Folding Forum

foldinghomealone wrote:during initial testing the base credit is definied according what they would expect in returns of PPD with their benchmark HW (...) Then there is beta test and maybe the beta folders will be too happy about PPD returns or are very sad and maybe the base credit will be redefined.

That seems like an ok compromise considering the variables involved with providing "fair" credits. Effectively attempting to catch abnormal WUs with beta testing. I am not attempting to change the way they give points, but predicting GPUs that are able to complete WUs as quickly as possible. To be honest, right now I don't know enough to feel passionately about any one method of doing that.

NoMoreQuarantine wrote:...My biggest question is why projects of similar sizes perform differently and how large the variation can be.

But first you have to define what you mean with 'perform differently'.
You mean in terms of TPF?

foldinghomealone wrote:

NoMoreQuarantine wrote:...My biggest question is why projects of similar sizes perform differently and how large the variation can be.

But first you have to define what you mean with 'perform differently'.
You mean in terms of TPF?

I mean in general. What are the factors that go into a project? What types of projects are there and how do they work? What instructions are given to the GPU? How do the different drivers and architectures handle the instructions?

I want to know the variables involved so I can say with certainty "this is how it works and this is what we can do". Also because it's fun to learn and imagine the possibilities.

The biggest variable is the number of atoms.

P13401 had 2069 atoms. It performed completely differently on Windows than Linux with a PPD in Windows half that of Linux for identical hardware.

Then you have P14562 for example that has 371771 atoms and cards with more cores will suddenly leap ahead in PPD because they are able to return the units much faster than cards with fewer cores and so the Quick Return Bonus kicks in aggressively (some would say unfairly).

So the reason why we have to use long-term averages for cards is because we can't choose what work we get. I'd love to limit my TitanX to only get units that let it run at 1.15m but sadly units like P14310 come along and it gets 726K. As a beta tester, sometimes the units aren't great and they get adjusted by the time they hit the main body of users but also sometimes beta units give too many points too. It all averages out I think

foldinghomealone wrote:New updates for:
RX Vega 11

RX Vega 64
RX 580
RX 570
GTX 1050
RX 460
i7-6700HQ @7t
i5-8400
Xeon E5649 @18t

Thanks to:
BobWilliams757
almcdsmi
AOD_N3URAL

Thanks for doing all the work to compile this data for everyone. It's great stuff for anyone looking for new hardware, and much appreciated. I personally also find it very interesting to see how various hardware either excels or struggles with certain specific WU's as the project list populates. It's an insight into how potentially complex choosing hardware for folding really is.

Also, I noted that you rightfully placed the mighty RX Vega 11 on the top of this contribution list, above all others. As it should be!

Kebast wrote:HFM is showing project 13872 at 727k PPD for my CPU. Normal range is around 160k.
That an HFM issue?

NoMoreQuarantine wrote:I just checked my work history, I also had amazing performance on project 13872. Not as good as yours mind you. With my Ryzen 2600 I would normally get around 55K, but on that one I got 191K.

This is another thing the spreadsheet might help us all with. I just recently had two Core 21 runs of the same project that returned close to twice my average PPD. But they were then followed by a Core 22 project that is returning less than half my average PPD, and will actually exceed the Timeout by a couple minutes.

But looking at the data in the spreadsheet, another member has uploaded data for the Core 21 project (14158) and the PPD they reported is in line with their averages. It doesn't even seem remotely possible to me that certain hardware either excels or struggles to the extent of that much variation. If project xxxxx always returned higher (or lower) points across multiple members we might be able to assume there was just a human error in the beta phase IRT points calculations. But unless there was a change made after the project was released, I would think the only variation would be in line with each persons normal hardware deviation.


ETA - the WU for the project that was returning very low PPD (16435) finished, and the next WU picked up was the same project again. But it appears that it is now working towards a PPD that is more in line with the average for my machine. I have had multiple runs of the same project before, and the PPD average is usually very close.

BobWilliams757 wrote:...ETA - the WU for the project that was returning very low PPD (16435) finished, and the next WU picked up was the same project again. But it appears that it is now working towards a PPD that is more in line with the average for my machine. I have had multiple runs of the same project before, and the PPD average is usually very close.

It could be a possibility that the generation of the next WU could be a genuine bad WU. Alternatively, it could be a bad trajectory. It is expected that WUs within the same Project behave similar and have similar TPF.

New data uploaded for:
RTX 2060 Super
RX Vega 64
RX 5700 XT
RX 580
R9 3900X @21t
i5-8400
i7-8550U

Thanks to:
skydivingcatfan
AOD_N3URAL
SHeg

foldinghomealone wrote:
R9 3900X @21t

I think you can merge with the R9 3900X as it's always 21 cores for the 3900X due to fah...

I also thought about that but PPD are quite different. Want to investigate further

foldinghomealone wrote:I also thought about that but PPD are quite different. Want to investigate further

Does someone has their 3900 set for 2 slots, 12+9?
So 24 threads isn't a possible configuration?

Here's an updated log, the first 90 were uploaded before. R9 3900X @ 21 cores, RX 5700 XT factory OC

http://www.mediafire.com/file/e6a2jufpv ... T.csv/file

If there is a GPU, 24 cores end up at 21. Lower PPD could be due to the stock cooler causing throttling. 24 cores would work for most projects.

NoMoreQuarantine wrote:

foldinghomealone wrote:during initial testing the base credit is definied according what they would expect in returns of PPD with their benchmark HW (...) Then there is beta test and maybe the beta folders will be too happy about PPD returns or are very sad and maybe the base credit will be redefined.

That seems like an ok compromise considering the variables involved with providing "fair" credits. Effectively attempting to catch abnormal WUs with beta testing. I am not attempting to change the way they give points, but predicting GPUs that are able to complete WUs as quickly as possible. To be honest, right now I don't know enough to feel passionately about any one method of doing that.

HaloJones wrote:The biggest variable is the number of atoms.

P13401 had 2069 atoms. It performed completely differently on Windows than Linux with a PPD in Windows half that of Linux for identical hardware.

Then you have P14562 for example that has 371771 atoms and cards with more cores will suddenly leap ahead in PPD because they are able to return the units much faster than cards with fewer cores and so the Quick Return Bonus kicks in aggressively (some would say unfairly).

So the reason why we have to use long-term averages for cards is because we can't choose what work we get. I'd love to limit my TitanX to only get units that let it run at 1.15m but sadly units like P14310 come along and it gets 726K. As a beta tester, sometimes the units aren't great and they get adjusted by the time they hit the main body of users but also sometimes beta units give too many points too. It all averages out I think

To add to all the above, even lesser hardware does appear to be impacted by atom count, but it seems to be possibly the inverse of what is stated above by Halo Jones in some cases. The WU's I had recently that returned almost double my PPD average had only 5000 atoms. In Beta people with higher powered cards reported lower than usual PPD. It could be that hardware architecture, drivers, OS, etc all play into the hand and occasionally there will still be some unexpected results, even after going through the Beta stage. But I tend to think that just as some of the higher end cards love big atom counts, my lower end stuff seems to love the lower atom counts. But in both hardware cases our averages might be skewed based on whatever the averages are at the time for atom counts.

It almost seems that any given card might have a "sweet spot" based on atom count, where it delivers the best. That might be the driving factor in the variations of cards, depending on which WU's they just happened to get.

New uploads for:
RX Vega 64
RTX 2080 Super
RX 5700 XT
Quadro M2000
GTX 1660 Ti
RX 580
R7 3700X
i9-9900KF @12t
i7-7700
i5-8400
i7-8550U

Thanks to:
Tohya
absolut_zero3
Juggy
NoMoreQuarantine
AOD_N3URAL

I created a rudimentary comparison tool in Google Sheets that links to the current database. You can download it from here: viewtopic.php?f=14&t=35071