The Folding@Home GPU Statistics Database

[P5-133XL]

compdewd, can you add power consumption to your stat? I wanna find the max PPD per watt.

Thanks

http://forums.atomicmpc.com.au/index.php?showtopic=264 and apply PPD and you've got it assuming that your card is listed.

[Hotmit]

Thanks alancabler & P5-133XL for the warm welcome and the info. I am considering to get the GTX 560 Ti, I just like to know what other contender out there with the similar horse power and efficiency.

I want to fold but I am also green & green conscious.

[alancabler]

Thanks alancabler & P5-133XL for the warm welcome and the info. I am considering to get the GTX 560 Ti, I just like to know what other contender out there with the similar horse power and efficiency.

I want to fold but I am also green & green conscious.

If you want to optimize GFLOPS/watt, then study that first link I gave you and see what you can fit within your budget. You'll see that a 560Ti processes @ 1263 GFLOPS/170watts while a 650Ti is rated ~1420 GFLOPS/110 watts.
Initial purchase price is about the same, but you'll save greenbacks with daily energy usage with the Kepler (650Ti). Is there any other "green conscious" aspect of this issue?
BTW, Kepler drivers aren't completely optimized for f@h yet, so their performance will increase in future.

[P5-133XL]

I wouldn't be using the GPU comparison database list to base PPD on till the QRB for GPU's is finalized and incorporated. Using conventional non-QRB WU's is going to distort the long term PPD/W calculation towards the lower end cards. My point is that it is all going to change ... radically.

[devpao]

CLUB RADEON HD6670 1GB
clocked = 850 Core 742 Mem
Driver Ver = CCC 11.12
Client Ver = 7.2.9
Project = 11292
Base Credit = 2224
TPF = 6m 34s
PPD = 4876

[GreyWhiskers]

Thanks alancabler & P5-133XL for the warm welcome and the info. I am considering to get the GTX 560 Ti, I just like to know what other contender out there with the similar horse power and efficiency.

I want to fold but I am also green & green conscious.

I just installed this week a new GTX660Ti to replace an ailing MSI GTX560Ti whose TwinFrozr fans siezed up for the second time in 18 months. The 560Ti is on its way back to MSI for RMA repair - and have some really interesting results. I have a much more detailed review in the Beta forum - but I will give the high points here.

the 660Ti is a Kepler card - which is a rearchitecture of the Fermi CUDA core system. The 660Ti is the low end of the GK104 Kepler core - a family of GPUs (660Ti, 670 and 680). the 660Ti has disabled some of the CUDA cores and has a narrower memory path than the other higher members of the family. See comparison chart from AnandTech at end of post. Note that the card I chose came from the factory with a much higher base clock than the Nvidia reference for the GTX660Ti.

Bottom line - as it came from the factory with the 1046 MHz base clock, my GTX660Ti card produced maybe 5% faster TPF than the 900 MHz 560Ti it replaced. While getting these results, it used 117 watts over the idle state, vs 162 watts for the 560Ti - much greener. And it ran about 53 deg C vs the 560Ti at 70 deg C. [Windows 7 Home Premium; Nvidia 306.97 drivers; Folding at Home v7.2.9]

Now, the Folding at Home Core 15 Nvidia software isn't optimized to take advantage of the Kepler characteristics - so there is the hope for even better performance if and when the Gromacs and/or Core 15 software gets those optimizations.

I'm happy so far.

Some basic info follows:

Factory OC settings:
Base Clock: 1046 MHz
Boost Clock: 1124 MHz
Memory Clock: 6008 MHz
Cuda Cores: 1344
2048MB GDDR5 192-bit
The spec core clock on my GTX660Ti is 1046 with boost to 1124. MSI AFterburner hardware monitor shows core clock at 1045 MHz during Core 15 loading the card at the beginning of a new WU, jumping to 1241 once the GPU usage goes to 99% and the card is into the actual computation.

For comparison, to the specs above, the old GTX560Ti - 384 CUDA Cores- reference clock - 822 MHz - my installation OC to 900 MHz

Overall - the card seems to be a quiet, low power consumption effective folder - with expectation that it will probably do even better if and when the GROMACS and FAH core 15 software gets Kepler optimization.

Caveat: YMMV

Comparison chart from AnandTech GT660Ti Review

Code: Select all

                  GTX 680     GTX 670    GTX 660 Ti     GTX 570
Stream Processor    1536        1344        1344          480

Texture Units       128         112          112          60


ROPs                 32          32          24           40
Core Clock        1006MHz      915MHz      915MHz       732MHz

Shader Clock        N/A         N/A          N/A        1464MHz

Boost Clock       1058MHz      980MHz      980MHz         N/A

Memory Clock     6.008GHz    6.008GHz      6.008GHz     3.8GHz
                    GDDR5       GDDR5       GDDR5        GDDR5
Memory Bus Width  256-bit     256-bit      192-bit      320-bit

VRAM                2GB         2GB          2GB        1.25GB

FP64             1/24 FP32   1/24 FP32    1/24 FP32    1/8 FP32

TDP                 195W        170W        150W         219W

Transistor Count    3.5B        3.5B        3.5B          3B

Manufacturing    TSMC 28nm   TSMC 28nm    TSMC 28nm    TSMC 40nm
Process

Launch Price       $499        $399         $299         $349

[kiore]

I am interested to see a comparison between the GTX 660ti and the GTX 670 (for folding) which really have very similar specs but not similar prices where I could buy one.

[bruce]

I am interested to see a comparison between the GTX 660ti and the GTX 670 (for folding) which really have very similar specs but not similar prices where I could buy one.

My simple-minded observation: The main difference between the two is in the quantity of rasterization and memory control. The extra hardware in the 670 would probably be significant in some games but would be pretty much useless for FAH. (well, except maybe for the Viewer.) For me, the hardware reduction saves power and heat but probably wouldn't change FAH's productivity.

[kiore]

I am interested to see a comparison between the GTX 660ti and the GTX 670 (for folding) which really have very similar specs but not similar prices where I could buy one.

My simple-minded observation: The main difference between the two is in the quantity of rasterization and memory control. The extra hardware in the 670 would probably be significant in some games but would be pretty much useless for FAH. (well, except maybe for the Viewer.) For me, the hardware reduction saves power and heat but probably wouldn't change FAH's productivity.

That is just what I was thinking, that for folding the GTX 670 seems to hold no advantage over the 660ti, this is different for the previous 5 series where the 570 is significantly more powerful that even the 448 shader version of the 560ti.

[csvanefalk]

Does anyone have some hard data on the output of the GTX690? Trolling the web, mostly EVGA forum, I have seen figures ranging from 33k-45k PPD.

On a more curious note: if we take the upcoming Kepler optimizations and QRB for GPUs into account, what could be the theoretical PPD output of a 690? A 660 chewing one of the QRB test WUs produced around 100K PPD without Kepler optimizations...would it be unrealistic to assume we could see 500k+ PPD from the 690 with optimizations on and the QRB in effect??

[Ben_Lamb]

I am interested to see a comparison between the GTX 660ti and the GTX 670 (for folding) which really have very similar specs but not similar prices where I could buy one.

My simple-minded observation: The main difference between the two is in the quantity of rasterization and memory control. The extra hardware in the 670 would probably be significant in some games but would be pretty much useless for FAH. (well, except maybe for the Viewer.) For me, the hardware reduction saves power and heat but probably wouldn't change FAH's productivity.

At the moment there is not a huge ppd difference between 670 and 660ti only 6.5%. Bandwidth is important for gpucompute as is L2 cache, 670 has 512mb - 660ti 384mb. Due to kepler not being optimized in core 15 we are only using 75% of the gpus resources. If 100% resources are used the 670s extra cache and memory bandwidth will come into play more and we will see a larger performance gap. Due to gpu work units migrating to qrb a 20@% reduction in tpf by the 670 will lead to significantly higher ppd.

[csvanefalk]

Will we be seeing some updates here following the new benchmark scheme?

[compdewd]

Do you mean simply with regards to PPD, csvanefalk? Short answer, yes. Long answer, yes, but I have not been able to keep up with the recent news here about many things, one of which being the new QRB for GPUs, so I would have to figure out the formula for the QRB (which I think I read may be the same as the formula used for SMP?) in order to convert the old data to be able to match that of new incoming data. I certainly have some reading to catch up on before I can give you a solid answer though.

If you mean with more than a PPD update, then that answer is probably also yes. I am in the process of choosing a domain name for a website that will have the data stored in a MySQL Database and will allow some user filtering of the data via PHP In fact I already have it running on a free hosting site, however I have not disclosed that site due to the fact that COMODO DNS Servers flag the hosting site as malicious and I don't want to scare anyone off by directing them to that site.

So yes, some updates should be coming in a week or two, once college lets out

[P5-133XL]

Just a side note, the QRB part of the new benchmarking system was suspended. However, the GPU WU's are still going to be benchmarked against the SMP machine, See: http://folding.typepad.com/news/2012/12 ... nted-.html

[csvanefalk]

Great stuff Compdewd! Looking forward to it

By the way, moderators: doesn't this thread deserve a sticky or the like? After all, having a central repository for data about GPU:s would be an important and useful tool for folders and enthusiasts looking to make a wise investment with regards to the project.