• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

Experimenting with NVIDIA GPU2 performance

Parja

Parja

[H]F Junkie
Joined
Oct 4, 2002
Messages
12,671
I did a clock speed experiment over the weekend on my system to really get a grip on how core, shader, and memory clocks effect GPU2 client performance.

I started with clock speeds at stock and then varied them upward and downward from stock to see their effect on the time required to finish 1% of a WU. For each setting, I let it finish about 5% and averaged the time (rounded to the nearest half second).

The results were pretty much as expected. Core and memory have a small effect on performance, but the shader clock is where the performance is (by roughly a factor of 5 over the other two clocks by percentage change). From maximum overclock to a 20% underclock, core and memory only cost about 1 additional second per percent. I don't have a Kill-A-Watt or similar to test the power consumption (yet...ordered one today...yay!), but I've got to believe the card is going to use significantly less power with those two clocks dropped that much.

With the shader overclocked 21%, time per % dropped a full 2.5 seconds (from 18.5 to 15), or almost a 20% drop in time per %. So performance changes almost linearly with the change in shader clock.

gpu2.jpg
 
So what are you running it at now? And how high can you get your shaders if you run default 8800GT clocks of 600core and 900 mem?

BTW, nice job on the graph, I like the results :)
 
Imitation said:
So what are you running it at now? And how high can you get your shaders if you run default 8800GT clocks of 600core and 900 mem?

BTW, nice job on the graph, I like the results :)
Click to expand...

I'm running with everything maxed right now. I can't actually get my shaders any higher than 1890, even with the core and memory underclocked, though the Accelero S1 I've got cooling my 8800GT might have something to do with that. Those with stock cooling might have different results.
 
With rivatuner, you can have overclocking profiles, one for gaming and one for folding. For gaming, it make sense to overclock all 3 for the best performance but if it's just folding, drop the core/mem to stock or even below and increase the shaders clock so it won't take that much power and generate heat.

Your observations is consistent with what I saw around.

 
Thanks for spending the time to put a plot together and post it for all of us. I have a question about how you set up your test rig.

On one if my boxen (Win XP 32-bit, Q6600@3.2) I have been running the 174.55 drivers from NVIDIA. Everything works ok and I am getting ~4500ppd from a 8800GT (with SMP going as well), however the computer lags so bad as to be rendered unusable except for folding. How did you install these new 177.35 drivers? I got them from NVIDIA's CUDA site and when trying to load them on it said it wouldn't work with my hardware which is the same as yours. It was easier getting this running on Vista 64 bit using the hacked drivers. :confused:

Any help would be greatly appreciated.
 
HighYield said:
Thanks for spending the time to put a plot together and post it for all of us. I have a question about how you set up your test rig.

On one if my boxen (Win XP 32-bit, Q6600@3.2) I have been running the 174.55 drivers from NVIDIA. Everything works ok and I am getting ~4500ppd from a 8800GT (with SMP going as well), however the computer lags so bad as to be rendered unusable except for folding. How did you install these new 177.35 drivers? I got them from NVIDIA's CUDA site and when trying to load them on it said it wouldn't work with my hardware which is the same as yours. It was easier getting this running on Vista 64 bit using the hacked drivers. :confused:

Any help would be greatly appreciated.
Click to expand...

Actually, I still had lag with the 177.35 drivers. I installed the 177.39s last night and was able to play Mass Effect with the GPU2 client running with no noticeable choppiness.

Anyway, all I did to get them installed was copy the nv4_disp.inf file from the 174.55 driver directory to the 177.35/39 directory. I was then able to install them no problem. The only thing I changed was the "DriverVer" entry on the inf file to display the correct version.

 
Parja said:
Actually, I still had lag with the 177.35 drivers. I installed the 177.39s last night and was able to play Mass Effect with the GPU2 client running with no noticeable choppiness.

Anyway, all I did to get them installed was copy the nv4_disp.inf file from the 174.55 driver directory to the 177.35/39 directory. I was then able to install them no problem. The only thing I changed was the "DriverVer" entry on the inf file to display the correct version.

Click to expand...

Finally getting around to trying this out with the 177.35 drivers. Got the 177.35s unpacked. Copied over the 174.55 nv4_disp.inf and then pasted in the date and driver version from the 177.35 file. When I go to set up the driver it is asking for nvdispsr.dll file saying it is not found. Any ideas how to fix this?

 
Does anybody know how to track these? When i left this morning, my p68 had finished two wu's and was almost done with the third and although my log file says puppeteernessus and team 33, i dont see them as added to my stats. I have a program that tracks my stats on my e2180 smp computer, but i think it just tracks that machines ppd. Is there a program that tracks the GPU2 performance or is it too new?
 
I use RivaTuner v2.09.
Its the only tool I've found that lets you alter the three clock speeds.
Most others I've tried just let you alter the core & memory speeds but not the shader speed.

I then use "scan for artifacts" in ATITool 0.26 to test the overclock.

Luck ........... :D
 
I just stared using Rivatuner for my lappy...kept the clocks stock (the gpu2 client doesn't seem to like it when I overclock them) and bumped the shaders from 900 to 1000...will check results when I get home. Currently getting 10 minutes per %.
 
PuppeteerNessus said:
Does anybody know how to track these? When i left this morning, my p68 had finished two wu's and was almost done with the third and although my log file says puppeteernessus and team 33, i dont see them as added to my stats. I have a program that tracks my stats on my e2180 smp computer, but i think it just tracks that machines ppd. Is there a program that tracks the GPU2 performance or is it too new?
Click to expand...

I use FahMon to track all my clients but FahSpy works just as well.

You've got to be able to point the monitoring software at your data files to track the client.
Look in C:\Documents and Settings\<your name>\Application Data\Folding@home-gpu.
Then point FahMon at this folder to track the client.
Depending on how you've set your computer up this could be a hidden folder.
So you may have to alter show hidden files in the Control Panel -> Folder Options to see it.

Luck ........... :D
 
lelliott731 said:
What tools do you use to modify the different clock speeds?
Click to expand...

RivaTuner for me as well. You can modify the clocks on the fly (without any sort of stupid stability test to set them) and even save clock speed presets.

 
Tigerbiten said:
I use RivaTuner v2.09.
Its the only tool I've found that lets you alter the three clock speeds.
Most others I've tried just let you alter the core & memory speeds but not the shader speed.

I then use "scan for artifacts" in ATITool 0.26 to test the overclock.

Luck ........... :D
Click to expand...

You can also try the NVIDIA system tools. I can't vouch for all of them but I downloaded it when I started using the hacked vista 64 driver. It let me independently increase my shader speed and leave the core and memory at stock on my 8800 GTX.

http://www.nvidia.com/object/nvidia_system_tools_6.02.html
 
You Rivatuner fans that don't use all the complexities RT has to offer might give EVGA's Precision Overclocking tool a try. It was developed by the RT author but has a significantly better interface.
 
Tigerbiten said:
I use FahMon to track all my clients but FahSpy works just as well.

You've got to be able to point the monitoring software at your data files to track the client.
Look in C:\Documents and Settings\<your name>\Application Data\Folding@home-gpu.
Then point FahMon at this folder to track the client.
Depending on how you've set your computer up this could be a hidden folder.
So you may have to alter show hidden files in the Control Panel -> Folder Options to see it.

Luck ........... :D
Click to expand...

Great! I got it too work. Took me a second to realize all i had to do was click in the open space to tell it where the client was, but i got it. Thank you sir! :p
 
Thanks for the info. This is obviously different than how one would OC for gaming and makes it much easier if you're going to set up dedicated GPU2 folders

 
Ok, got my Kill-A-Watt in today.

Watts pulled from the wall was measured on the "Main Box" in my sig with the NVIDIA GPU client and one console client running. The clocks were varied on the video card to see the difference they have on power consumption.

Core/Shaders/Memory clock - Watts (GPU2 client running) - Watts (GPU2 client not running)
626/1566/900 (stock) - 190W - 140W
626/1890/900 (maxed shaders, rest stock) - 195W - 140W
500/1890/720 (maxed shaders, rest -20%) - 191W - 138W
756/1890/1053 (all maxed) - 200W - 142W

So with the client running, looks like the clocks don't make too big of a difference in power consumption.

Coincidentally, going from the single console client running to not running only amounted to a 19W drop in power consumption.

 
Parja said:
<snip>

Coincidentally, going from the single console client running to not running only amounted to a 19W drop in power consumption.

Click to expand...

Well then.... fold on! The power man wont get you today :p

 
You must log in or register to reply here.
Back
Top