Folding Forum

Grandpa_01 wrote:I am not at all sure GPU's can do bigadv so far the largest is quite a bit shy of a bigadv WU and takes several hrs to complete. There have not been many core 17 projects released and there have been several smp projects released since the core 17 WU's were released, so that in itself makes me question the validity of GPU being capable of doing the same work at this point in time.

Allot of people are choosing either to do GPU work at this time or to build a bigadv capable machine. The reason is simple ( cost vs reward ) I tried to explain (= pay for equal work) before and it got turned into (= sience done for = pay)

Well that does not work it needs to be an = pay for = work system, say it takes (x amount of energy to do x task = y value) if a cpu takes 200 watts and cost x to run 24 hrs and get's 30K points and a GPU takes 200 watts and cost x to run 24 hrs and get's 80k which one are you going to run.

Now enplane to Joe Cpu why you want him to work for 24 hrs and receive 50k less that Joe Gpu for working the same amount of time and using the same amount of energy.

That is = pay for = work

Many have chose not to run cpu because there is no good or logical answer to that question.

DING DING DING we have a winner!!!!!!!!!!!!
All PG has to do is ask the folders to help with the SMP backlog and just bump the QRB a couple multiples 3x to 5x. or what ever it is.
It must be true GPUs can NOT do SMP, bigadv-smp they would have ported them over.
So CPUs are needed there are hundreds of cpus idle with just a thread or two feeding a GPU. I always hear it is not worth it to run them, PPD/W wise.
So make it worth their while to run them and ask the bigadv folders to help. I am sure many would if you ask them, and not push then into it.
If the 8 core desktop are being released soon 8+8=16 go from 16 to 24, then they will have to cheat and spoof core count.

bruce wrote:So they're saying that if FAH just changed 1 points to 10 points (and 10K to 100K) for all projects and for all points earned in the past, people would leave other DC projects and flock to FAH?

I don't think so.

Folding@Home gives zero BOINC-credits (or rather Cobblestones), so multiplying with 10 or for that matter 1 million still gives zero BOINC-credits. Sorry wasn't clear before, but the project-jumpers to whatever is the highest-paying project is between BOINC-projects, either to get to #1 in the Users combined BOINC-stats and/or to get their team to #1 in the Team combined BOINC-stats.

So, until FAH releases a BOINC-application to open beta-testing, it's unlikely you'll get any benefit from your suggested point-change.

Note, since DiRT has less than 1 % of the active BOINC-users, the wast majority of users doesn't care about a project giving significantly higher credit than their choosen project(s). So even if FAH does release a BOINC-application any extra boost due to higher crediting will probably not give a huge boost since FAH is a fairly large project.

bruce wrote:So you're comparing three parts of FAH: SMP, GPU and bigadv. You're not comparing FAH to BOINC (or some other DC project) [It sounded like somebody suggested the heavy hitters would flock from FAH to some project X that awards more PPD.

A few does try to compare across projects, but this is completely futile between FAH and BOINC since BOINC gives equal credit for same wu regardless of platform used to crunch the wu(*).

While directly comparing between BOINC-projects isn't possible, as long as limits to cpu-only, if your intel-i5-dual-core gets N credits/day in project X and your intel-i7-8-core gets 4N credits/day in project X, chances are if the dual-core gets M credits/day in project Y the 8-core will get roughly 4M credits/day in project Y.

bruce wrote:How you suggest Stanford resolve the disparity between those three parts of FAH? Arbitrarily inflating the points for one part is probably not the answer (FAH has a pretty high inflation rate rather than maintaining the former value of a point.) An anti-inflationary action like reducing the pay of the nation's CEOs is not going to be popular either.

Grandpa_01 wrote:Allot of people are choosing either to do GPU work at this time or to build a bigadv capable machine. The reason is simple ( cost vs reward ) I tried to explain (= pay for equal work) before and it got turned into (= sience done for = pay)

Well that does not work it needs to be an = pay for = work system, say it takes (x amount of energy to do x task = y value) if a cpu takes 200 watts and cost x to run 24 hrs and get's 30K points and a GPU takes 200 watts and cost x to run 24 hrs and get's 80k which one are you going to run.

Now enplane to Joe Cpu why you want him to work for 24 hrs and receive 50k less that Joe Gpu for working the same amount of time and using the same amount of energy.

That is = pay for = work

Many have chose not to run cpu because there is no good or logical answer to that question.

Answering both, under BOINC as already mentioned it doesn't matter if a wu is crunched by a GPU using 10 minutes on the wu, a cpu using 1 hour on the wu or an Android using 24 hours on the wu, everyone will get N credit for this. When it comes to credit/day on the other hand, the cpu will obviously get 24x more credit/day than the Android and the GPU will get 6x more credit/day than the cpu but for most users this isn't a big problem.

Then "Help Cure Cancer" (HCC) got their GPU-version, a single GPU could get roughly 15x more credit/day than a cpu, but I only remember one user complaining about GPU-version being much faster on the project-forum, and AFAIK the non-GPU-projects didn't get a large decrease in production due to users only running HCC...

For FAH, maybe I'm a little bit off but my i7-2700K gets roughly 15k PPD running SMP-7 and the Amd-7950 gets roughly 70k PPD so a difference of roughly 5x more for the GPU. With a "bleeding-edge" GPU the advantage is... hmm, is it 10x or something for a single-GPU-card compared to i7-2700K?

Since FAH now can run same wu's on cpu and gpu, it's fairly simple to fix the point-system, just switch-back to using a benchmark-computer for deciding the points and don't screw-up the point-system with all the various bonuses.

Bill1024 wrote:DING DING DING we have a winner!!!!!!!!!!!!
All PG has to do is ask the folders to help with the SMP backlog and just bump the QRB a couple multiples 3x to 5x. or what ever it is.

Ah, you basically means a message like this: "Thanks for running FAH in the past, we need your help again and to this end we've just made your previous contribution nearly worthless"...

... and I should increase my FAH-contribution because...


(*): GPUGRID reportedly does give some kind of QRB, but since this project is Nvidia-only I've never looked more into this so no idea how large the eventual bonus is.

Grandpa_01 wrote:I am not at all sure GPU's can do bigadv so far the largest is quite a bit shy of a bigadv WU and takes several hrs to complete. There have not been many core 17 projects released and there have been several smp projects released since the core 17 WU's were released, so that in itself makes me question the validity of GPU being capable of doing the same work at this point in time.

Allot of people are choosing either to do GPU work at this time or to build a bigadv capable machine. The reason is simple ( cost vs reward ) I tried to explain (= pay for equal work) before and it got turned into (= sience done for = pay)

Well that does not work it needs to be an = pay for = work system, say it takes (x amount of energy to do x task = y value) if a cpu takes 200 watts and cost x to run 24 hrs and get's 30K points and a GPU takes 200 watts and cost x to run 24 hrs and get's 80k which one are you going to run.

Now enplane to Joe Cpu why you want him to work for 24 hrs and receive 50k less that Joe Gpu for working the same amount of time and using the same amount of energy.

That is = pay for = work

Many have chose not to run cpu because there is no good or logical answer to that question.

Well said, my sentiments entirely

And some of you are starting to touch on the current problem with equal work for equal pay. While it may be true that GPUs can do the same science as CPUs, the platforms are not quite equal yet.

If you give GPUs 5x-10x the points, many people will not invest in CPU hardware and reduce/stop running what they have. This is not an unreasonable thing to do. The only way we have to determine the scientific value PG places on each WU is the point value. Therefore if my GPU is getting 5x the points at the same cost (electricity), then I will logically assume the GPU is getting 5x the science per (my) dollar. Folding uses a lot of electricity and I (and many others) try to do the most science for the least electricity.

Now if the platforms were completely equal this would be a good thing for PG. And I think the platforms are moving together, and probably technically are already there, with one major limitation: existing projects. As far as I have heard, it is very difficult (if not practically impossible) to port over existing projects/results to the Core 17 platform. Starting up new isn't bad, but if you have everything setup for the CPU cores or Core 15 then it will be way more effort than it is worth to switch. I mention Core 15 as this is also effecting GPUs. As many NVidia owners have noticed, there are now plenty more Core 15 projects than Core 17 even though Core 15 is supposed to be phasing out. In fact, there has even been a new Core 15 project while we haven't had a new Core 17 in months. Now, something exciting may be happening behind the scenes (but that is a different topic ), but this does illustrate that the old Cores still have higher scientific value than the newer Cores to researchers trying to finish existing projects.

So, where is all this rant that only 'sort of' applies to the topic going? If PG wants to get more SMP running, they need to shift some of the points there and clearly tell us why. If they said "Look, we really need to get more of these SMP units done. We are going to boost the points to show there importance. Don't go buy a bunch of hardware as this is probably temporary...". Everyone who has a GPU(s) has a CPU and if they dedicated their CPU resources the SMP project could get moving much faster. Right now people assume with the low points SMP gets relative to GPUs/Bigadv that PG views them as worthless, when it sounds like the researchers don't.

Of course this could cause some problems on the other end when they decide to drop the SMP points back down to old levels when the crisis abates, but that is in the future so no need to worry about it yet
(This is somewhat like the current bigadv situation - the points were inflated relative to SMP as nobody would have run them otherwise, then as PG wants to bring the points down as the need for it (apparently) has reduced, they face backlash).

Wonder what PG's opinion is on the proposal to higher value the efforts on non-BA-CPU projects.

ChristianVirtual wrote:Wonder what PG's opinion is on the proposal to higher value the efforts on non-BA-CPU projects.

It's the same as when this same idea was proposed the last two times that BA was adjusted. The first time, BA bonus was reduced, and the second the Core threshold was raised. Neither time was SMP adjusted, so very unlikely it will change this time either.

powerarmour wrote:

Grandpa_01 wrote:I am not at all sure GPU's can do bigadv so far the largest is quite a bit shy of a bigadv WU and takes several hrs to complete. There have not been many core 17 projects released and there have been several smp projects released since the core 17 WU's were released, so that in itself makes me question the validity of GPU being capable of doing the same work at this point in time.

Allot of people are choosing either to do GPU work at this time or to build a bigadv capable machine. The reason is simple ( cost vs reward ) I tried to explain (= pay for equal work) before and it got turned into (= sience done for = pay)

Well that does not work it needs to be an = pay for = work system, say it takes (x amount of energy to do x task = y value) if a cpu takes 200 watts and cost x to run 24 hrs and get's 30K points and a GPU takes 200 watts and cost x to run 24 hrs and get's 80k which one are you going to run.

Now enplane to Joe Cpu why you want him to work for 24 hrs and receive 50k less that Joe Gpu for working the same amount of time and using the same amount of energy.

That is = pay for = work

Many have chose not to run cpu because there is no good or logical answer to that question.

Well said, my sentiments entirely

Sentiment and supposition are nice, but let's try to stick with facts.

Fact: Protoneer tabled developing the GPU BidAdv project because there was no interest in it from the current researchers. And not because it doesn't work as was intimated. Check the #fah IRC log on that one.

Fact: Researchers do NOT switch fahcore types in the middle of their research project. That explains two things, the above fact, and also why SMP projects continue to come out from the same researchers instead of them switching to GPU and releasing new core 17 projects. Also, as more proof, the reason they just released a new core 15 project instead of switching to core 17.

Fact: You can define equal work for equal pay any way you want, but the only one that counts in the one defined by Pande Group. It has never factored watts, electricity cost, or hardware cost in their definition of equal (scientific) work for equal pay (points). Optimizing PPD/$ has always been up to the donor.

Old horse. New year is coming.

Crap Points and GPU WUs
So, I quit Folding@Home
Unbalanced Scoring
Points are (or should be) social engineering
Uniprocessor client or SMP?
[Debate] Revamped points system based on early return of WUs
benchmarking F@H
F@H and World Community Grid
Comparing performance between folding clients
Curious, why are points so important to some?
Open letter to PG [with response]
What is going on with BIGADV?
Fair distribution for PPD BigAdv / GPU
Is PPD adjusted for inflation as time passes? [No]
Lopsided points when folding -smp
Pande Group announcement about Big Advanced Projects
Suggested Change to the PPD system
PG's use of points to influence how the science is done
Bigadv Reform
Blog post "Unified GPU/SMP benchmarking scheme..."

That is in this forum section alone. Out with the old, in with the new. Happy New Year!

nothing as compelling as the:

"we've always done it this way" argument.

It is a classic... how often we have heard it from the likes of the great venerable, GM,Xerox,Kodak,city of Detroit...the nostalgia alone just chokes me up...

ah, auld lang syne...will the 100K donors that left us this year ever be seen again?...

oh well, we'll always have the memories....

Fact: Protoneer tabled developing the GPU BidAdv project because there was no interest in it from the current researchers. And not because it doesn't work as was intimated. Check the #fah IRC log on that one.

Fact: until you show me it works it does not work the current biggest WU is 40,000 atoms and takes 4 hrs to complete on the fastest cards so extrapolated to bigadv size hmmm does not look so good. bigadv is how many atoms

Just a guess but I would say it is going to be a while before this is a reasonable goal and most likely with coprocessor. I also know that my GPU's are not surviving these latest Core 17 WU's too well whether it is GPU quality or WU related idk but I have lost 3 out of 6 GTX 580 and GTX 680 to GPU failure in the last year so I am not real excited about GPU folding at this time (I have never had this kind of failure in the past). This is quite possibly nothing but it may be something that becomes an issue in the future.

Fact: Researchers do NOT switch fahcore types in the middle of their research project. That explains two things, the above fact, and also why SMP projects continue to come out from the same researchers instead of them switching to GPU and releasing new core 17 projects. Also, as more proof, the reason they just released a new core 15 project instead of switching to core 17

I will give you that but why are all the new projects smp

It's the same as when this same idea was proposed the last two times that BA was adjusted. The first time, BA bonus was reduced, and the second the Core threshold was raised. Neither time was SMP adjusted, so very unlikely it will change this time either.

I believe most are talking about the need for smp work getting done (one of the reasons listed for the change) and the lack of incentive to do them. I know I no longer run SMP on my 980X rigs when I do run them on F@H they run GPU only I can drop them to low wattage settings on the CPU and only loose 20K PPD and use 200 watts less electricity, so there is no incentive to run them. In fact there is an incentive to not run smp there are far more productive choices.

mdk777 wrote:nothing as compelling as the:

"we've always done it this way" argument.

It is a classic... how often we have heard it from the likes of the great venerable, GM,Xerox,Kodak,city of Detroit...the nostalgia alone just chokes me up...

ah, auld lang syne...will the 100K donors that left us this year ever be seen again?...

oh well, we'll always have the memories....

Correction. 100K active clients, assuming you didn't exaggerate.

How many people shut off a farm of single CPU systems to build 1 or 2 Bigadv systems? How many 1 core clients were shut down with the last server move when core_78 went away? How many older GPUs got shutdown when GPU2 ended?How many people sold multiple low end GPUs to buy 1 GTX Titan? That certainly doesn't explain all of that 100K change, but please don't pretend to know why they all left, or that they all left for one single reason.

All 100K was certainly not because of a fah running as status quo, either in policy or plan. If anything, any drop was from too much change, not the other way around.

There was this interesting tongue and cheek post on my forum... it went something like this:

If all the folders who fold left folding what would Stanford do?

Answer: Require 64 cores.

Again it was sarcastic and a tongue in cheek moment but it reflects what I believe is the reality we perceive in dealing with the folding in general and management in particular.

Machines may have been turned off to fund 4P builds (some have done this) but now those who built their anklebiters (24C machines) will no longer feel inclined to fold perhaps. At least not SMP... so maybe 100 more machines drop out?

Sure the recession hit hard but FAH faces far bigger challenges from BOINC, WCG and coin mining. Unless there is some sort of recogintion from the top that these problems exist and they act upon solving those problems, then FAH will go the way of SETI... occasional funding and hope you survive mentality. If folders cease production, loose interest or find a better value for their donation then I guarantee you research money will dry up because you cannot produce enough papers and research to justify your existence to those who fund you.

End of line....

Grandpa_01 wrote:

7im wrote:Fact: Protoneer tabled developing the GPU BidAdv project because there was no interest in it from the current researchers. And not because it doesn't work as was intimated. Check the #fah IRC log on that one.

Fact: until you show me it works it does not work the current biggest WU is 40,000 atoms and takes 4 hrs to complete on the fastest cards so extrapolated to bigadv size hmmm does not look so good. bigadv is how many atoms

I agree. Until GPU BA finishes being developed, we don't know what hurdles will need to be overcome, nor what speed it will run. Please don't assume atom count scales in a linear manner with time to process. Atom count has never been a good indication of processing time, even using the same core types, let alone different core types.

Grandpa_01 wrote:I will give you that but why are all the new projects smp

Because all the current researchers started their projects using SMP. Again, they don't switch cores mid-project. And Core 17 is a very recent development. How many new PG researchers have even joined FAH since core 17 was available?

Grandpa_01 wrote:In fact there is an incentive to not run smp there are far more productive choices.

I agree there also. But only if everyone was a FAH enthusiast like you, who buys hardware specifically to maximize FAH production, and to do nothing else. Don't forget there are large numbers of home and corporate donors that fold part time, or fold on a single computer. SMP is a good choice for those donors. Points are not the only incentive for them to fold, nor their biggest incentive. Clearly that doesn't apply to everyone.

When AVX comes to SMP, the whole landscape changes again.

With 100k ACTIVE donors now, if 50K stop folding, PG would say: "Screw them, we still have 50k others. We rule because they run our program. They voluntarily run F@h, we didn't ask them to .. lol......" --- That's how I look that what 7im said.

7im wrote:When AVX comes to SMP, the whole landscape changes again.

https://github.com/SimTk/openmm/pull/251