Folding Forum

I base my theory on the fact that basically every new supercomputer gets more and more GPUs but reducing the number of CPUs and in addition that every higher piece of software gets GPU optimised.

But they are currently not doing anything close to what a CPU can when it comes to FAH and GPU's are currently getting far more points per atom processed than CPU's.

From my understanding Stanford has designed the point system to deliver points in proportion to the scientific value. Extremely high point values are equatable to extremely high scientific value. The fact is that the range of capabilities between CPU's/GPU's combined with people usage patterns is extremely large. Further, as time marches on the top end gets further and further away from the bottom end of the performance curve because technology advances. There are people with part-time P4's (or in a few cases even older CPU's) to full time folding using incredibly high performing multi-socketted CPU's and everything in-between. There is scientific value for it all.

I do not pretend to know relative scientific values other than faster returns are much more valuable than slow. In this respect I trust Stanford for they design the WU's, assign point values, and produce the scientific papers. They really have a great incentive to match the point values to the scientific value because many people optimize for points not knowing what is best for the science.

While I understand that point inflation exists and that it seems to drastically devalue old contributions. When I started folding, I had a very high-end machine and it took over a year to get 60K that I now get in a day. I still don't believe in handicapping Stanford's ability to assign point values to match the scientific value by dictating to them what should be done for all it can do is create a distortion to the concept of points being proportional to the science.

All I care about is that the point system is relatively fair so that the same rules apply to everyone. Anyone that wants to invest in a super high performance machine can get those incredibly high point values. Generally people are not willing to spend that kind of money but more power to those that do. I do not begrudge them their points. It is not a zero-sum game where their points get taken from me. So why should I care?

I can buy a SR2, or I put 4 GPUs each in to 2 cheapo systems. The cost and PPD are about the same. What's the problem?

And it's not so much points inflation as it is Moore's law affecting how fast we can fold. Does everyone really want to go back to the days of folding 1 WU for 2 weeks on a Celeron 300a? Now compare that PC hardware to what we have today. The current CPU clock speeds are 10x faster than back then. And the Floating Point Unit in today's hardware can processing multiple instructions per clock cycle, where the Celeron could not. So 10x becomes 30x today. Throw in a quad or 8 core processor, and we have 120x or 240x processing power today. What is 240x the old 110 PPD? The answer is not that far off what we are actually getting in PPD.

As P5-133XL said, the points follow the science getting done.

And I'll say it again, the QRB should be capped at X times the number of cores you dedicate to processing. Seems the fairest way to determine the top end points. quads get 4x, quad quads get 16x, etc.

Putting a cap on the points is contrary to the points being proportional to the science for there is no corresponding cap on the science value. All a cap does is create a distortion in the proportionality. Further, assuming that CPU/GPU capabilities increase over time that distortion increases over time too as more and more get affected by that cap. Eventually the distortion will get where it is more valuable to run multiple instances where each instance completes 1/n slower which results in less scientific value being done because the returns are not being returned as fast. If there is one constant in folding it is that the faster the return the higher the scientific value. Caps don't work in any situation where scientific value is related to the speed of WU return and both need to be proportional to the points value.

Please note that any discussion of the Beta Project 6903 is against Forum Rules (viewtopic.php?f=66&t=18031). So please make sure that future discussion doesn't include any Beta Projects. Thanks.

And I'll say it again, the QRB should be capped at X times the number of cores you dedicate to processing.

Threads vs core?
1536 cores on a graphics card ?
Cores and GPU on the same die?

Seems like an over-simplification to me.

k1wi wrote:

Grandpa_01 wrote:So I am trying to figure out what you are baseing your argument on. ? Clearley based on the amount of science done based on the newer Openmm GPU vs new bigadv (New bigadv = 2,533,797 atoms / Openmm = 292 atoms = 912 points) So if we take the new bigadv 2,533,797 and divide that by the new GPU 292 that = 8677 new GPU WU's have to be completed to = 1 new bigadv. (GPU Science completed = bigadv Science completed) so 8677 GPU WU's x 912 points ea. = 7,913,424 points according to the amount of the science completed the point value of the GPU is absurd it should be around 120 points. So that argument is unfounded. So is your argument based on emotion or what.

I thought it was always unwise to measure purely on the number of atoms - what if the GPU work units were 7 times more steps?

Plus Openmm is newer code than bigadv and they've mentioned they're using projects with small atom-sizes to test it, I guess you chose them on that basis over GPU2 so that it supports your claim (measuring just on atoms)?

I guess what you're saying grandpa is that GPUs are terribly inefficient at realising their theoretical computational power? and are thus being overcompensated in points?

Nope just using what appears to be apples to apples to a simple folder. I do not pretend to know anything about the atoms other than there posted size and the assigned values of the WU's and, a few of them there relative upload size. And using the readily available information supplied by Stanford ie Project Summary page which is the first place where most folders go for information about a WU, it appears that the GPU's are overcompensated. I do not believe they are but that is the way it appears. And it does not matter if you use the OpenMM WU's or the other ones it is still the same results. Perhaps you can enlighten me as to why they have such a high value to atom ratio. ?

I cannot "enlighten" you as to why they have such a point to atom ratio, but I think it is naive to just look at atom number as the only measure of how many points a work unit should have. Why? Because it's like looking at a race and saying "it's 10 kilometres long" while not also pointing out that it is up a 2000m hill or that it's being run in a desert. Atom count is just one dimension of difficulty in folding. (On ATI it doesn't seem to list the number of steps, so I cannot check that)

It *may be* that because GPUs are so fast at folding, but are limited for some reason to a limited number of atoms, they have a much larger number of steps per work unit (perhaps ten times fewer atoms, but 100x more steps). After all, the theoretical native & 86x TFLOPs are much higher on a GPU than a CPU (the Nvidia 560 has 1TFLOP whereas the i7 920 @ 3.4Ghz only has 70GFLOPS; but they are apple GFLOPS to orange GFLOPS). Just look at the number of FLOPS recorded by cores on the stats page and you can see where the bulk of Folding@Home's raw computing power comes from.

The reason why I said that it matters that you used OpenMM WU's is because you picked out the smallest possible work unit from which to make that claim, which I believe are of a small atom count because the core is still very new, using brand new interfaces & coding and thus still being tested. They are much newer and experimental than even the new CPU cores.

One other significant factor, besides the number of atoms, is the simulation duration (ref Runs, Clones and Gens An explanation by Dan Ensign). How long is each GEN along the time axis?

One of the few project descriptions that defines that is the one I've quoted below - half a nanosecond. Parenthetically, I've picked up two of these babies on my P4/HT 3.2 GHz one core Uniprocessor, and it's taking 5 days or so to get the 493 points. . This must be one patient team, since they've assigned a 136 day deadline!!

But, the same project series can also go out as SMP with only a two day deadline and a k-factor of 3. I haven't seen one of these in SMP, but fast folders will get, relatively, lots of points.

The science value is the same in either case - same number of atoms, same WU duration. base points for the SMP case is less than the base points of the single-core clients. I would guess that the points were assigned relative to performance on the benchmark machines. I would have thought that getting results in two days would be a lot more valuable than in 136 days.

I can understand the purpose of the point system, more points for more scientific value to stimulate people to return work units fast and use certain clients, but personally, important: personally, I don't really like this. I would rather have a point system that actually represents the amount of work everyone did and not the scientific value. I am not a scientist, and although I'd find it interesting to know how useful my contributions are, I'd like more to be able to see how much work I actually did.

Some people earn hundreds of thousands of points every day, and they aren't doing for example 200000 / 500 = 400 times more work than me if I earn 500 points per day using the standard client, and someone else earns 200000 using an advanced client and the latest hardware. Therefore the point system doesn't represent enough how much work everyone does. On the other hand, the scientists need to stimulate people to use folding@home in such a way that the scientific value is high, because in the end that's what this project is all about, so I understand why the point system currently is the way it is.

Maybe it's a good idea to have 2 types of points, one that represents the amount of work done and one that represents the actual scientific value of the work. Both are interesting things to keep track of for participants I think. Now it just feels to me as though something is missing. The amounts of points everyone has isn't the full picture.

The amounts of points everyone has isn't the full picture.

What I think everyone in this tread is looking for is "ownership" of a piece of a project.
For example: A million points four years ago was a significant contribution to that years work.

Without knowing the total points generated in any given period of time, it is impossible to know what percent a given amount of points represents to the whole.

Even a million points per day today equals what percent of today's total? No one knows.

Point "inflation" is only a problem if it does not reflect an actual increase in the amount of work done. Getting 200K PPD is not a problem if the actual amount of work done is 1000 times greater than it was when people 200 ppd. (which I think is happening, for example cost per GFLOP(hardware only) has gone from something like $1000 in the year 2000 to $1.18 or less today http://en.wikipedia.org/wiki/FLOPS#Records )

Anyway, a yearly, or project percent becomes an accounting nightmare I realize. But it would put things in perspective for people.

Yeah, your 1 million points were a great percentage of the contribution when the yearly production was billions; but today when the yearly production is trillions, you have to expect to produce more to stand out.

Perhaps just a running daily account of total PPD somewhere would help people evaluate their contribution more accurately.
JMHO

Actually, you can quite easily see what FAH as a whole is doing, then measure individual contributions against this. [H]ardfolding's stats site includes a section on overall project stats (F@H Statistics >> Project Information). I've extracted a snippet below from this morning's stats. The point averages column is probably what mdk777 is looking for.

Thus, your one million point per day folder would be getting, on average, 0.548% of Folding@Home's total production.

I don't know exactly how each stats site defines ACTIVE folders, but [H]ardfolding's site says 26,202, and Kakao's site 77,094.

EDIT: I went into EXCEL and computed the stats per folder, using [H]ardfolding's site value of 26,202

EOC uses a 7 day window- any result sent within the past week and you are marked active. i imagine those 2 do something similar.

well, there you go.

Chart it on the stats page along with the total # of donors and people would see the exponential increase in work done in comparison to the linear # of folders.

http://folding.stanford.edu/English/Stats

The entire question of "point inflation" or "capping" would become self evident.