Folding@home 3850 AGP

[Kougar]

Gratz on those performance numbers GFreeman, I would'nt have thought a GT 425 would even give you that much in games!

Rob, DarkStarr, it's a tough call to answer. I'm sure some part of it is optimizations... but CUDA is just more efficient at this specific type of workload. Even taking into effect the doubling of performance for AMD cards... reported CPU usage is still ~25% for AMD folders. Guess how much CPU overhead is required for Folding on my GTX 480 @ 15,000 PPD? Less than 1%. Even if I drop down to a Quadcore system, it still measures the CPU processing overhead at <1% for the GPU client. Basically the entire Folding@home project can remain self-contained inside the GPU, with none of it getting outsourced to the CPU for computations it can't handle.

A very, very long time back NVIDIA's GPU folding used to eat ~10% of a four-thread capable system, so I am sure AMD's 25% processing overhead can be dropped significantly. But it has a huge way to go to still be comparable to NVIDIA, even assuming the PPD comparisons were equally matched, which they still are not.

 

[GFreeman]

Thanks Kougar! That is thanks to the overclock I run on it. It's overclocked to 37% I've also been playing with my fan speed settings so it runs 80 degrees tops when folding now. For folding I'm running this system at the lowest possible multiplier and vcore. According to AIDA64 it's using 6 watts running@700 Mhz with the lowest memory devider That's when the load is only on the GPU folding.

Here's some AIDA64 benchmark. I'm suprised what this little puppy puts out with a little tweaking / overclocking. I've included a zip file (contains PDF benchmark report).

 

[GFreeman]

Gratz on those performance numbers GFreeman, I would'nt have thought a GT 425 would even give you that much in games!

Rob, DarkStarr, it's a tough call to answer. I'm sure some part of it is optimizations... but CUDA is just more efficient at this specific type of workload. Even taking into effect the doubling of performance for AMD cards... reported CPU usage is still ~25% for AMD folders. Guess how much CPU overhead is required for Folding on my GTX 480 @ 15,000 PPD? Less than 1%. Even if I drop down to a Quadcore system, it still measures the CPU processing overhead at <1% for the GPU client. Basically the entire Folding@home project can remain self-contained inside the GPU, with none of it getting outsourced to the CPU for computations it can't handle.

A very, very long time back NVIDIA's GPU folding used to eat ~10% of a four-thread capable system, so I am sure AMD's 25% processing overhead can be dropped significantly. But it has a huge way to go to still be comparable to NVIDIA, even assuming the PPD comparisons were equally matched, which they still are not.

What I'm about to say now is how I feel about it and not factual so feel free to correct me haha! I have a feeling that nVidia get more out of there cards using technologies like CUDA.. If you look at the AMD cards it looks like they are about raw power. They have a lot more shaders if you compare them to the nVidia cards. Ofcourse nVidia is using a much different architecture compared to the AMD cards. It seems that AMD should even get a lot more performance out of there cards I guess..

 

[Kougar]

What I'm about to say now is how I feel about it and not factual so feel free to correct me haha! I have a feeling that nVidia get more out of there cards using technologies like CUDA.. If you look at the AMD cards it looks like they are about raw power. They have a lot more shaders if you compare them to the nVidia cards. Ofcourse nVidia is using a much different architecture compared to the AMD cards. It seems that AMD should even get a lot more performance out of there cards I guess..

In a nutshell, that's true. The important thing for everyone to remember is that Fermi was the first GPU architecture where NVIDIA combined their gamer & Tesla/Quadro market designs into a single core, a single product.

Fermi had to deliver all the GP-GPU and other workstation workloads, while at the same time being a powerful desktop GPU for gamers. Obviously in trying to be both, some compromises got made... but the point still stands Fermi was built for GP-GPU. AMD's designs are built almost entirely with gaming in mind. Until AMD decides to increase their core-size by incorporating additional GP-GPU hardware, there's always going to be a difference.

 

[DarkStarr]

See however you look at it its not CUDA, cuda is ~= to OpenCL which is what AMD uses now. Then take a look at all other distributed computing, AMD pwns typically several times over. I think they are not really trying because OCL is pretty similar to CUDA so therefore they could easily transfer optimizations that do the same thing on both cards and start on specific AMD ones. They have not. They implemented a BAD version of OCL, its stupid. I used a OCL bench tool and my 5750 scored higher than my 275 yet folding, my 275s each get ~9k while the 5750 doesn't even do half that.

The combined score (each 275 did about 50%) for the Nvidia cards was C1753.
The ATI alone was C2772.8.
This was done with the 275s highly OCed and the 5750 at scock 700/1150, I could easily push it to 870/1370.

So in essence my 5750 should kick ass in folding but it doesn't. Wonder why?

 

[GFreeman]

Thanks for the detailed explanation Kougar! Well I think AMD cards should do much better as they now score as well.. Let's hope there will be better future support for the AMD cards too..

Anyways, I have been fiddling with my bios settings and got a better overclock out of the CPU with adding a tiny little vcore The CPU can be overclocked more, but then I have to settle with a lower memory devider and I don't like that I had to sacrifise a few Mhz on the ram, but it's still running @1400Mhz at CAS8. I'm happy with the performance it's much better then I expected. No throttling, it ran at 3150Mhz all the way This is with stock cooling and this machine is tiny. It runs a little bit hotter on full load, but not by much.

 

[GFreeman]

That makes quite a difference to the score in 3dmark06

Here's 8819: http://3dmark.com/3dm06/16010672

 

[marfig]

Umm... the CPU will not affect your folding performance, GFreeman. F@H GPU client doesn't make use of GPU-CPU multithreading facilities in CUDA. It makes exclusive use of the GPU.

 

[GFreeman]

I know that When folding I'm only running@700Mhz with the lowest vcore setting This configuration is when I use it for gaming every now and then. This system has become my main rig. I prefer it to the older X2 actually..

 

[GFreeman]

It's actually folding for 12/14 hours a day average. It folds when I'm at work. It's a little too noisy for night folding I have one room all in it at the moment. It is silent, but not silent enough for that purpose haha! Still looking for an apartment. If I have a seperate bedroom then it will fold at night as well.

 

[marfig]

Oh gotcha. Didn't realize you were thinking of a different purpose for it.

 

[GFreeman]

I have become a little too fond of it as a system too haha! But that doesn't mean it won't be folding. It will for the most part. This thing has 3d vision as well.. and a great 7.1 soundcard with digital output and USB 3.0.. Some features I really like. I've allready watched Avatar in 3D with it that was cool! I'm thinking making it my primair system for watching TV on the internet as well. A company is working at making TV channels online available in Holland. That would be fun having one multimedia system for such wide use. It even has a remote!

 

[GFreeman]

Windows is now reporting these scores with the i3 370m clocked at 3150 Mhz 1.176v of vcore Not so bad for this little HTPC. It has folded about 45 WU's so far.

 

[DarkStarr]

I reran the OpenCL bench on my new 480 and got

GTX 480 C4281.1 at Stock on it.
GTX 275's (SLI) C1753.
ATI 5750 C2772.8.

So from those results their is no reason the 5750 should be folding that low.