Folding Forum

Nvidia has released a new set of drivers (http://www.guru3d.com/news_story/downlo ... river.html).

Have anyone else tried these drivers on their folding GPUs? If so, please comment on your experience.

I am using 331.82 on a GTX 660 Ti (967 MHz default, 1136 boost clock). HFM.net shows 41645 PPD (should be around 67 K). The FAH Control shows 31361 PPD (should be around 56 K). This is with only about 6 percent of a P7810 work unit finished, but the PPD has stabilized now and it should be enough to show that OpenCL is still not working right.

EDIT: Going back to the 327.23 drivers (on Win7 64-bit), at 22% complete this same work unit has stabilized at 80 K PPD on HFM.net, the best I have ever seen. On FAHControl, it is currently at 56 K PPD as expected, but bounces around and has been higher (up to 102 K PPD).

But I did delete the previous version of FahCore_17.exe before starting these tests, so it may have helped to get the latest version.

Same as the other 331's on my 780s.

One question slightly off topic. Currently I use driver 314.07 (GTX 650Ti; Win7/64). Is there a newer driver which gives advantage of the TPF?

I've seen many recommendations for 326.80 beta for GK104 Keplers, but I'm not sure how they compare to drivers that old.

artoar_11 wrote:Currently I use driver 314.07 (GTX 650Ti; Win7/64). Is there a newer driver which gives advantage of the TPF?

I don't think so. I have used 314.07 (also on Win7 64-bit) on a GTX 660 Ti, GTX 600 and GTX 650 Ti at various times. Upgrading to the later ones doesn't seem to change anything, except past 327.23 you fall off a cliff. Any improvements otherwise are probably due to the work units, though I would also suggest deleting FahCore_17.exe (C:\ProgramData\FAHClient\cores\www.stanford.edu\~pande\Win32\AMD64\NVIDIA\Fermi\Core_17.fah) and allowing it to download the latest version, which might help.

I just hope Nvidia fixes this issue soon for the GK104's, it's the same thing in Linux also.

I installed driver - 327.23. GTX 650Ti @ 1.125 GHz; Mem @ 5600 MHz.

p8900: Avg. TPF: 00:08:54; Avg. PPD: 31 720

Old driver - 314.07
p8900: Avg. TPF: 00:09:09; Avg. PPD: 30 430

About 4% faster

So do we all think it's because of Nvidia folding results are very poor past 327.x drivers? I know the CUDA version changes to 6000 (from memory it shows up as 5050) on the ones where folding results are "normal", can this have any impact?

error-id10t wrote:So do we all think it's because of Nvidia folding results are very poor past 327.x drivers? I know the CUDA version changes to 6000 (from memory it shows up as 5050) on the ones where folding results are "normal", can this have any impact?

Core 17 performance is the issue at hand, and as core 17 runs exclusively OpenCL code, the CUDA version of the drivers should have no effect.

error-id10t wrote:So do we all think it's because of Nvidia folding results are very poor past 327.x drivers? I know the CUDA version changes to 6000 (from memory it shows up as 5050) on the ones where folding results are "normal", can this have any impact?

As Zagen30 just said, Core_17 doesn't use CUDA; it uses OpenCL. Both NV and AMD package OpenCL with their drivers, and a driver update from either of them can alter performance, fix (or create new) bugs, etc. and WHQL testing is supposed to validate that no new ones have been introduced.

Any GPU core can be adapted for OpenCL (perhaps with some accommodation to allow for differences between what works and what doesn't work in the subset of OpenCL supported by each OEM). Since CUDA is nvidia proprietary, adapting a core for CUDA may or may not be possible, again depending on the features that NV supports compared to what the FahCore really needs and whether it's worth the extra programming and testing time to make an NV-only version of that core.

Development carefully rations any possible benefits vs. those costs. A future CUDA version is possible, of course, but as always, the future is unpredictable.

I have found that using the latest nVidia drivers 331.82 or 331.58 cut my production nearly in half versus 314.07

If I reinstall an older version like 314.07 or 310.70 my total PPD using a 660Ti and 680 +3370K with cpu-5 goes up from 100-120K to 160-190K PPD. I let the driver exe file do the "remove older driver" so there might be remnants of the previous installs somewhere in the registery but since I don't know where to look I assume the install program takes care of all the details. The system reboots between each driver install and nothing else is running in the background of any significance.

Using FAHBench from Proteneer to benchmark the different drivers the test times in OpenCL
for the 331.82 driver Using MSI TwinFrozer
Single Precision data only
.........GTX680............GTX660Ti (averaging at least 5 samples)
Explicit= 14.92___________13.61
Implicit=137.21__________135.59 ns/day

for the 314.07 drivers.
Explicit= 33.31 (seems very high compared to other benchmarks I have seen but...)
Implicit=132.2

what is wierd is that when I ran the benchmark using 314.07 a month ago and exactly the same hardware with the same 314.07 driver it gave me
single precision
..........GTX680...........GTX660Ti (averaging at least 5 samples)
Explicit=17.55___________16.24
Implicit=87.61___________84.62 ns/day
no idea what I did different data collected after a reboot with nothing else running in background according to task manager

testing on 11/28/13- (I gave F@H 3 hours working with each WU-these take 9-10 hours to complete on a good day)
Using WU 8900 (142,5,6) the newer driver (331.82) TPF=8 min 51 sec
Using same WU but with the older driver (314.07) TPF=4 min 56 sec
Using WU 8900 (421,2,16) on the 660 card with same drivers:
TPF=9 min 37 sec versus 5 min 19 sec.

testing on 11/29/13- using 314.07
8900 ( 41,8,1) GTX680 TPF=4 min 48 sec with credit=26376 (estimated by F@Hv7.3.6)
8900 (197,5,2) GTX660Ti TPF=5 min 20 sec with credit=25019

data today:
8900 ( 0, 47, 280) GTX680 TPF=4 min 51 sec Credit=26364 PPD=79091
1111MHz Core/1502 GPU Memory Clock, 1.160VDDC, 12.02v, 35ampVDDC, 63C, 96-98% GPUload, 58% TDP

8900 (438, 2, 26) GTX660Ti TPF=5 min 20 sec Credit=25039 PPD=67500
1189MHz Core/1502 GPU Memory clock, 1.1750VDDC, 12.05v, 28ampVDDC, 64C, 97-98% GPUload, 68%TDP

I use all devices at native speeds-no deliberate overclocking as that seems to make things unstable.
I am running FahCore_17.exe 8/3/2013 using 314.07 drivers for my MSI cards.
The slots are configured for client-type=beta and I download a new WU when I have completed 99% of the work.
Slider is on FULL
3770K@3.5GHz cpu-5

The total system is pulling 504 watts from a Seasonic Platinum 1000W PSU which seems low but that's what the meter says.
Cards are at 60-65C and CPU at 70C.

ThePretzelGuy
ASUS P8Z77WS-3770K >2x8GB Mushkin Redline 1600 > CM 212+ Cooler > MSI TwinFrozr GTX680 2GB & GTX660Ti MSI TwinFrozr 2GB @stock > Samsung 840Pro 256GB > WD 2TB Caviar Black > 2x ASUS DVD > Seasonic Platinum 1000W > Win7 Pro x64> CoolerMaster HAF X Full Tower

PS 8900WUs seem to provide much fewer points than 78xx projects for the same amount of time. Use 314.07 usless you need a patch for gaming

Welcome to the F@H Forum ThePretzelGuy,

Please note that setting client-type to Beta isn't doing anything for FahCore_17 since it has been released to Full F@H (doesn't need any client-type set) on Windows (http://folding.stanford.edu/home/change ... -full-fah/). Moreover, its usage is restricted to Beta Forum as per our Forum rules (viewtopic.php?f=66&t=18031).

Regarding the loss of performance, thanks for providing us with additional data. We are aware of the situation and the development team is looking into this issue. I hope that this issue will be resolved in the future. I am currently using WHQL 327.23 on my GTX 660 Ti and GTX 650 and it is folding along nicely.

I hope you have reported this problem on the NV forum. Drivers are the responsibility of the manufacturer and for the most part, FAH developers can do no more that report the problem to them.

780Ti's love, and require, 331. 660Ti's hate them and run at half PPD. As long as i don't mix them, I can run 327 on the 660's and 331 on the 780's. Everybody's happy.

edit- A regular 780 will run on 327.