Estimate personal share on overall compute power in FAH

[ThWuensche]

Guess that question has been discussed before, however could not find that information.

I would be interested to estimate the part of compute power I'm contributing to the project. However all attempts have severe uncertainties. Taking the PPD and comparing it to the overall PPD (as published on EOC for example) has the influence of quick return bonus, as a result fast hardware will indicate to high, slow hardware to low share. Part of completed WUs probably is also misleading, as CPU WUs seem to finish much faster on my system than GPU WUs and in general the compute power requirements differ widely over the WUs. Base points per day I could not find published, furthermore I don't know statistics which show base points completed instead of PPD.

So in my case the estimate ranges from 1/5000 of compute power (200WU/d compared to a total somewhat above 1MWU/d in total) to 1/3000 (11MPPD compared to about 30GPPD in total). The PPD based calculation does not properly take into account the compute power supplied by clients without passkey.

Is there a better way of estimation?

This calculations however, even in the weaker case of 1/5000 of total power, lead to some interesting commercial considerations. If I roughly estimate my expenses for hardware in the range of EUR 7000, then the hardware value for providing the compute power of FAH is about 35M EUR. That seems quite cheap compared to supercomputers providing the same performance. That of course works only in case of workloads that can be divided in the way FAH does.

[UofM.MartinK]

I am actually working on that, and I am currently using base_credit as the basis of actual work performed.

[ajm]

For these considerations, ultimately, I think that we will have to take the power consomption into account? Wouldn't it be a good "investment" to integrate in the client some secure ways to monitor the effective power consomption? This would flow into the stats and be freely available to the folders, possibly with some "official" FAH declaration.

If it works, it would provide a solid basis for folders to optimize their systems' efficiency and their FAH-related acquisitions (computers and/or power production hardware like PV panels), and negotiate subsidies, tax-deductions, and special tarifs with their local state authorities and/or power producers.

[Neil-B]

In the big scheme of things I am contribution about 0% - but I am quite proud of the statistically insignificant amount I am doing and will continue to do it

[PantherX]

...I think that we will have to take the power consomption into account? Wouldn't it be a good "investment" to integrate in the client some secure ways to monitor the effective power consomption? This would flow into the stats and be freely available to the folders, possibly with some "official" FAH declaration...

You're not alone and there are some thoughts around it... however, let's see what happens since Nvidia (and potentially AMD) GPUs might get that feature before the CPUs do: https://github.com/FoldingAtHome/fah-issues/issues/1554

[UofM.MartinK]

It's all connected Although I have higher hopes in getting nice(r) absolute contribution numbers (relative to FLOPS or similar), than reliable power.

But logging/reporting what the GPU drivers give us by default is a very good start!

[ipkh]

It's a trick problem figuring out a good number. Especially as crazy folks like me and others are running 5+gpus at home. A weighted daily/weekly/monthly producer dashboard might be useful in that.
It's especially problematic to factor in users such as myself with 5 GPUs running mostly 24/7 at hone with 3.2 billion points already logged and ranking 250 overall.

[MeeLee]

You'll have to account for a couple of things,
one being: System downtime. Not always having a fast and stable system, means consistent PPD. Sometimes the system is waiting for WUs.
Another being that CPU WUs generally are smaller in size, and quicker to finish, than GPU WUs. So you can't count on 'finished WUs' either.
Third, People upgrading. Perhaps once they spent years running WUs on older hardware (GTX series GPUs), and now will be buying an RTX 3000 GPU, that will basically get the same PPD in a tenth to a hundredth (or even 1/1000th) the time.
Fourth, core optimizations. We've gone through a few core optimizations. Whatever ran on older hardware, has been upgraded, to further make use of newer hardware feats; faster folding (both on CPU as GPU).

Considering the amount of WUs completed, and the amount of PPD distributed, even the highest folders with billions of PPDs, won't see anything more than 0.001% as their contribution.
There are about ~+250k teams, and some teams make more points per day, than even the fastest users do per entire year.

[ThWuensche]

@Neil-B Of course any contribution is good, independent of the share in the total. But this still does not exclude interest in the share in the total. I also will go on folding independent from the figure representing the share.

@MeeLee Of course there are many things influencing the share. But that does not mean there is (or might be) no way of calculating it. The way of estimating it based on numbers was my initial question/goal.

@UofM.MartinK If you have progress with your effort on absolute contribution numbers, please let me know. Would be good to get that into official statistics in addition to PPD.

[UofM.MartinK]

@UofM.MartinK If you have progress with your effort on absolute contribution numbers, please let me know. Would be good to get that into official statistics in addition to PPD.

Yes, that's my goal as well. I will soon have some things to show using a preliminary study based on the "imperfect" base credit, hopefully weighted with some calibration data per project.

And I think we can spark interest within the FAH team to "properly" count or estimate "floating point operations" or something similar per project or WU. They might even have something like that already, just no time yet to publish, make an API call for it or store it in the database.

[HaloJones]

The contribution statistic will be so small for 99.9999% of users that it will simply put people off