F@H Web Client Dark Skin with PPD comparison by GPU / WU

[TVwazhere]

On 9/12/2020 at 11:01 AM, LAR_Systems said:

Looks like your hard work is starting to move the needle as WU calculation window slides.
https://folding.lar.systems/folding_data/gpu_ppd_datasheet?gpu_model=gp104_geforce_gtx_1070_ti_8186

Spoiler

 

 

Admittedly I dont think I've opened it on ITNOS yet

I keep remembering at 8am when I'm at work and dont get home till 5pm. I should set a phone reminder....

[TVwazhere]

On 9/14/2020 at 8:50 AM, TVwazhere said:

Admittedly I dont think I've opened it on ITNOS yet

I keep remembering at 8am when I'm at work and dont get home till 5pm. I should set a phone reminder....

Finally did it. You should now see +2 1070ti entries  

 

Curious to get home after almost a day now and see how 60% power limited 1070ti's stack up against unknow power limited/boosted cards. 

[TVwazhere]

3 hours ago, TVwazhere said:

Finally did it. You should now see +2 1070ti entries  

 

Curious to get home after almost a day now and see how 60% power limited 1070ti's stack up against unknow power limited/boosted cards. 

Neat. Other card isnt boosting as high so it's ONLY 30% *edit it's now about 45-50%*, but both cards are rank 26/90?

[Kilrah]

47 minutes ago, TVwazhere said:

but both cards are rank 26/90?

It's not your card's rank, it's the average rank of all cards of the type.

[TVwazhere]

8 minutes ago, Kilrah said:

It's not your card's rank, it's the average rank of all cards of the type.

Did not know that. 

 

So I guess there's no way to know on our end how many other cards of our type there are and how we rank in our "class" so to speak?

[Kilrah]

Nope

[TVwazhere]

4 minutes ago, Kilrah said:

Nope

[tep]

Is it possible to reject submissions which takes more than double the average time of lets say top 20 finished WUs for a specific project. Just a suggestion to make the database more accurate in terms of PPD.

[A-J-M]

Not sure whether you guys know of have noticed, but yesterday FAH launched a new version of their core22 (0.0.13) that takes advantage of Nvidia GPU's CUDA cores. For me, it brings some 40% overall increase in PPD.

You don't have to do anything - the new core should be downloaded automatically. But I guess it will produce a sudden and massive discrepancy in your DB. 

https://foldingforum.org/viewtopic.php?f=38&t=36165

[LAR_Systems]

4 hours ago, A-J-M said:

Not sure whether you guys know of have noticed, but yesterday FAH launched a new version of their core22 (0.0.13) that takes advantage of Nvidia GPU's CUDA cores. For me, it brings some 40% overall increase in PPD.

You don't have to do anything - the new core should be downloaded automatically. But I guess it will produce a sudden and massive discrepancy in your DB. 

https://foldingforum.org/viewtopic.php?f=38&t=36165

Thanks for the heads up.  F@H had published a blog on it, and the system will just slowly start re-calculating the averages.

To help visualize this update I have added average historical PPD going back 12 weeks to GPU profile page.


 

[LAR_Systems]

9 minutes ago, LAR_Systems said:

Thanks for the heads up.  F@H had published a blog on it, and the system will just slowly start re-calculating the averages.

To help visualize this update I have added average historical PPD going back 12 weeks to GPU profile page.


 

P.S.  This curve is much steeper when you cherry pick the people clearly running the latest Nvidia drivers that allow taking advantage of optimized core.   It will be very interesting in a few weeks as people update to see the previous weeks flatten out... as I see 2060 Supers on Linus getting 2.5 mil on some projects.   But the average is brought down for the time being with people running the old drivers, windows,  etc.

[A-J-M]

7 minutes ago, LAR_Systems said:

P.S.  This curve is much steeper when you cherry pick the people clearly running the latest Nvidia drivers that allow taking advantage of optimized core.   It will be very interesting in a few weeks as people update to see the previous weeks flatten out... as I see 2060 Supers on Linus getting 2.5 mil on some projects.   But the average is brought down for the time being with people running the old drivers, windows,  etc.

I'm running Windows, now on all machines. 

I got interested in that so famous Linux superiority last Spring and made some serious tests, folding on two GPUs (1 Nvidia, 1 AMD) on a one machine, one week long (24/7) with Windows, then one week long with Ubuntu 18.04 (and then through WSL2 and Vbox), collecting all WUs with their TPF. There is a significative advantage when folding on the CPU with Linux indeed, but I could not find any sustained difference for GPU folding. And I spent many stupid hours setting up and fixing the damn system for Linux (mixing GPUs is a nightmare on Linux). So I wouldn't take for granted that Windows users will slow down this progression here. 

[Sequence211]

I downloaded the plugin, and installed the newest NVidia drivers, and the new folding@home app.

I can say that the new cuda support, kills it.

Compared to last time i folded, (february 2020) my PPD is up from 800k PPD to pt 1680k PPD, that is over 100% perf increase on my gtx 1080.

The current project is 13426, and i will see what the final stats end up being. But this is realy nice.

 

[A-J-M]

11 minutes ago, Struck said:

I downloaded the plugin, and installed the newest NVidia drivers, and the new folding@home app.

I can say that the new cuda support, kills it.

Compared to last time i folded, (february 2020) my PPD is up from 800k PPD to pt 1680k PPD, that is over 100% perf increase on my gtx 1080.

The current project is 13426, and i will see what the final stats end up being. But this is realy nice.

 

13426 is a Moonshot WU. These are experimental in some ways and deliver odd results in terms of PPD. As the differences were mostly very negative at the beginning, the researchers artificially raised their PPD. It is very difficult to estimate to what this would have corresponded in February.

[GOTSpectrum]

1 minute ago, A-J-M said:

13426 is a Moonshot WU. These are experimental in some ways and deliver odd results in terms of PPD. As the differences were mostly very negative at the beginning, the researchers artificially raised their PPD. It is very difficult to estimate to what this would have corresponded in February.

This isn't exactly true....

 

The Point reward were not 'Artificially altered' please see the below quote for explanation 

 

  

On 9/18/2020 at 6:59 PM, GOTSpectrum said:

WUs are tested on a standard benchmark machine and are given a 'difficulty' based on the length of time it takes for that machine to complete the WU, it is then assigned a points award based on that difficulty level. This means that getting WUs that vary in difficulty and thus in time should not have a major effect on your average points over a protracted length of time, eg over days or a week. This applies to most projects due to the linear nature of most simulations.  

 

There are some projects however that due to the type of work they are doing, there is no way to gauge the difficulty of each individual WU, the most recent example of this is the COVID moonshot WUs that due to the very nature of what it is doing (testing many different samples against a given protein to find low binding energies) there is no way to test every permutation of the work. In these cases a representative sample of the WUs is tested and an average used. This mean that the points can vary greatly as the project progresses. The idea is that over longer periods the average should kick in meaning points should be relatively stable. In practice however, we have seen that this isn't the case.  

 

[A-J-M]

https://foldingforum.org/viewtopic.php?p=342185#p342185

Quote

@cine.chris: We hear you, and we will adjust upwards again for the next Sprint as a compromise. Please do stick with us for Sprint 4, starting Sunday!

~ John Chodera // MSKCC

John Chodera is the main researcher behind the Moonshot folding, and the actual researcher responsible for the 13426: https://apps.foldingathome.org/psummary

[GOTSpectrum]

Just now, A-J-M said:

https://foldingforum.org/viewtopic.php?p=342185#p342185

John Chodera is the main researcher behind the Moonshot folding, and the actual researcher responsible for the 13426: https://apps.foldingathome.org/psummary

But they weren't artificially altered, they were corrected due to the WUs having low PPD. The point of PPD is that it is meant to be generally average across a period of time. 

 

I would say if they made it significantly higher than the average, then it would be an artificial alteration, but in this case it was simple an aware correction, which does happen time to time. I guess I'm just being a little pedantic over the wording but then again, that's just who I am!  

[A-J-M]

You would be right in a normal situation, but Moonshot is really very special. It is an experiment. It was not really a project for FAH, actually, and the researchers had to tweak many things to make it work. 

[LAR_Systems]

1 hour ago, A-J-M said:

I'm running Windows, now on all machines. 

I got interested in that so famous Linux superiority last Spring and made some serious tests, folding on two GPUs (1 Nvidia, 1 AMD) on a one machine, one week long (24/7) with Windows, then one week long with Ubuntu 18.04 (and then through WSL2 and Vbox), collecting all WUs with their TPF. There is a significative advantage when folding on the CPU with Linux indeed, but I could not find any sustained difference for GPU folding. And I spent many stupid hours setting up and fixing the damn system for Linux (mixing GPUs is a nightmare on Linux). So I wouldn't take for granted that Windows users will slow down this progression here. 

OS reporting is going to be coming to the site soon enough, have to collect a baseline of data before chart / tables can be introduced reliably.

That said, Linux PPD is generally better than Windows depending on GPU, project etc. based on test data I have been working with so far backing up what the community tends to report.  But you need weeks to months of data samples across various projects to really get an average due to fluctuations in F@H projects and individuals settings, thermals, work server upload times etc. that impact PPD.

Here is a backend pull from the system, highlighted color is bigger.   Based on what I was observing a week ago vs now I would suggest the windows based systems that are higher than the Linux based ones today are due to the windows users having likely updated to the new Nvidia driver and the averages are adjusting accordingly on the my testing system while Linux users in many cases run headless and are not likely to upgraded yet.

As more Linux user update their systems, they should be pulling more PPD once again.   Also of note, GPUs on Linux Mint has been very straight forward to setup in my test environment, seems to be much simpler to maintain than Ubuntu.


 

[A-J-M]

Thank you! You've got a very good data basis indeed! 

This is what I got back in Spring, before the Moonshot: https://docs.google.com/spreadsheets/d/1dUXspeOY0pGZQvqaPVUSkjL9O_D4afcE_N8pvDWMXp0/edit#gid=0

 

EDIT: And this is the result of the launching of the new fahcore yesterday for my systems (2 20280S, 1 2070S, 1 1080ti, 2 1050ti and total 76 active CPU threads): https://folding.extremeoverclocking.com/user_summary.php?s=&u=911175 (until 09.08, there also was a 5700XT, now sold)

[Sequence211]

About 13426 i can say that my AMD gpu, got the same project 13426, and it produces about the same PPD as always. But for that there was no driver updates etc. It is a 560X on a Macbook pro 2018,with driver 19.3 in bootcamp windows 10 2004.

Soon my gtx 1080 is finished with the WU, and next WU will show if the PPD stays high. (currently 1.7M PPD on 13426)

 

EDIT: next wu is also 13426

[LAR_Systems]

20 minutes ago, Struck said:

About 13426 i can say that my AMD gpu, got the same project 13426, and it produces about the same PPD as always. But for that there was no driver updates etc. It is a 560X on a Macbook pro 2018,with driver 19.3 in bootcamp windows 10 2004.

Soon my gtx 1080 is finished with the WU, and next WU will show if the PPD stays high. (currently 1.7M PPD on 13426)

 

EDIT: next wu is also 13426

Keep in mind when dealing with any laptop GPU/APU they are thermal throttling when folding pretty much guaranteed making them not great for relative PPD comparison as the biggest fluctuation factors will likely be heat (ambient room temperature and ventilation near the laptop)  and what else you are doing on the laptop causing further thermal throttle bottleneck of the of CPU.

Bootcamp on a Macbook is also known to have strange fan curves compared to when running native OSX which can negatively impact performance of the GPU/CPU.  Apple does optimize for windows at the hardware level.  If you have not already I would suggest you try to get the bottom of the laptop raised off any flat surface and get some air passing over/under it even if just from a little USB fan.

I have experienced this issue first hand folding and gaming on Apple systems across different models of laptops and mac minis and you see big drops about 10-15 mins in.

[A-J-M]

I had a look at the 14326 WUs on my HFM DB, with cards between 1050ti and 2080S. Looks like the PPD gain is bigger for average GPUs. I see an almost incredible peak at the 1080ti level (~75%). The gain is still around 45% for the 1050ti (albeit I have less of those 14326 WUs there) but drops at around 30% for the 2080S (where the bulk of "my" 14326s is concentrated). Albeit the PPD spread is also much larger for bigger GPUs and I have not looked at averages there. 

[Sequence211]

55 minutes ago, LAR_Systems said:

Keep in mind when dealing with any laptop GPU/APU they are thermal throttling when folding pretty much guaranteed making them not great for relative PPD comparison as the biggest fluctuation factors will likely be heat (ambient room temperature and ventilation near the laptop)  and what else you are doing on the laptop causing further thermal throttle bottleneck of the of CPU.

Bootcamp on a Macbook is also known to have strange fan curves compared to when running native OSX which can negatively impact performance of the GPU/CPU.  Apple does optimize for windows at the hardware level.  If you have not already I would suggest you try to get the bottom of the laptop raised off any flat surface and get some air passing over/under it even if just from a little USB fan.

I have experienced this issue first hand folding and gaming on Apple systems across different models of laptops and mac minis and you see big drops about 10-15 mins in.

I know about the throttling issues. but atm the gpu is power throttling at about 25W.  

It is not a full time machine, and was only to test this one WU.

i have good ventilation, and the macbook is raised so there is about an inch underneath of free air. the gpu is at 75-80C, and the fans are mostly silent, since there is no load on the cpu.

[LAR_Systems]

2 hours ago, Struck said:

I know about the throttling issues. but atm the gpu is power throttling at about 25W.  

It is not a full time machine, and was only to test this one WU.

i have good ventilation, and the macbook is raised so there is about an inch underneath of free air. the gpu is at 75-80C, and the fans are mostly silent, since there is no load on the cpu.

Sounds like you're doing it all right, sorry if it came off preachy.
I tend to warn people to be on the safe side, rather than someone brick their laptop.

Less of an issue when talking to people on this forum I have to say.