Blog post: "Unified GPU/SMP benchmarking scheme ..."

[Grandpa_01]

I have been doing some benchmarking the last week or so and thought I would post the results for those that are wondering which way to go. If things stay the way they currently are it looks like GPU is the only way to go. A good example of the potential folding power of the GPU's is one team that is particularly heavy in GPU's production has almost doubled since the release of just this 1 WU when all GPU WU's are given QRB with the current benchmark scheme there PPD will effectively triple or Quadruple. As you can see from the results below GPU folding is currently the way to go, running smp with GPU is a waste of power. The 4P folding on anything besides the 8102 which is very rare it is a waste of power also. So there you have it looks like Stanford needs to port the smp and bigadv over to GPU and just git rid of them.

2700k / GTX 580 / GTX460 all folding
375,000 PPD = 690 Watts at the wall
= 543 ppd/w

2700k / GTX 580 / GTX 460 GPU folding only
359,000 PPD = 470 Watts at the wall
= 763 ppd/w

4P folding 48 cores at 3Ghz
8101 = 480,000 PPD = 1150 Watts at the wall
= 417 ppd/w

4P folding 48 cores at 3Ghz
8102 = 797,000 PPD = 1080 Watts at the wall
= 737 ppd/w

[patonb]

Just remember the qrb multi isn't set yet, and maybe they'll finally run out of the shotty 8101s.

It is sad to see the "team" taking all the beta units instead of the proper beta members.

[7im]

I will suggest that Gramps is either being reactionary, or short sighted.

Sure, BA will decline for a while, and maybe even come close to an end, but not all BA folders are ONLY points hounds. Philanthropy continues.

And what happens when the Gromacs core adds support for AVX and the SSE performance doubles? Your 750,000 PPD system suddenly does 1.5 Mil PPD and starts looking interesting again. Don't fire sale your 4P boxes just yet...

SMP? Not likely to end. With all the integrated GPUs, and every new system shipping with multiple cores, and increasing numbers of multiple cores, SMP will keep going strong as a common option. And even when there is a discreet GPU, SMP folding impacts the GPU folding so little, everyone will keep running their SMP slot, although maybe at a lower priority than their GPU slot.

And AVX support potentially kicks SMP up a notch here as well.

Finally, the spectre of OpenCL running on your CPU and GPU at the same time looms in the distant future. It's anyone's ballgame then...

[Grandpa_01]

The 8101 has been stated as the norm for bigadv not the 8102 as far as non beta members running the 8057's there is nothing wrong with it, PG wanted everything open to the public, they do not need our input (beta team) they know and see what is happening with the WU's, the beta team has pretty much been defunked since it went public. I do believe if they were going to adjust them they would have already done it or made a statement to that effect, I would believe at this time this is pretty much the way it is. It will make my folding less risky and save me some $$$ my cost for equipment and power are going to go way down.

[Grandpa_01]

I will suggest that Gramps is either being reactionary, or short sighted.

Sure, BA will decline for a while, and maybe even come close to an end, but not all BA folders are ONLY points hounds. Philanthropy continues.

And what happens when the Gromacs core adds support for AVX and the SSE performance doubles? Your 750,000 PPD system suddenly does 1.5 Mil PPD and starts looking interesting again. Don't fire sale your 4P boxes just yet...

SMP? Not likely to end. With all the integrated GPUs, and every new system shipping with multiple cores, and increasing numbers of multiple cores, SMP will keep going strong as a common option. And even when there is a discreet GPU, SMP folding impacts the GPU folding so little, everyone will keep running their SMP slot, although maybe at a lower priority than their GPU slot.

And AVX support potentially kicks SMP up a notch here as well.

Finally, the spectre of OpenCL running on your CPU and GPU at the same time looms in the distant future. It's anyone's ballgame then...

7Im I am not selling my 4P's I will just retire them and start my Gulfies back up with GPU only folding that is 12 GPU's when and if PG gets AVX and SSE working for them then the GPU's will go in them, Just righ now the most production per $$$ spent is GPU only.

[Zagen30]

I don't doubt that this will change the spending habits of the affluent contributors, but they aren't the majority. I'm guessing that most contributors aren't able to spend lots of money on a regular basis to upgrade their boxes and/or build dedicated folding rigs, and therefore fold with what they have. They are definitely going to have a CPU; whether they have a folding-capable GPU is much less certain, let alone a relatively powerful one. If you look at the client counts, 88% of active clients are x86 CPUs. Unless I'm wrong on the average donor's finances, I don't foresee something like an 80% exodus there to GPUs. Also, and I'm speculating here, wouldn't the fact that CPUs and GPUs can now do the same work indicate that it's relatively easy to generate the work for either core? If it doesn't cost them much time or effort to put out CPU work, I don't see why they'd drop it, even if there were some major shift among the rank-and-file to GPUs.

[mdk777]

Last number I saw was 15 million discrete graphics cards sold :

LAST QUARTER

While many people don't have the newest or most powerful, SLI, or quad SLI.....
there are a shed load of single graphic card computers out there

Recently, people stopped GPU to maximize SMP.

That really wasn't the best choice for maximizing the total work that could be achieved.

Perhaps giving up entirely on SMP will be a pendulum swing too far the other direction.

However, looking at the TOTAL FLOPS chart...keeping the pendulum moving is certainly a consideration.

[kiore]

Although I think it is possible that new builders may shy away from committing to new multi cpu dedicated rigs for BA I am not certain that all who have current capable ones will retire them so easily, however adding GPUs has always been the easiest upgrade path so many people buying new hardware may just do this. Personally, and I don't claim to be representative in any sense, I have just bought the components for a new rig knowing these changes are underway and have still bought a i7 3820 as my cpu and will smp fold it alongside two fermi gpus. This will be my general purpose system and should be a great folder as well. I have seen the results for the new trial of QRBs for GPU and as a long term GPU folder I welcome them, I am again not so certain that the highest of the current PPD reported will be the standard, we have seen in the past initial work units bringing high returns only to have them become harder and harder to score easy points on as they become more complex and the subsequent complaints that the scores are not as good or the units run hotter etc and even as with the BA they become far too big for the systems that could initially run them. Already we have seen a significant drop off in points awarded when using the 2.25 core over the 2.22 with in my case seeing my GTX 570 significantly outscored by older and theoretically less capable cards like GTX 460's using the older cores. I very much doubt that this is set in stone yet, but equal points for equal work is hard to dispute if modern fast GPUs can now produce the results that the project requires well good.. Results are the main objective after all and the points competition is fun and useful as a driver, but they really do need to be awarded as the project requires to be an effective currency. I do understand peoples annoyance if they have invested money in a specific system that they now feel does not reward them proportionally as much as it has, but points are not being reduced, they are being expanded to compensate and encourage participation with hardware that the project desires to have access to.
In the years I have been involved in the project I have seen these swings and roundabouts, the pendulum seems to be swinging back towards GPU rigs being the hot thing again, but exciting new changes seem likely in increasing productivity and it is impossible to say which way the pendulum with swing next, in favour of GPUs. SMPs or what I (wild crazy guessing opinion) anticipate combinations of both.
Foldon.

[csvanefalk]

Hello, I am a Linux user, I cannot run GPU folding. Please remember me the next time you mix yourself a blueberry smoothie.

[kiore]

Hello, I am a Linux user, I cannot run GPU folding. Please remember me the next time you mix yourself a blueberry smoothie.

This would be a huge step forward to have GPUs work in Linux easily, rather than trying to do wine wrappers etc, I would certainly change my OS on my old MSI k9a2 gpu rig from winxp to linux if this occurred. I do understand that most of those gaming gpu's out there that could be recruited are on M$ flavours and so that is where the easy recruiting can be done, but a Mac/Linux GPU function is very desirable IMO.

[csvanefalk]

Yep, I am very much looking forward to it, especially if the changes to marks for GPUs will be this drastic. For all we know, GPUs might just work even better under Linux, just like SMP does.

[Punchy]

To misquote Mark Twain, reports of SMP's death are greatly exaggerated. It's too early to make any predictions until the benchmarking is finalized.

Oh, and you didn't include price in the equations, so...

2P E5 folding 32 "cores" at 2.9Ghz
8101 = 230,000 PPD = 300 Watts at the wall
= 766 ppd/w

Oh, and at the blueberry smoothies, but I do agree

[Grandpa_01]

If you look through the stats a very large portion of the daily production is done by enthusiast both in # of WU's and points, these folders try to optimise for PPD why would any enthusiast build for or run smp when relativity speaking you can get so much more by turning off smp and adding a cheap GPU and more efficiently use the electricity, I can currently buy 12 used 580 GTX GPU's for less than the price of 1 - 4P and get 3,000,000 PPD with them and use less than half the electricity that I am currently using. I would venture to say that bigadv folding will die rather rapidly (There just are not that many bigadv rigs and they are enthusiast ) while smp will suffer a slower death. It really should not matter since they are supposed to be able to run all types of work on GPU's. If that is not the case then I am sure Stanford will adjust the points accordingly.

I am not looking at this as a bad thing just a change I have gone from AMD CPU to ATI GPU to Nvidia GPU back to Intel CPU to AMD MP CPU and looks like back to nVidia GPU in my folding career, Things change that is the nature of the beast people will follow the carrot that is just the way it is. Stanford appears to want the nVidia carrot right now so that is what will happen who knows about tomorrow or the next day thing will change with hardware advancements.

[Jesse_V]

To misquote Mark Twain, reports of SMP's death are greatly exaggerated. It's too early to make any predictions until the benchmarking is finalized. ...

Exactly. This entire thread rests on the premise that p8057's PPD is relatively finalized. Maybe it is, maybe it isn't. Who knows. The beta team members are reporting the PPD as they see it. I'm sure the PG is smart enough to figure out the final PPD based on the actual science performance and it's importance relative to the other platforms. Sure, I like to maximize PPD too, but every little bit helps. Fold with what you have, and if one wants to make hardware changes based on the relative PPD, it's probably a safer bet to first check to see if the PPD numbers will remain consistent. Right now, at least for the GPU, it's probably not.

[Grandpa_01]

To misquote Mark Twain, reports of SMP's death are greatly exaggerated. It's too early to make any predictions until the benchmarking is finalized.

Oh, and you didn't include price in the equations, so...

2P E5 folding 32 "cores" at 2.9Ghz
8101 = 230,000 PPD = 300 Watts at the wall
= 766 ppd/w

Oh, and at the blueberry smoothies, but I do agree

Punchy
i7 2700k 2 - GTX 580 no smp
8057 = 500,000 PPD = 600 watts at the wall (which should be accurate) I do not have a second GTX 580 yet but soon
= 833 ppd/w