CPU x86 FLOPS = GPU x86 FLOPS?

[Jonathan]

Whenever there is the question "What is faster: GPU or CPU?" the response to this has always been "they are different, one is slow but gives alot of results, the other fast but little."

The reason I don't like this answer is because it is just avoiding the question, there IS one that is faster. I don't know alot of computers, but is x86 FLOPS seem like the comparison unit of measurement.

So with the GTX 560 Ti being a whopping 1263.4 GFLOPS, and the i7-2600 being ~ 50 GFLOPS, the obvious answer to me is the GTX 560.

Note I do not care about ppd, I care about the actual output of what I am getting (as I will be making a computer that maximizes FLOPS/$, compared to PPD/$).

I was thinking of something like this:

Code: Select all

Intel Core i7-2600 = $300
ASUS P8P67 WS REVOLUTIONKingston HyperX Grey Series 8GB (2 x 4GB)  (4 PCI Express 2.0) = $260
2 X G.SKILL Ripjaws X Series 8GB (2 x 4GB)= $150
4x GeForce GTX 560 Ti = $1040
OCZ ZX Series 1250W Fully-Modular 80PLUS Gold = $240
Western Digital Caviar Blue 250GB = $39
Antec Nine Hundred Black Case: $100

Total: 2129


GFLOPS: 5100

GFLOPS/$: 2.4
$/GFLOP: 0.42

[bruce]

Welcome to foldingforum.or, Jonathan.

Your assumption that GFLOPS is a universal standard doesn't apply for several reasons, the most important being the software.

The code that runs on GPUs and PS3s is significantly limited in what it can do. The work that hardware can do requires a much simpler model of the protein. The results are important, scientifically, but often the approximations that have to be made to get results on a GPU turn out to be important enough that once the trends are established, a more precise model can be devised with fewer approximations using code that only runs on the CPU.

There's nothing wrong with your hardware choice, but you're basing it on an assumption that doesn't reflect the actual scientific results of those GFLOPS.

I think you'll find that even if you dedicate part of the i7 to sending and receiving data from the GPUs, there's still enough processing power to run SMP.

I would make two suggestions that you can consider before finalizing your choice.
1) Get 2600K. It costs more, but it's relatively easy to overclock to more than make up for the difference in cost.
2) if you pay your own electric bill, the power costs for GFLOPS on a GPU far exceed the power costs on a CPU. After a couple of years, that adds up.

Again, there's nothing wrong with your first choice.

[Racer43]

Bruce, to crystallize in my mind, would this be correct? GPU and PS3 allows a quick read if a theory is viable without a huge investment; CPU then refines the theory to a precision usable for reporting for peer review? If so, then would bigadv just be a bigger chunk of refinable data vs smp's medium chunk vs classic's smaller chunk of refinable data, most of the refinable data coming from the quick reads of the GPU/PS3 models?

[Jonathan]

Ah that makes sense. Repeatable results are key.

This makes the decision even harder now! But I guess the GPUs have a "wow" factor with the FLOP calculations... the reason I'm doing this is I'm starting a business and this is sort of my "charity", and the more flops the better it looks. While a lot of groups focus on world hunger, poverty, etc, I like to focus on science because once there is a result, it is fairly permanent.

Not that those issues are not important, but giving them food isn't going to solve the source of the problem. Woah, way off topic...

Thank for the lengthy response, more details the better, appreciate you spending that much time on a person who just joined. (I've actually been folding for 2 years now anonymously)

[Jonathan]

Also would like to point out I'm studying biology and interested in becoming a computational biologist. I know about proteins but not so much the physics of it, it astounds me that we actually know how a protein reacts and the forces upon them and actually put it into equations and code. Polar, Non-polar, Van der Waals, hydrogen bonds, etc. Simply amazing scientists figured this out... you guys are my heros... sounds corny, but really it's true.

[GreyWhiskers]

i found this post by Dr. Izaguirre quite promising - especially that the method for potentially accelerating the MD computations by 100 fold being applied to the OpenMM GPU Core - with the GFlops capability of relatively cheap Nvidia GPUs, like the 560Ti, being north of 1 TFlops! As Dr Izaguirre says things are looking good! Stay tuned for an exciting summer (in northern hemisphere).

Re: Projects 10012-10085 -> accelerate F@H 100 times?
by izaguirr » Sat Jul 02, 2011 12:27 pm

@7im: yes, you've summarized our approach well.

For those of you who are curious, by the end of the summer 2011 we hope to have characterized the performance / accuracy tradeoff of this methodology. Things are looking good. Then we hope to deploy it for select projects on an OpenMM GPU core by the end of 2011. That means that the methodology will be at first restricted to the types of simulations that can use GPUs.

A second stage would include extending the methodology to work with general simulations that can be then incorporated into core A4 (or its successor). We plan on doing this in 2012.

Even at its most successful though, different scientific projects require different levels of modeling fidelity/accuracy. Thus, while we hope that this methodology can enable projects spanning much longer scientific timescales and eventually much larger systems, don't expect the standard CPU and GPU cores to go away ever.
Prof. Jesus A. Izaguirre, Ph.D.
Department of Computer Science and Engineering
Interdisciplinary Center for Network Science and Applications
University of Notre Dame, IN, USA
Member, Folding@Home Consortium, Stanford University, CA, USA

Folding@home project descriptions

Project 10012

This project is a test of a novel means to significantly accelearate Folding@home Molecular Dynamics calculations by 100x without any additional increase in hardware. Thus, this method has great promise to push Folding@home way beyond what it can do now, although there is much work to do to test it to make sure that the results are scientifically valid and useful. You can find more information about the method by reading one of our recent papers.

MULTISCALE DYNAMICS OF MACROMOLECULES USING NORMAL MODE LANGEVIN

[bruce]

Bruce, to crystallize in my mind, would this be correct? GPU and PS3 allows a quick read if a theory is viable without a huge investment; CPU then refines the theory to a precision usable for reporting for peer review? If so, then would bigadv just be a bigger chunk of refinable data vs smp's medium chunk vs classic's smaller chunk of refinable data, most of the refinable data coming from the quick reads of the GPU/PS3 models?

I'm not qualified to give that a yes or no answer but I'm pretty sure the answer is some of both. I think that does happen sometimes. I also think that a lot of GPU results can stand on their own. Scientists don't use a single tool for all studies but choose the best one for the task at hand . . . while simultaneously inventing new tools such as the one mentioned by Dr. Izaguirre (above).

[toTOW]

Don't forget to read the Flop FAQ, some differences between Native Flops (for GPUs) and x86 Flops are detailed with examples of instructions : http://folding.stanford.edu/English/FAQ-flops

[derrickmcc]

I was thinking of something like this:

Code: Select all

Intel Core i7-2600 = $300
ASUS P8P67 WS REVOLUTIONKingston HyperX Grey Series 8GB (2 x 4GB)  (4 PCI Express 2.0) = $260
2 X G.SKILL Ripjaws X Series 8GB (2 x 4GB)= $150
4x GeForce GTX 560 Ti = $1040
OCZ ZX Series 1250W Fully-Modular 80PLUS Gold = $240
Western Digital Caviar Blue 250GB = $39
Antec Nine Hundred Black Case: $100

Can I suggest the Antec 902, it will be a bit more expensive than the Antec 900 but worth the extra.
You will need to seriously consider cooling.
I found that to fit 4 GTX 260 cards I had to remove the cooling shrouds, as the depth of each card was 43mm, so check the dimensions of cards you are intending to use.
Fitting additional 120mm internal Antec fans is an option, but you will have to check on the length of the GTX 560 cards.

[Jonathan]

I spent extra for the Antex 900 because of cooling, I thought i was generous because it is a $100 case :S

As for the depth of the case, It would be a bad day for me if i set that up and found out the 4th gpu didn't fit, i hope they have exact spec for that as it will be a pita to figure out (as where the screws are that mount the mobo, then how far the mobo 4th pcie 2.0 slot is, then how deep the gpu is, lol)

Thanks for the suggestion