GPU Folding Guide

[Garret Jax]

The 8800GTS 640MB can get to the 4300 ppd area with the 177.35 driver and shader domain overclocked to 1700 MHz. I have one currently running

Sweet! I've got my 8800 GTS 640MB overclocked to 652/1000 (don't recall what the shader is at). Do you OC the shader separately from the core? I have them linked. Maybe I should unlink them and see if I can push the 8800 up to 4000 ppd.

 

[Garret Jax]

What nvidia driver are you using ?

I'm using 177.66.

 

[APOLLO]

Well, I may not have been clear in my last post. Sorry.

Prior to building my new PC, I was getting 3000+ ppd with an NVIDIA 8800 on Vista. Now I have the NVIDIA 8800 on XP and I also have a new PC with an ATI 4870 on Vista. I am only getting about 3000 ppd - so either the 8800 has slowed and the 4870 is slow as well, or somehow one or the other card says it's crunching but in fact is not.

Vista is a little faster than XP for GPU clients, so that alone can explain the difference you're seeing with the nVidia card. The Ati box is a new build but as Kendrak mentioned, the client is not optimized yet. Your total PPD with the two GPU clients should be in the neighborhood of 6000-7000, give or take. If your monitoring apps report that amount, then all your clients should be processing their WUs. It will take at least several days before your stats show the corresponding increase in average PPD on the stats pages, though. Sometimes longer.

 

[Xilikon]

Sweet! I've got my 8800 GTS 640MB overclocked to 652/1000 (don't recall what the shader is at). Do you OC the shader separately from the core? I have them linked. Maybe I should unlink them and see if I can push the 8800 up to 4000 ppd.

I overclock the shaders separately by unlinking them. I keep the core at default of 550 (F@H doesn't benefit from core speed, only shader speed help) so overall, it run cooler.

 

[Garret Jax]

Vista is a little faster than XP for GPU clients, so that alone can explain the difference you're seeing with the nVidia card. The Ati box is a new build but as Kendrak mentioned, the client is not optimized yet. Your total PPD with the two GPU clients should be in the neighborhood of 6000, give or take. If your monitoring apps report that amount, then all your clients should be processing their WUs. It will take at least several days before your stats show the corresponding increase in average PPD on the stats pages, though. Sometimes longer.

The alpha-version [H] stats page (which is excellent, by the way) shows updates at 3 hour intervals over the last 24 hours, and my updates show that over the last 24 hours, I've amassed 3100 points - and that is with BOTH pc's running constantly and little to no use of them other than letting them run the GPU2 client.

 

[Garret Jax]

I overclock the shaders separately by unlinking them. I keep the core at default of 550 (F@H doesn't benefit from core speed, only shader speed help) so overall, it run cooler.

Cool! I just unlinked them in rivatuner and set my shader to 1600. Maybe later I'll see how much I can push it.

 

[Xilikon]

Don't forget that the nvidia servers is currently overwhelmed so you may not be able to upload or download units as much as usual so that expain the drop. Use Fahmon instead to give a theorical PPD output.

 

[Garret Jax]

OH YEAH BABY!

So I setup FahMon and it was showing an estimate of 2221 ppd for my ATI box, and had an error for the NVIDIA box. I checked the NVIDIA box - the client appeared to be working, but just in case I 'downgraded' to the 177.35 drivers for the NVIDIA box, and now FahMon estimates that my NVIDIA box will do 4822 ppd! So apparently 177.66 drivers don't work with F@H on my machine, but for some reason they give the illusion of working. Bastards! The only feature I wanted in the 177.66 drivers was the ability to set the video color space, which is not available in 177.35, but that's a small sacrifice to make for an extra 4,822 ppd.

But given that the ATI 4870 kicks the crap out of the NVIDIA 8800 GTS 640 MB in real-world gaming (in my experience) and in 3dMark06, I would expect the ATI box to be getting more ppd than the NVIDIA box (perhaps 6000?) so for now I have to conclude that the above posts were correct: the ATI client is crap. It gets less than half the ppd of the NVIDIA client despite the NVIDIA client having only half the raw GPU power of the ATI client.

Hopefully ATI and Stanford get this resolved soon so we can all crank out even more frames.

Thanks for the help, mates.
-Garret

 

[APOLLO]

But given that the ATI 4870 kicks the crap out of the NVIDIA 8800 GTS 640 MB in real-world gaming (in my experience) and in 3dMark06, I would expect the ATI box to be getting more ppd than the NVIDIA box (perhaps 6000?) so for now I have to conclude that the above posts were correct: the ATI client is crap. It gets less than have the ppd of the NVIDIA client despite the NVIDIA client having only half the raw GPU power of the ATI client.

Hopefully ATI and Stanford get this resolved soon so we can all crank out even more frames.

Mike Houston is currently working on it. Ati will be back in the folding game but can't say for sure when the optimizations will arrive. Glad to see you got your issues resolved.

 

[mle]

You guys are not doing yourselves any favors by talking about overclocking the card for folding. If the person needs a guide to get this working, then it's a sure bet the same person isnt capable of overclocking the card safely, so that there are no calculation errors.

For folding, i believe someone consistently sending incorrectly calculated tasks slows down the project as a whole, and wont be actively participating.

 

[Xilikon]

You guys are not doing yourselves any favors by talking about overclocking the card for folding. If the person needs a guide to get this working, then it's a sure bet the same person isnt capable of overclocking the card safely, so that there are no calculation errors.

For folding, i believe someone consistently sending incorrectly calculated tasks slows down the project as a whole, and wont be actively participating.

There are safety guards in the core which will stop with a UNSTABLE_MACHINE if you overclock too much. Otherwise, the results is pretty valid.

 

[Sunin]

Cool! I just unlinked them in rivatuner and set my shader to 1600. Maybe later I'll see how much I can push it.

Don't forget to jack up the fan speed to 90-100%... I just set it to 100% and forget it myself...

 

[Garret Jax]

For folding, i believe someone consistently sending incorrectly calculated tasks slows down the project as a whole, and wont be actively participating.

Proof or it didn't happen.

 

[mle]

I dont recall which article years ago, where the authors of folding mentioned they feared purpose sabotage and said everything calculated had safeguards, and all units are calculated more than once, then results are compared within the same units. Computers that consistenly send back incorrect results compared to the other controls are still sent units to do, but results are discarded.

Hell if you care enough look it up yourself - what do YOU think will happen if you send in faulty results? You seriously think they dont have safeguards??

As for someone saying if you overclock too much it'll self correct using the core - I guess you have never overclocked your cpu to point where the OS blue screens, and harddrive corruption occurs. And i guess you never overclocked to point where the GPU starts exibiting artifacts.

 

[Xilikon]

I dont recall which article years ago, where the authors of folding mentioned they feared purpose sabotage and said everything calculated had safeguards, and all units are calculated more than once, then results are compared within the same units. Computers that consistenly send back incorrect results compared to the other controls are still sent units to do, but results are discarded.

Hell if you care enough look it up yourself - what do YOU think will happen if you send in faulty results? You seriously think they dont have safeguards??

As for someone saying if you overclock too much it'll self correct using the core - I guess you have never overclocked your cpu to point where the OS blue screens, and harddrive corruption occurs. And i guess you never overclocked to point where the GPU starts exibiting artifacts.

Let's throw the question back to you : Do you think the computer can fold if it will bsod, freeze, crash or artifact ?

 

[Garret Jax]

I dont recall which article years ago, where the authors of folding mentioned they feared purpose sabotage and said everything calculated had safeguards, and all units are calculated more than once, then results are compared within the same units. Computers that consistenly send back incorrect results compared to the other controls are still sent units to do, but results are discarded.

Hell if you care enough look it up yourself - what do YOU think will happen if you send in faulty results? You seriously think they dont have safeguards??

As for someone saying if you overclock too much it'll self correct using the core - I guess you have never overclocked your cpu to point where the OS blue screens, and harddrive corruption occurs. And i guess you never overclocked to point where the GPU starts exibiting artifacts.

Chill, brah. Of course they have safeguards, which means I don't have to worry about it.

Plus, I am not sure that what you say is accurate. Yes, I've OC'd my GPU enough to cause artifacts, but if a few textures are displayed wrong, it's no big deal. As a GPU designer, if I could get greater FPS by accepting a small amount of error (i.e. artifacts), then I would do it. However, NVIDIA and ATI are really pushing the GPU as a GENERAL purpose parallel processor. It seems to me that if either company were to produce a runtime environment for non-graphical work (e.g. F@H), they would be sure to turn off the fault tolerance of the GPU. The error propogation that could occur if you allow even a small percentage of procedures or data points to be in error and not crash the program or throw an error message would be such a mess, no one would use it! So I have to conclude that when OC'ing the GPU, if the card becomes unstable, it will close the program, crash, lockup, or do anything other than continue processing as normal because this would lead to disastor in a parallel programming paradigm.

Think about it like this: NVIDIA and ATI would love for business organizations that do lots of number crunching to make use of GPU's rather than traditional super computers or CPU clusters. Would companies adopt such a system if the results weren't all but guaranteed accurate?

I would be quite surprised - and disappointed - if NVIDIA's CUDA or ATI's CTM actually allowed executing code to have bit switches and lost data. As you say, if you OC your card too much, it produces some artificts, if you OC further, it produces more artifacts, still further and the whole thing locks up. There is some (unknown) inherent function that predicts error rates based on things like GPU load and shader clockspeed. Whatever that function is, if OC'ing your card increases the error rates, which it does, then the CUDA or CTM environment damn well better increase the error catching mechanisms to compenste so that GPU software developers know their data and instructions are uncompromised.

But I could be wrong.

Thoughts?

 

[mle]

Let's throw the question back to you : Do you think the computer can fold if it will bsod, freeze, crash or artifact ?

yes. Prime95 will pass 10 mins if a cpu is moderately to severely overclocked but fail at 1 hour mark with errors. It doesnt mean much for stability but will run. I've seen prime95 fail at 22 hour mark. It doesnt crash, but since it is in diagnostic mode it'll tell you there's an error since it's comparing what the computer calculated against what is pre-calculated. This isnt the same thing for folding.

Now we're talking about very likely about 2 things overclocked: 1)cpu 2)video card. This is an incident waiting to happen unless the system has been thorougly tested.

 

[mle]

"I would be quite surprised - and disappointed - if NVIDIA's CUDA or ATI's CTM actually allowed executing code to have bit switches and lost data. As you say, if you OC your card too much, it produces some artificts, if you OC further, it produces more artifacts, still further and the whole thing locks up. There is some (unknown) inherent function that predicts error rates based on things like GPU load and shader clockspeed. Whatever that function is, if OC'ing your card increases the error rates, which it does, then the CUDA or CTM environment damn well better increase the error catching mechanisms to compenste so that GPU software developers know their data and instructions are uncompromised."


It's speculation whether the folding program catches any errors relating to overclocking. The authors arent going to release this info - as they are very tightlipped about their safeguards as they should be. You're asking an overclocked card to do more calculations to make sure the data is not corrupt, assuming such safeguards within the program exists. Would you even trust it to do that?

It comes down to this - are you folding to brag about being able to fold quickly, or are you there to help the project? I still think suggesting people to overclock is a bad idea for a project such as this, but that's just me.

 

[Garret Jax]

"I would be quite surprised - and disappointed - if NVIDIA's CUDA or ATI's CTM actually allowed executing code to have bit switches and lost data. As you say, if you OC your card too much, it produces some artificts, if you OC further, it produces more artifacts, still further and the whole thing locks up. There is some (unknown) inherent function that predicts error rates based on things like GPU load and shader clockspeed. Whatever that function is, if OC'ing your card increases the error rates, which it does, then the CUDA or CTM environment damn well better increase the error catching mechanisms to compenste so that GPU software developers know their data and instructions are uncompromised."


It's speculation whether the folding program catches any errors relating to overclocking. The authors arent going to release this info - as they are very tightlipped about their safeguards as they should be. You're asking an overclocked card to do more calculations to make sure the data is not corrupt, assuming such safeguards within the program exists. Would you even trust it to do that?

It comes down to this - are you folding to brag about being able to fold quickly, or are you there to help the project? I still think suggesting people to overclock is a bad idea for a project such as this, but that's just me.

I'll make you a deal (seriously): if you can provide any solid evidence (not just rhetorical arguments) that overclocking the CPU and/or the GPU will result in corruption/invalid results in submittable F@H data (i.e. it is not caught and discarded), I will immediately cease and desist OC'ing my systems whenever they are running F@H, and I will help spread the word.

I think this thread is the perfect place to do this, because the [H]ard community consists of a bunch of rabid performance junkies who love to overclock our gear and stroke our e-penors with PPD stats, FPS bragging, and general e-flexing. If you can back up your big talk with grounded evidence, you may single-handedly benefit the bio-modeling distributed computing community more than any other single user.

Cowboy up, mother fucker.

 

[Xilikon]

Anyway, we are not the official forum for F@H, only a team forum. If you want answers to your questions, go to http://foldingforum.org, which is as close as official as you can get, with professor Vijay Pande posting there. Ask him this question and I'm sure you will get the right answer.

Explain to me why the Stanford staff often ask if a user overclocked further than normal to see the variance in production ? If overclocking indeed cause bad science, they should post a stern warning about overclocking but why they didn't until now ? Prime95 is a poor stability test and F@H is more stressful (I often see a computer 100% prime stable fail at the first workunit with a bang).

Unless you have proof or words from the Pande Group, you are just trolling the forum.

 

[mle]

I'll make you a deal (seriously): if you can provide any solid evidence (not just rhetorical arguments) that overclocking the CPU and/or the GPU will result in corruption/invalid results in submittable F@H data (i.e. it is not caught and discarded), I will immediately cease and desist OC'ing my systems whenever they are running F@H, and I will help spread the word.

I think this thread is the perfect place to do this, because the [H]ard community consists of a bunch of rabid performance junkies who love to overclock our gear and stroke our e-penors with PPD stats, FPS bragging, and general e-flexing. If you can back up your big talk with grounded evidence, you may single-handedly benefit the bio-modeling distributed computing community more than any other single user.

Cowboy up, mother fucker.

The point is that you wont know if the results are bad or good. You do know why, dont you. Has to do with what was said about the folding project concerned about people sabotaging the results.

If you for example wanted to run the project into the ground by purposely submitting wrong results, then you would keep trying so long as you were sure the results were getting through and them letting you know that your results arent getting through allows you to try different methods to screw them up.

By the way, how EXACTLY is what im saying trolling. If im trolling then by all means do you best to get me banned.

What i said should be of concern if you truly cared about the project. Do you care about the project? I get the feeling you guys dont.

 

[Thor]

after haveing issues getting dual single clients running, and not being able to keep my box up 24/7 for SMP clients , im happy to report that im seeing 480pts every 2.5 hours on my brand new asus 9600gso top, just so you guys know, the DD tyee fits this cards gpu mounting holes, but not its ram retintion holes.. seeing as how this card did not come with any sort of ram cooling, it dose not worry me, the card dose make contact with the ram , with the exception of 2 chips where it only makes about 3/4 contact, which was good enought, so the card is in the loop and showing 54°c temps (600/1700/1800) on 34° ambient temps not bad for a sub 100$ card

(nothing says love like cooling a 88$ graphics card with a 125$ cooler.. lol)

 

[Thor]

The point is that you wont know if the results are bad or good. You do know why, dont you. Has to do with what was said about the folding project concerned about people sabotaging the results.

If you for example wanted to run the project into the ground by purposely submitting wrong results, then you would keep trying so long as you were sure the results were getting through and them letting you know that your results arent getting through allows you to try different methods to screw them up.

By the way, how EXACTLY is what im saying trolling. If im trolling then by all means do you best to get me banned.

What i said should be of concern if you truly cared about the project. Do you care about the project? I get the feeling you guys dont.

being that we are # 1... and have ALWAYS overclocked our pcs to get the results faster, im pretty sure that stanford would have requested that we not OC at all because they keep seeing bad data come from our team, being that no-one at stanford has EVER said anything to any of us about bad data (its not like they dont know where to find us) i would have to say that the occasional bad data packet that they do get is not enought to worry about. If Ocing caused bad data packets, they would see them from us first and formost.

 

[Xilikon]

being that we are # 1... and have ALWAYS overclocked our pcs to get the results faster, im pretty sure that stanford would have requested that we not OC at all because they keep seeing bad data come from our team, being that no-one at stanford has EVER said anything to any of us about bad data (its not like they dont know where to find us) i would have to say that the occasional bad data packet that they do get is not enought to worry about. If Ocing caused bad data packets, they would see them from us first and formost.

Exactly, we are the biggest team so if we produce bad results, they will know about that and warn about us. Incidentally, they are after a few members with ATI cards since it's disturbing their server : http://foldingforum.org/viewtopic.php?f=24&t=4242 and none of us was in this list.

Remember that if you overclock and get invalid results, it also mean that it can cause a bsod, data corruption, freezes, etc... If you can run the box overclocked 24/7 stable, it mean it produce valid results as well. You are not the official staff so you have no business telling us what to do We also care about results and we already refused to do the multi-GPU setup created by someone since it's not Stanford who did it.

 

[Dan-515]

being that we are # 1... and have ALWAYS overclocked our pcs to get the results faster, im pretty sure that stanford would have requested that we not OC at all because they keep seeing bad data come from our team, being that no-one at stanford has EVER said anything to any of us about bad data (its not like they dont know where to find us) i would have to say that the occasional bad data packet that they do get is not enought to worry about. If Ocing caused bad data packets, they would see them from us first and formost.

I would have thought that Stanford would have made it clear if they did not want people to OC their kit and done something to stop it. Something like if my E6400 is at 3GHz it would know not to run as it expects 2.13GHz. Same thing with an 8800GTX or whatever, if the stock clocks are say 500/1500/1000 up to 600/1600/1050 for some factory OC version it would only run within these clocks. At least this is what I would imagine they would implement.

 

[Garret Jax]

The point is that you wont know if the results are bad or good. You do know why, dont you. Has to do with what was said about the folding project concerned about people sabotaging the results.

If you for example wanted to run the project into the ground by purposely submitting wrong results, then you would keep trying so long as you were sure the results were getting through and them letting you know that your results arent getting through allows you to try different methods to screw them up.

By the way, how EXACTLY is what im saying trolling. If im trolling then by all means do you best to get me banned.

What i said should be of concern if you truly cared about the project. Do you care about the project? I get the feeling you guys dont.

Apparently you did not see the "SERIOUS" tag in my last post. I was not joking. You are not trolling, not at this point, and I did not say you are trolling... in fact, you might be right. However, I think you're wrong for the reasons discussed above. If you are not a troll, then get some evidence to support your theory. If you are correct, than the largest folding team in the history of distributed computing is making a terrible, correctable mistake and should know about it so we can fix it. If you are incorrect, then you should admit it and we should all celebrate by overclocking our GPU's an additional 3%. YEAH!

Now, as I said before:

Cowboy up, mother fucker.

 

[mle]

hey I also overclock - I have nothing against overclocking. I just mentioned that some posts in here advocated overclocking. Without intensive testing the overclocked machine, it could be detrimental to the project.

Post #2, what does it say right in the quick guide? Overclocking. Ya such a super idea to teach the noobs who dont have a clue that intensive testing is needed.

Most of you however do know better. I was concerned about post#2.

 

[Garret Jax]

hey I also overclock - I have nothing against overclocking. I just mentioned that some posts in here advocated overclocking. Without intensive testing the overclocked machine, it could be detrimental to the project.

Post #2, what does it say right in the quick guide? Overclocking. Ya such a super idea to teach the noobs who dont have a clue that intensive testing is needed.

Most of you however do know better. I was concerned about post#2.

OK, for the third and final time: WHAT PROOF DO YOU HAVE THAT OC'ING THE GPU CAN INVALIDATE F@H RESULTS AND HARM THE F@H PROJECT?!?!?!

Stop repeating your comments are start proving them.

 

[Tigerbiten]

Post #2, what does it say right in the quick guide? Overclocking. Ya such a super idea to teach the noobs who dont have a clue that intensive testing is needed.

Whats the matter with ..................

Overclocking the card.
I use RivaTuner to overclock the card and the scan for artifacts test in ATITools to test for stability and find out how hot the card will run under load.
GPU2 folding relies on shader speed then core speed then finally memory speed.
So thats the order to raise the clocks in.

That basically states what I do.
I gave the basic tool needed to overclock a nVidia vid card, also a quick way to test the overclock.
Added are the facts on what alters the PpD the most.

Also I think by keeping it very basic, anybody not to sure on how to overclock, is more likely to ask questions on how to do it correctly.
If you want to expand on that and give a compleate walk through of how to overclock a vid card, with all the pros and cons that go with it, feel free .........
But its a quick guide.

One thing about overclocking and the GPU2 client.
The GPU2 client has a counter which shuts the client down for 24 hours after 5 EUE's.
Thats unlike the GPU1 client that did get stuck in a loop of EUE's and could send in hunderds of results for 0 points.

Luck ..............

 

[mle]

OK, for the third and final time: WHAT PROOF DO YOU HAVE THAT OC'ING THE GPU CAN INVALIDATE F@H RESULTS AND HARM THE F@H PROJECT?!?!?!

Stop repeating your comments are start proving them.

For the final time (minus the caps *g* ), what proof do you have that it isnt harming the project.

You assume this assume then assume that, then say that the folding client will somehow adjust for errors or whatever. That's one hell of an assumption.

Short of just saying you're an overclocker, what else have you said.

 

[APOLLO]

For the final time (minus the caps *g* ), what proof do you have that it isnt harming the project.

Well, for one thing the burden of 'proof' is on you since you're the individual making the controversial claim. I think this subthread has gone long enough. It is way off topic and inappropriate to have it derail a guide. You can start another thread if you wish on the same topic, or perhaps raise the issue in the FCF, where such matters will elicit the type of response I think you're looking for.

 

[Xilikon]

mle, in case you are stubborn or just plain blind, here's what I already said :

Go to http://foldingforum.org/ and ask the same question there. There will be answers to you questions from the official staff at Stanford.

 

[nomad8u]

@mle: How about starting a separate thread if you're so insistent on pursuing this and NOT crapping up the GPU Folding Guide!

To the Team members, maybe just ignoring this thread crapping will stop it from continuing. A full page of this crap in a stickied thread/guide is bullshit IMO!

 

[Xilikon]

From now, please don't feed mle. I asked a admin to do the cleanup.

mle, please make a new thread since you are now threadcrapping and the admins is not fond of this behavior

 

[Garret Jax]

From F@H:

What about hardware clocks?

On 3xxx hardware, 3D clocks will be set automatically when FAH runs and you can adjust the clocks rates for core and memory in Catalyst Control Center in the Overdrive panel. Note that stable clocks for graphics may not imply stable clocks for Folding, overclock at your own risk. On 2xxx hardware, setting to 3D clocks is not reliably automatic, so a third party tool like ATI Tray Tool can be used to adjust clocks. Once again user beware. The recommendation is to leave the settings alone and fold at the clocks set by the driver.

So basically they don't want you to break your machine, but they are NOT worried about overclocked GPU's compromising F@H data. Simply, if your machine is unstable it will trigger the "UNSTABLE_MACHINE" or "EUE" error messages and your results will not be accepted.

LET'S FIRE IT UP MOTHER FUCKERS!!!


I am totally [H]ard for Humanity!!!

 

[Diablo2K]

This thread is to long to read the whole thing. but I have a couple quistions.

I have been running 2 clients for a while now. I just downloaded the GPU client and installed it and set it up, in the settings I chose to raise the priorty slightly as I am also running the 2 CPU clients.

My GPU useage is averageing about 60% useage and deviates alot, and dips down to near zero frequently for a breef moment but never goes above 60%. My temps are holding steady at 60-61deg, and fan speed is around 40-50%. I have riva tuner running but cant figure out how to make adjustments to the fan speed. anyone know where I can find a good tutorial on setting the fan speed, I use to do it with ATItool but it don't work right anymore with my configuration.

 

[APOLLO]

My GPU useage is averageing about 60% useage and deviates alot, and dips down to near zero frequently for a breef moment but never goes above 60%. My temps are holding steady at 60-61deg, and fan speed is around 40-50%. I have riva tuner running but cant figure out how to make adjustments to the fan speed. anyone know where I can find a good tutorial on setting the fan speed, I use to do it with ATItool but it don't work right anymore with my configuration.

There's a 'Fan' tab in the overclocking window of RivaTuner. Select 'direct control' in the drop-down window and just move the sliders where you want them. Most people leave the fan speed at 100%.

 

[Diablo2K]

There's a 'Fan' tab in the overclocking window of RivaTuner. Select 'direct control' in the drop-down window and just move the sliders where you want them. Most people leave the fan speed at 100%.

I know how to set a single speed, thing is I would like to adjust the auto speeds, like increase the fan speed by 10% for each temp lvl. Theres no way I want this video card with the fan runing at 100% just way to loud.

i guess my main quistion is is only getting less than 60% GPU usage normal? i have turned off both CPU clients so that all that is running is the GPU client and the % did not increase at all, its still less then 60% GPU usage.

 

[Sunin]

Generally when you get low GPU usage you need to remove teh drivers, reboot into safe mode, use your favorite driver remover, reboot load 177.35 or higher cuda enabled drivers...

On my recently rebuilt system (moved from xp32 to vista64) I had to do this. I went from frames in the High 2m or low 3m to 1m 16s per frame...

Well unless your talking ATI, low gpu usage on ATI's is common they are "working on it"

 

[Xilikon]

With ATI cards, a low GPU usage mean that you are CPU bound. If your sig is correct, you have a X2 4600+ so it's a bit weak to feed he ATI card. With NVIDIA, you are not CPU bound at all since it have a very low CPU usage.

A way to improve the GPU usage is to overclock the CPU.