us DC/HPC folks; so any improvements they make are going to be geared almost exclusively towards gaming performance, while DC/HPC concerns would likely be regarded as lower priority.
It might also be the case (but has a lower probability than crappy WUs) that some of the functionality may have been mis-tweaked by Nvidia.

It's too early to pass judgment, but it *IS* appropriate to closely watch folding performance under this version of the driver.

For me, I cannot fall-back because there was something funky in the prior version (327.23) causing intermittent problems with the display freezing-up.
334.89 cured the freeze-up problems.

I have the following GPUs:

• Rig-1: Two GTX-690s (operating like four GTX-680s)
• Rig-2: Two GTX-660ti
• Rig-3: One GTX-770

Before the upgrade, I was averaging about 550,000 PPD over the past few months.
Now, I am lucky to get 150,000 PPD - but I cannot tell if it is due to the driver, or [coincidentally]an influx of poorly designed WUs.

If anybody has accurate before/after stats (same WU project [PRCG]) or running side-by-side (same GPU type & WU project, but different drivers) where the results are obvious and definitive, I think we'd all like to see those.
PLEASE POST if you have solid info - Thanks!

If you post the Project (PRCG), TPF and frequencies (Core/Memory) from each of your GPUs, other donors running the same/similar GPUs but with the older drivers can give you a reply so you can compare the data easily and quickly since performing a driver comparison can take time to do it properly.

Windows 8 64-bit
Driver 320.18 WHQL
EVGA GTX 660 Ti SC @ 1178 MHz & 3121 MHz
Project 9401
Min. Time / Frame : 00:09:58 - 73,418.02 PPD
Avg. Time / Frame : 00:10:18 - 69,883.03 PPD

KWSN_PToT wrote:I recently updated my 3 folding rigs to 334.89 WHQL.
My PPD has dropped SHARPLY - but it's too early blame this version of the driver.
It could be a rash of poorly designed WUs that got into the AS hopper (some of which take over 24~72 hours to complete on a GTX-770 or GTX-690)!

Nvidia pays a lot more attention to the gaming community than it does to us DC/HPC folks; so any improvements they make are going to be geared almost exclusively towards gaming performance, while DC/HPC concerns would likely be regarded as lower priority.
It might also be the case (but has a lower probability than crappy WUs) that some of the functionality may have been mis-tweaked by Nvidia.

It's too early to pass judgment, but it *IS* appropriate to closely watch folding performance under this version of the driver.

For me, I cannot fall-back because there was something funky in the prior version (327.23) causing intermittent problems with the display freezing-up.
334.89 cured the freeze-up problems.

I have the following GPUs:

• Rig-1: Two GTX-690s (operating like four GTX-680s)
• Rig-2: Two GTX-660ti
• Rig-3: One GTX-770

Before the upgrade, I was averaging about 550,000 PPD over the past few months.
Now, I am lucky to get 150,000 PPD - but I cannot tell if it is due to the driver, or [coincidentally]an influx of poorly designed WUs.

If anybody has accurate before/after stats (same WU project [PRCG]) or running side-by-side (same GPU type & WU project, but different drivers) where the results are obvious and definitive, I think we'd all like to see those.
PLEASE POST if you have solid info - Thanks!

It's the drivers. Every 331, 332, and 334 driver has a serious bug that kills folding performance on anything other than GK110 (780, 780 Ti, Titan). They still haven't fixed it.

Judgement passed on several forums already. As mentioned above, still not fixed.

Well I won't be using 334.89 NV driver then.. Had hoped when I saw a new one that it would be the fix.. Am running my GTX 680 alongside my GTX 780ti anyway but performance less than half what it should do.. My other rig out of action so have to live with this for now.

KWSN_PToT wrote:If anybody has accurate before/after stats (same WU project [PRCG]) or running side-by-side (same GPU type & WU project, but different drivers) where the results are obvious and definitive, I think we'd all like to see those.
PLEASE POST if you have solid info - Thanks!

I didn't keep tabs on what PRCGs I ran at the time, but my GTX 680 dropped from 75-90k PPD to 25-35k with anything newer than 327.23.

What I do as a quick test is run FAHBench and look at the explicit single-precision results (see my first post for an example). For me, they mirror any change in TPF in the actual program.

Many other people on many other sites have experienced the same performance issues. No need to reinvent the wheel.

I´ve got the Info from a member of my team that he has got a boost in PPD folding with a GTX560Ti under Win 7 with driver 334.89WHQL from 18,2K auf 21,3K PPD (WU 8900).
Maybe someone is able to verify this result?

I will try to get more results from my member.

folding_hoomer wrote:I´ve got the Info from a member of my team that he has got a boost in PPD folding with a GTX560Ti under Win 7 with driver 334.89WHQL from 18,2K auf 21,3K PPD (WU 8900).
Maybe someone is able to verify this result?

I will try to get more results from my member.

Without knowing the previous version of driver, it's hard to draw any conclusion.

folding_hoomer wrote:I´ve got the Info from a member of my team that he has got a boost in PPD folding with a GTX560Ti under Win 7 with driver 334.89WHQL from 18,2K auf 21,3K PPD (WU 8900).
Maybe someone is able to verify this result?...

I think that I read somewhere that Fermi GPUs aren't effected by this bug. However, I can't be sure if that was a fluke case or not.

i added a 750ti today so am using the 334.89 drivers.

9401: 560ti - 16' 44" - 33,750/ppd

ud

i7-3770
Win7 x64
MSI GTX 660ti overclocked to 1084MHz

I doubled my GTX 660ti GPU throughput by rolling back my nvidia drivers from 331.82 to 327.23; i.e. ETA for a new GPU work unit changed from 20 hours to 10 hours. So, if you are interested in FAH performance on the GTX 660ti, the 327.23 drivers are the way to go.

Now all I have to do is figure out why CPU folding throughput has dropped to about 50% of where it was a few months ago ...

Anyone else having the issue where coming out of screen saver I get hard lockups that require power-down/hard reset to recover from? I have a GTX 260 and 334.89.