GPU/PS3 folding better than CPU folding?

[Jacko87]

I was just wondering if GPU folding is better for "the cause" than CPU folding, points aside. For example, lets say a GPU/PS3 and a CPU can both fold one WU in 24 hours, both get the same amount of points, but did one actually do more calculations and issue the same points for fairness to other clients? I'm asking because someone told that while a GPU may get similar points for folding, it actually does more calculations because of its architecture, and I want to know if its true or not. Thanks for any info!

[bruce]

I was just wondering if GPU folding is better for "the cause" than CPU folding, points aside. For example, lets say a GPU/PS3 and a CPU can both fold one WU in 24 hours, both get the same amount of points, but did one actually do more calculations and issue the same points for fairness to other clients? I'm asking because someone told that while a GPU may get similar points for folding, it actually does more calculations because of its architecture, and I want to know if its true or not. Thanks for any info!

You've asked a good question, and it's not quite the one most people have asked.

First, the points are supposed to represent scientific value. There are some known inaccuracies in the current system and Vijay Pande has mentioned that there will be a new system one of these days -- presumably one which will reduce the inaccuracies -- but we still need to deal with the problem of defining what "better for the cause" means. If we assume that the Pande Group is entirely responsible for the points system, and whatever gets more points is truly worth more to them, then your question no longer has an answer.

Q1) Is GPU folding better for "the cause" rather than CPU folding if both fold one WU in 24 hours and are the same value to science?
A1) It doesn't matter.

We do know that faster submittals are better than slower submittals, even if the points do not change, but you took that into account when you said 1 WU per 24 hours.

We also know that the WUs assigned to the CPU client or the GPU client or the PS3 client or the SMP client are different. On the average, you'd have to have a pretty fast CPU and a pretty slow GPU to find a WU for each of them that would earn the same number of points per day, but maybe you're thinking of running them different numbers of hours per day so that the points work out to be equivalent.

If somehow you got a fast enough CPU to keep up with a slow GPU in terms of PPD, AND if the same protein happened to be assigned to both, it's true that the GPU would perform more calculations. The GPU hardware is very efficient and performing lots of calculations in parallel and it's less efficient in simple things like memory access and branching so when you optimize the throughput, it's sometimes more efficient to perform extra calculations in order to avoid memory access or branching, as compared to equivalent processing on the CPU where memory access and branching are quicker but extra calculations are slower.

[Insidious]

I know this isn't a simple question and depends on many things....

I didn't really get a good feel (my fault, not yours) and have a similar curiosity.

I hope you'll let me ask a different (but similar) question:

If the Scientists who use the results of our work had their choice..... which client would we use (assuming we had the latest and greatest hardware required to run our choice of client).

[Obcy_from_Poznan]

I was just wondering if GPU folding is better for "the cause" than CPU folding, points aside.

Hi!
According to the information written in this report, calculations using GPUs are much more valuable for the Folding@home project.:
http://arxiv.org/PS_cache/arxiv/pdf/070 ... 3060v1.pdf
Quotation:
" Compared to the donations of CPUs from over 150,000 Windows computers currently producing 145 TFlops, we have 550 GPUs donated to the project producing over 34 TFlops. Thus each GPU is providing roughly 60 times the performance of the average donated x86 CPU.
6 Conclusion
We have successfully taken advantage of the processing power available on GPUs to accelerate pairwise force calculations for several commonly used force models in stellar and molecular dynamics simulations. In some cases the GPU is more than 25 times as fast as a highly optimized SSE-based CPU implementation and exceeds the performance of custom processors specifically designed for these tasks such as GRAPE-6A. Furthermore, our performance is compute bound, so we are well poised to take advantage of further increases in the number of ALUs on GPUs, even if memory subsystem speeds do not increase significantly. Because GPUs are mass produced, they are relatively inexpensive and their performance to cost ratio is an order of magnitude better than the alternatives."

Taking this info under account, the credit points for donating one's GPU should be at least 3 times higher than they are now. This can be also justified by the fact that apart from an active GPU a person participating in the project also donates a 100% activity of a CPU at the same time! (CPU has to feed GPU with data etc.) which means maximally increased power consumption for the CPU as well as GPU.
For excample, 333 points for 1 WU of project 2725 (GPU) is much too little taking under account that both graphics and processor work at 100% of their cycles.

[bruce]

Taking this info under account, the credit points for donating one's GPU should be at least 3 times higher than they are now. This can be also justified by the fact that apart from an active GPU a person participating in the project also donates a 100% activity of a CPU at the same time! (CPU has to feed GPU with data etc.) which means maximally increased power consumption for the CPU as well as GPU.
For excample, 333 points for 1 WU of project 2725 (GPU) is much too little taking under account that both graphics and processor work at 100% of their cycles.

Your conclusions are only partially correct.

The GPU is fast but it lacks versatility. The SMP version of FAH is much more versatile and it's providing some very impressive speed gains over the classic CPU client. Contributions from the classic CPU client are also important. The various projects will continue to be assigned to each platforms, depending on which is best for that project.

As far as points are concerned, we know Vijay has announced that some changes will be made to the points system. We don't know what those changes will be. Perhaps the GPU will be increased "at least 3 times higher" and perhaps not. We'll have to wait an see.

[Insidious]

.
.
.
If the Scientists who use the results of our work had their choice..... which client would we use (assuming we had the latest and greatest hardware required to run our choice of client).

So is it a secret or what?

I mean, I appreciate all the points and process based speculation, but won't someone representing the scientists please just tell me what they would like? How hard is that?

-Sid

[Ivoshiee]

.
.
.
If the Scientists who use the results of our work had their choice..... which client would we use (assuming we had the latest and greatest hardware required to run our choice of client).

So is it a secret or what?

I mean, I appreciate all the points and process based speculation, but won't someone representing the scientists please just tell me what they would like? How hard is that?

-Sid

Have you heard this interview?:
http://folding.typepad.com/news/2007/12 ... in-bi.html

At the moment they are after multi core systems running at top speeds. These multi cores can be CPU cores (4 core CPUs) or fast GPUs (new GPU client is supposed support latest ATI cards).

[bruce]

I mean, I appreciate all the points and process based speculation, but won't someone representing the scientists please just tell me what they would like? How hard is that?

The answer is BOTH.

They will use whatever resources that people will donate and put them to good use. If you're planning on buying more hardware, select whichever one you'll get the most use out of for other purposes.

I'm planning to buy a second PS3. I'm not a gamer, myself, but my first one got sent to someone who is using it for both games and folding. My second one will be a blu-ray player in addition to folding. When I have time, I'm going to build a new PC with better hardwre than the ones I have now. They'll all be folding.

[Insidious]

Thanks for the link to the interview, I'm looking forward to listening to it in it's entirety.

So far, we've narrowed it down to the SMP, the GPU or the PS3 (fast multicore systems)

I guess I'm more interested in which ONE of these the scientists would choose me to use if I am willing to run only one of them with 'latest and greatest' hardware.

-Sid

Bruce, I think we were typing at the same time.

I know the project benefits from all of them. I am not trying to minimize the usefulness of any of the various ways we fold.

These are great posts and I thank you for them.... good answers..... just not to the question that was posed. It really is a pretty straightforward question. Is the answer a secret?

-Sid

[Ivoshiee]

Thanks for the link to the interview, I'm looking forward to listening to it in it's entirety.

So far, we've narrowed it down to the SMP, the GPU or the PS3 (fast multicore systems)

I guess I'm more interested in which ONE of these the scientists would choose me to use if I am willing to run only one of them with 'latest and greatest' hardware.

-Sid

There is no single platform all Pande Group scientist are targeting. If you check the project table (http://fah-web.stanford.edu/psummary.html) then it is obvious (compare "Contact" and "Code" fields) that some of these scientist are PS3 "supporters", some are GPU etc.