Buy AMD or NVidia GPU for PPD?

[Michael_McCord,_M.D.]

With Dr. Pande blogging about the large advances for AMD GPU ppd due to Core 17, is that now considered the go-to GPU line for pure folding? Anybody have hard ppd data? Would appreciate input. Thanks.

[7im]

The jury is still out as all of the potential optimizations are not in yet.

And yes, you are missing the thread where there is hard data. http://foldingforum.org/viewtopic.php?f=38&t=24492

[bison88]

AMD's future looks more promising than Nvidia just because Nvidia is heavily behind in proper drivers (OpenCL 1.2) and a JIT compiler. Beta units make AMD look like a rockstar and Nvidia look like a chump, for the price > points.

If you plan on folding Core_15 WU's then Nvidia hands down wins and destroys.
If you plan on folding Core_17 WU's (Beta/Advanced) then AMD is where it's at, probably for the next year or so depending on how long Nvidia drags their feet. F@H was Nvidia's only real Distributed Computing performing GPU and that market is fairly big, so hopefully they get things fixed soon.

As 7im said. Things are in limbo. Really not the best of time for a video card upgrade dedicated to folding because nobody knows how its going to level out by the end of the year. Nobody is really the clear winner at this point if you're looking into a HD 7xxx or GTX 6xx/7xx series.

[bruce]

As 7im said. Things are in limbo. Really not the best of time for a video card upgrade dedicated to folding because nobody knows how its going to level out by the end of the year. Nobody is really the clear winner at this point if you're looking into a HD 7xxx or GTX 6xx/7xx series.

Things are ALWAYS in some kind of limbo. AMD and NV compete for market share and will continue to. The game market doesn't match the FAH market precisely and ATI has clearly been behind for a while in the FAH market. I expect that they'll leapfrog each other for a while now.

[7im]

More solid info: http://www.anandtech.com/show/7103/nvid ... -review/15

This is about the sweet spot for price/ppd. The single precision graph is the one best related to folding.



Like bruce said, if you bought a card today, these are good options. But it will continue to change for a while.

[VijayPande]

Note that all of this could change very easily in the future. If NV implemented JIT, we'd see a ~2x speed up there. However, ATI might be able to make their own advances in their OpenCL driver. It's very hard to predict. The best we can do is to look at how FAHbench is doing on existing hardware.

[SASinUtah]

While on the subject of GPUs, for I also have been waiting to upgrade my old 6570 until the Core_17 was out of beta testing. In the past, an AMD card required a full cpu core for Core_16, and I have read that NVIDIA cards now need a full core for Core_17 (an NVIDIA issue and not a F@H issue). Does an AMD card still require a full cpu core for Core_17 as it did for Core_16?

[7im]

A full cpu core is needed for both NV and AMD in core_17. NV more so than AMD, but that could change just as easily as the drivers situation.

[Stonecold]

I'm certainly no expert on the subject, but I'd buy an Nvidia GPU if I were you. I foresee Nvidia having, in the long run, more PPD. From what I hear, Nvidia has fewer cores than AMD but each core is more geared towards GPGPU activities (like Folding@home). I think this is also why AMD > Nvidia at password cracking (e.g. oclHashcat) and Bitcoin mining, but the opposite.

Someone correct me if I'm wrong of course.

[3D_Now!!]

It is my experience with core 17 that it does not use any CPU at all (around 0.2%). It uses a whole core for the first minute or two while the GPU is loaded (I assume) but then the WU runs without further overhead. I am running 2 AMD 7790's with around 75k ppd for the pair. I am also running rosetta@home on the CPU (i5 4670k). I am yet to test if leaving the CPU unloaded will affect the gpu ppd. However, I am impressed with the general performance of core 17....good job f@h.

[bruce]

The internal architecture of AMD and NVidia is quite different, but when you view them from the perspective of the OpenCL API, they appear to be quite similar. So much depends on how well the drivers can map OpenCL calls to efficient use of the hardware -- and that means that future driver improvements are very, very important.

NVidia sells more hardware by having people write to the CUDA API than the OpenCL API so it's in their best interest to develop CUDA

[k1wi]

AMD sells more hardware by having people write to the CUDA API than the OpenCL API so it's in their best interest to develop CUDA

Did you mean NVidia bruce?

[bruce]

AMD sells more hardware by having people write to the CUDA API than the OpenCL API so it's in their best interest to develop CUDA

Did you mean NVidia bruce?

Of course. Sorry.
Fixed

[MaddMutt]

I'm just tossing my two cent's worth in on this matter :

When using the Sapphire Toxic HD-6950 ( Factory overclocked to HD-6970 by BIOS switch) with the new Core_17, it still uses a Full Core. Estimated ETA is 12 hrs @ 945 core and 1350 memory, Estimated PPD is < 40K. This also applies to the Sapphire HD-6970 @ 945 Core and 1455 Memory.

When using the Gigabyte WF-3 HD-7950 ( Overclocked to 1200 Core and 1385 Memory) with the new Core_17, It uses all cores to Download and Prep the 79XX series. This lasts for about 1-minute or until the core and card are properly prepped. Then the GPU takes over and the CPU usage drops to less than 2% for the rest of the folding process.

How much Superior is the 79xx over the 69xx when using the new Core_17 :
Using 2x HD-7950's stock @ 1000 Core and 1250 Memory ( Version 1 with no Voltage Lock)

EDIT !!!!!! This is using WU 8900 Core_0x17 Zeta Version 2 Beta, AMD 13.8 Version 2 Beta Drivers and MSI Afterburner Version 3.0.0 Beta 11.
The BUG In The Software, is that the PPD is calculated for how much work is done in 24 HOURS.
My HS Edubacation, stated that there are only (4) FOUR (6) SIX Hour groups in 24 Hours. My Times Below do not add up to this, if I start them @ Exacly 0000 - 2359. I lose 27K - 30K per card per day if counted like this.

Gigabyte WF-3 HD-7950 @ 1200 Core and 1385 Memory : ETA = 6 Hrs and 10 Min - TPF = 3 Min and 47 Sec - PPD = 112~K
Gigabyte WF-3 HD-7950 @ 1145 Core and 1325 Memory : ETA = 6 Hrs and 30 Min - TPF = 4 Min and 5 Sec - PPD = 101~K
This gives an Estimated Total - PPD = 213~K

Thank You For Your Time Reading This ( Folding for MPC)

MaddMutt

[7im]

So a only 18 seconds per frame makes an 11,000 PPD difference. Interesting. My GT 430 takes 2 days to make that many points.

Helps explain some of the larger variances in this charting... http://folding.typepad.com/news/2013/07 ... rking.html