NVIDIA GPU2 Client performance -> important information

[axeman]

I thought I'd start a separate thread for this, I started rambling about my 8800GT 512MB not getting the PPD it should have in the "call to arms" thread. And the solution was:

DRIVER VERSIONS MAKE A BIG DIFFERENCE if you want the best possible output from the GPU2 client.

In the middle of a WU, I upgrade my drivers from the "current" WHQL 175.19 drivers to the latest beta NVIDIA has publicly available, 177.79

With the same WU, PPD went from ~2900 to ~4600 Avg. time per step went from over 2 minutes to 1:30

I don't know why I never stumbled upon this little tidbit anywhere else, maybe because most geeks tend to update software and drivers frequently anyhow? In any case, I thought I'd pass it along in case anyone else was unaware, and it might help 2630's PPD output.

edit: Here's some links, because navigating around NVIDIA's site to find things is not always straight forward...

1) Do an advanced driver search here: http://www.nvidia.com/Download/Find.aspx?lang=en-us
In the drop down under "recommended/beta", choose "all", and you'll get a listing of several choices for recent drivers including beta drivers

2) The official "CUDA" drivers can be found here: http://www.nvidia.com/object/cuda_get.html
The driver on this page is the one recommended for Geforce 2xx series cards.

3) The 174.55 driver that is Recommended for 8/9 series cards by Stanford is here: http://www.nvidia.com/object/get_cuda2_beta.html

4) Stanford's FAQ for NVIDIA cards is here: http://folding.stanford.edu/English/FAQ-NVIDIA
If I would have read through it carefully to begin with, this thread might not exist.

[SNM]

It looks to me like the PhysX release came with 177.83, and that's gotta be a WHQL release because it installed on my Vista 64 system.

I haven't been getting 4600ppd on my 8800GT, but then again I've been playing several hours of COD4 each day, and still getting well over 2900.

[TheEmrys]

I thought I'd start a separate thread for this, I started rambling about my 8800GT 512MB not getting the PPD it should have in the "call to arms thread". And the solution was:

MAKE SURE YOU ARE USING THE LATEST DRIVERS if you want the best possible output from the GPU2 client.

In the middle of a WU, I upgrade my drivers from the "current" WHQL 175.19 drivers to the latest beta NVIDIA has publicly available, 177.79

With the same WU, PPD went from ~2900 to ~4600 Avg. time per step went from over 2 minutes to 1:30

This is odd. It isn't that the latest drivers are what you want, its the latest CUDA drivers you need. I don't believe 175.19 was EVER recommended. The recommended CUDA driver was 177.35. As long as the drivers are a CUDA driver, you should be fine. Also, the latest physix/CUDA drivers are great. Gave my 9600GT a bump from 3700ppd to 3800ppd.

[axeman]

I don't believe 175.19 was EVER recommended.

I didn't read anything about which drivers are recommended, other than that if you had the wrong ones, the client would complain. It didn't with 175.19, so I thought I was good. Then again, I'm not so good at RTFM. According to http://folding.stanford.edu/English/FAQ-NVIDIA#ntoc8 I should be using 174.55, the 177.35 is recommended for the 2xx series. I wonder if I would get more PPD yet with the driver...

[Ragnar Dan]

[...] The recommended CUDA driver was 177.35. As long as the drivers are a CUDA driver, you should be fine. Also, the latest physix/CUDA drivers are great. Gave my 9600GT a bump from 3700ppd to 3800ppd.

And thus the natural question for those of us who don't normally play around with updating drivers: where does one find these latest CUDA drivers? All I can find on nVidia's site are the 177.35 version which I've had since July 10 or so when I began using them in my 8800GT-256MB.

It shouldn't have to be said that if one is going to claim a thread has important information, it would be helpful to include such things. Even if you're replying with a correction.

[heyal256]

I think there was another thread that mentioned that you needed to get the latest driver (177.83) from this link
http://www.nvidia.com/content/forcewith ... wnload.asp

[Ragnar Dan]

Thanks for the link.

[axeman]

[...] The recommended CUDA driver was 177.35. As long as the drivers are a CUDA driver, you should be fine. Also, the latest physix/CUDA drivers are great. Gave my 9600GT a bump from 3700ppd to 3800ppd.

And thus the natural question for those of us who don't normally play around with updating drivers: where does one find these latest CUDA drivers? All I can find on nVidia's site are the 177.35 version which I've had since July 10 or so when I began using them in my 8800GT-256MB.

It shouldn't have to be said that if one is going to claim a thread has important information, it would be helpful to include such things. Even if you're replying with a correction.

Ack! You're right. I've edited the first post in the thread with some links.

[jeffry55]

Good thread Axeman! here are some links from the Folding@Home forum on setting up a 9800 GX2.

by grumpydaddy on Thu Aug 14, 2008 10:44 pm

It's been said elswhere but to get me and my evga 9800gx2 ssc up and running I used the info in these:
http://www.hardforum.com/showthread.php?t=1330889
http://forum.xcpus.com/folding-home-...e-nv-gpus.html
http://www.forums.pureoverclock.com/sho ... php?t=3929

When only one GPU was folding I reinstalled the 177.79 drivers TWICE (each card needs setting up with drivers) This in turn allowed Rivatuner to "see" the base clocks and I was able to both Fold and overclock on each card.

[axeman]

Good thread Axeman!

Thanks, it's nice to contribute to the cause! GPU-based folding has really renewed my interest in folding, especially after being frustrated by SMP issues, and the output from the standard client as far as PPD goes is so little by comparison now. I know it's for a good cause, and the standard clients are still important, but we all know the competition of the whole thing is really what drives interest in it.

As an aside, I've started to wonder if this whole effort is a wise use of resources... Cancer research is important, but is conserving energy usage (global warming, not to mention rising energy prices) something that should take precedence in the short-term? I've made my decision based largely on the fact that I end up leaving at least a couple of machines on 24/7 in my house already as it is, so the added energy consumption is probably negligible, especially compared to the kWH the A/C uses this time of year. And in the winter, we need to heat the house anyhow... ; Although electricity is more expensive than the natural gas the furnace uses, it is somewhat more likely to come from a more renewable source (hydro-electric), though it could come from coal (even worse, at least from current coal plants). Not to mention I drive a 12 year old V8-powered gas guzzler on my daily commute, so I'm probably just a hypocrite.

[jeffry55]

Good thread Axeman!

Thanks, it's nice to contribute to the cause! GPU-based folding has really renewed my interest in folding, especially after being frustrated by SMP issues, and the output from the standard client as far as PPD goes is so little by comparison now. I know it's for a good cause, and the standard clients are still important, but we all know the competition of the whole thing is really what drives interest in it.

As an aside, I've started to wonder if this whole effort is a wise use of resources... Cancer research is important, but is conserving energy usage (global warming, not to mention rising energy prices) something that should take precedence in the short-term? I've made my decision based largely on the fact that I end up leaving at least a couple of machines on 24/7 in my house already as it is, so the added energy consumption is probably negligible, especially compared to the kWH the A/C uses this time of year. And in the winter, we need to heat the house anyhow... ; Although electricity is more expensive than the natural gas the furnace uses, it is somewhat more likely to come from a more renewable source (hydro-electric), though it could come from coal (even worse, at least from current coal plants). Not to mention I drive a 12 year old V8-powered gas guzzler on my daily commute, so I'm probably just a hypocrite.

12 years old?! That's no gas guzzler............................. that's a classic!!

[Usacomp2k3]

Not to mention I drive a 12 year old V8-powered gas guzzler on my daily commute, so I'm probably just a hypocrite.

12 years old?! That's no gas guzzler............................. that's a classic!!

So my 19-year old V6 is beyond classic? Technically 25 is when it can be considered a classic in Florida, I believe.

[CampinCarl]

How much are the GPU WU's worth? I just turned in my first one for a whopping...480 points. Plus it took 10 hours to complete. Is that pretty normal for an 8600GTS?

[Pegasus]

Sounds about right for that card. Although doing a quick search of 8600GTS ppd shows a few doing 1400 or 1700 ppd with that model. 7-8 hours time. What drivers are you using? Search cuda 177.35 drivers or maybe the newest ones that came with the physX stuff.

[CampinCarl]

I'm running 177.83. But at least it's performing as it should. Go go gadget FOLD!

[Ragnar Dan]

axeman: You'll have to decide on the value of your energy use and its cost for yourself, but if you're concerned about the media-popularized worries of the day about energy scarcity or AGW, I'd just say there are ways to find out the truth about such matters. Official theories (politically chosen, taxpayer-funded outside of the normal competitive pressures proper science faces) don't have a very good track record over extended periods. Of those two, the latter is now hanging by a thread, and only doing that well because of such funding.

[Pegasus]

Have you guys noticed different WUs for the Nvidia client giving different PPD results.
For example, Project 5504 has been showing up in FahMon at ~400ppd higher than most of the other WUs. On both my 8800GTS 512MB and 8800GT 256MB cards.

[SNM]

axeman: You'll have to decide on the value of your energy use and its cost for yourself, but if you're concerned about the media-popularized worries of the day about energy scarcity or AGW, I'd just say there are ways to find out the truth about such matters. Official theories (politically chosen, taxpayer-funded outside of the normal competitive pressures proper science faces) don't have a very good track record over extended periods. Of those two, the latter is now hanging by a thread, and only doing that well because of such funding.

Okay, ignoring Dan's insane and R&P-diving-into opinions, the amount of energy you use on folding on one computer can't even come close to the amount of energy used in transporting yourself and your goods, or on your A/C, etc.
And we need to go nuclear anyway, so why worry about power?

[jeffry55]

Sounds about right for that card. Although doing a quick search of 8600GTS ppd shows a few doing 1400 or 1700 ppd with that model. 7-8 hours time. What drivers are you using? Search cuda 177.35 drivers or maybe the newest ones that came with the physX stuff.

177.92 is now the latest edition.