Is GPU folding more valuable than CPU folding scientifically

[donkom]

This is a question that has been on my mind for some time. I am folding on a powerful CPU as well as a powerful GPU (2x E5-2697v2 and R9 290X respectively), and within a windows environment the two folding slots pull in roughly the same amount of points. I understand that points are based on benchmark hardware, etc... and I understand that each slot processes different work units with different cores, etc.

The question is this: given equal points, is one work unit more sceintifically valuable than another? Are experiments that run on Core A4 or A5 impossible to redesign to run on Core 17? Is there some additional computational wizardry that a CPU lends to Folding@Home that it cannot do without?

Statistically, GPU horsepower seems to greatly overshadow CPU performance ( http://fah-web.stanford.edu/cgi-bin/mai ... e=osstats2 ), and with GPU performance-per-dollar increasing exponentially over CPUs, will we see a future where F@H only runs on graphics cards? Or will CPU calculations still be necessary to simulate molecular dynamics that are not computable any other way?

[PantherX]

Saying that Project A is "more scientifically valuable" than Project B depends on your perspective. If, for example, Project A deals with Lung Cancer while Project B deals with "understanding biochemical reactions within a cell", it may seem that Project A wins for someone who may have a personal experience with Lung Cancer. However, a biochemist would value Project B over Project A while a cancer researcher would value Project A over B. Right now, we may see these two projects as different but maybe, later in the future, a discovery might be made which relates Project B to Project A. Thus, IMO, all Projects are equally scientifically valuable since they attempt to increase our understanding. Do remember that in experimentation, a negative result can be just as important as a positive result. Having said that, for the donors, there is a planned feature where the donor may choose to contribute more towards a single disease (it may consist of different Projects which relates to that disease). This has been implemented in the F@H Clients but not yet on the F@H Servers (no time frame given as to when it will be implemented).

Regarding the CPU VS GPU, FahCore_17 now allows explicit modelling (treat water atoms are individuals) on GPUs. This was previously reserved only for CPUs since GPUs could only run implicit modelling (treat water as a continuous medium) with the older FahCores (FahCore_11, FahCore_15 and FahCore_16). If you look at it from the ns/day scale, GPUs are significantly faster than the CPUs with the current generation of hardware.

I highly doubt that F@H will ever drop CPU support since an average computer will always have a CPU while a GPU can be optional. Do note that F@H is primarily aimed towards home users running average systems. However, F@H enthusiasts may choose to contribute using the best metrics, be it PPD/watts, PPD/price, etc. It is optional and is prone to changes in the hardware released by AMD, Intel and Nvidia.

[7im]

Short answer: No.

Points are tied to scientific value, so equal points means equal scientific value. That is the whole basis of the points system, equal pay for equal work. Personal opinion has nothing to do with it.

[donkom]

Thanks very much for the thoughful response PantherX! I complete agree that the sceintific value of individual simulations are equal (or at least impossible to compare), but the question was more specific to the platform the simulations are running on.

If I understand you correctly, a CPU has no advantage over a GPU for running a simulation? Comparing Core A5 to Core 17, is there any advantage to designing a project to run on A5? Better accuracy, more options for variables, etc... is a CPU capable of generating results that a GPU is not able to create?

[Jesse_V]

It's easier to design and run code that works on a CPU rather than GPUs. GPUs are really tricky things. So in that respect, the researcher may end up spending less time resolving issues with CPU work rather than GPU work. This is becoming less of an issue now that core 17 is now fairly stable and runs on many platforms and architectures. Results wise, I don't think a CPU has any significant advantages.

[donkom]

Thanks for clearing it up 7im and Jesse_V!

[codysluder]

As a general rule the gpus that people have are computationally more powerful than the cpus that people have. Since faster is better and the gpu software now is just as capable as the cpu software has always been, gpus are preferred. That's a general statement, though and it may not apply to your cpu and your gpu.

[7im]

You can spend $5k and get a single bigadv box that makes 800 k PPD, or you can spend the same on a quad Titan box and get the same PPD.

Performance is a function of price.

[gimpy]

7im said:Performance is a function of price.

This is a universal truth.It should be made into a scientific law on the same importance as Newton or Einstein. Imho. Driver: "how fast can we go?". Engineer/mechanic "How much money do you have?". Replace driver and engineer with whoever is appropo/relevent to you needs and it's always true.

[7im]

Yes, but not always. Technological advances tend to break this "law". When JIT hits NV drivers for FAH, the GPU folding speed doubles, at no additional cost. Then again, if folding on AVX comes to FAH, then CPUs double in speed at no additional cost.

So again, depending on the swing of the performance pendulum, the answer to this thread will change from one day to the next. Today, CPU and GPU are about even. Tomorrow may be a different answer.

[PlayLoud]

Short answer: No.

Points are tied to scientific value, so equal points means equal scientific value. That is the whole basis of the points system, equal pay for equal work. Personal opinion has nothing to do with it.

Has this really been stated by the PG? Just looking at the Client Statistic page, it would seem GPUs are so many times faster than CPUs, yet the PPD differences don't seem to scale. If my GTX 760 can get 2x the PPD as my 3770k, does that mean my 760 is only 2x as fast?I thought the difference was greater than that. Even for a GPU that isn't top of the line. If they are both doing the same type of work (both explicit, with no drawbacks to the GPU), wouldn't PPD be roughly tied to FLOP count (assuming a 24/7 folder, so QRB should be similar).

[P5-133XL]

yes, PG has stated that scientific value is their goal when assigning points i.e. PPD. Theoretically, more PPD means more scientific value.

As a second point, with Core_17, GPU PPD values are assigned by running the same WU on the SMP benchmark machine. So theoretically there shouldn't be a disconnect between the CPU and GPU clients. Older cores such as the 11, 15, 16 were benchmarked against a GTX 460 so it was possible to have that disconnect.

[7im]

Short answer: No.

Points are tied to scientific value, so equal points means equal scientific value. That is the whole basis of the points system, equal pay for equal work. Personal opinion has nothing to do with it.

Has this really been stated by the PG? Just looking at the Client Statistic page, it would seem GPUs are so many times faster than CPUs, yet the PPD differences don't seem to scale. If my GTX 760 can get 2x the PPD as my 3770k, does that mean my 760 is only 2x as fast?I thought the difference was greater than that. Even for a GPU that isn't top of the line. If they are both doing the same type of work (both explicit, with no drawbacks to the GPU), wouldn't PPD be roughly tied to FLOP count (assuming a 24/7 folder, so QRB should be similar).

Yes, stated by Pande Group many times. Simply search the blog or forum posts by Dr. Pande.

You are making some bad assumptions about points. Vs performance, so please start with some basics, like the points FAQ.

[PlayLoud]

yes, PG has stated that scientific value is their goal when assigning points i.e. PPD. Theoretically, more PPD means more scientific value.

As a second point, with Core_17, GPU PPD values are assigned by running the same WU on the SMP benchmark machine. So theoretically there shouldn't be a disconnect between the CPU and GPU clients. Older cores such as the 11, 15, 16 were benchmarked against a GTX 460 so it was possible to have that disconnect.

Thanks for the info! I really haven't been following for a while. When I last frequented these forums, GPUs could only do implicit, so I knew my CPU folding had a special value.

What you said about the Core_17 benchmarking makes sense. If they are benching the same WU, it is very easy to base it on scientific value.

Yes, stated by Pande Group many times. Simply search the blog or forum posts by Dr. Pande.

You are making some bad assumptions about points. Vs performance, so please start with some basics, like the points FAQ.

Sorry that I came across so newbish. I have been away for a while. Searching on this forum for PPD and GPU vs CPU issues is very tough, since there are so many topics, and usually there are too many results to even display. It's hard to find exactly what you are looking for.
P5-133XL's answer was perfect. I have a better understanding now.

[Grandpa_01]

Yes, but not always. Technological advances tend to break this "law". When JIT hits NV drivers for FAH, the GPU folding speed doubles, at no additional cost. Then again, if folding on AVX comes to FAH, then CPUs double in speed at no additional cost.

So again, depending on the swing of the performance pendulum, the answer to this thread will change from one day to the next. Today, CPU and GPU are about even. Tomorrow may be a different answer.

More like tripple, the projects I run that use AVX generrally run about 3 time as fast as when AVX is not used on the same project.