Folding Forum

Have you used Tools > Download projects From Sanford in HFM to get updated point values for the various projects?

Tohya wrote:Have you used Tools > Download projects From Sanford in HFM to get updated point values for the various projects?

Yes, several times.

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

And to all the following posts regarding this issue:

This WU is somehow crazy.
You can check your credit here:
https://apps.foldingathome.org/wu
Divide the credit by the duration in days and you will see if your ppd value in HFM is correct.
I checked several times, it is.

You can also check the database we are trying to build and you will find out that a R9 3950X makes >1.2Mio PPD with this project.

foldinghomealone wrote: You can also check the database we are trying to build and you will find out that a R9 3950X makes >1.2Mio PPD with this project.

It's a stock R9 3900X, running on air (and doing better than many graphic cards for less Watts)
Sadly we don't have a R9 3950X in the database...

NoMoreQuarantine wrote:I believe the correct way to do it is to simply compare the atom count to the time it takes to process 1 frame (TPF). Which I have done here: https://imgur.com/gallery/nXFM3cV

TPF is preferred over PPD bonus or credit bonus because it is time independent to FAHClient. That is, it does not get screwed up by people pausing or interrupting the folding.

The plots are quite interesting!
For the radeon vii, it seems even with 450000 atoms she's not reaching saturation. Same seems to be true for the GTX Titan X (is it a pascal?), but hard to say with the resolution...
"Funny" are the plots from the 2080 OC vs. the 2080Ti... Also the 1080Ti vs the Titan Xp: wouldn't have expected such a difference at small atom numbers!?

A shame we don't have more data to feed the plots, I'm really curious what will come out...

NormalDiffusion wrote:

foldinghomealone wrote: You can also check the database we are trying to build and you will find out that a R9 3950X makes >1.2Mio PPD with this project.

It's a stock R9 3900X, running on air (and doing better than many graphic cards for less Watts)
Sadly we don't have a R9 3950X in the database...

totally true...

NoMoreQuarantine wrote:

NormalDiffusion wrote:What was wrong? Was having a quick look this afternoon...

I believe the correct way to do it is to simply compare the atom count to the time it takes to process 1 frame (TPF). Which I have done here: https://imgur.com/gallery/nXFM3cV

TPF is preferred over PPD bonus or credit bonus because it is time independent to FAHClient. That is, it does not get screwed up by people pausing or interrupting the folding.

I think your assumptions are wrong.
The HFM output for TPF is (in my understanding) the time (between download and upload) divided by 100. So it is depended on interruptions or other slow downs.

I don't understand what you want to find out by comparing atom count vs. frame time.
TPF doesn't say anything about performance when comparing different projects (what you want to do).
There are projects performing good with high TPF compared to other projects that have less TPF.

During beta phase the PPD of projects are adjusted to an 'expected range'. If too many folders complain the PPD is too low then the base credit will be changed to get expected PPD.
The same should happen to projects which give way too high PPD.
TPF is only used to calculate PPD with an certain base credit (and other constants)

Therfore I guess, PPD vs. atom count is the only way to find performance vs. atom count

foldinghomealone wrote:

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

And to all the following posts regarding this issue:

This WU is somehow crazy.
You can check your credit here:
https://apps.foldingathome.org/wu
Divide the credit by the duration in days and you will see if your ppd value in HFM is correct.
I checked several times, it is.

You can also check the database we are trying to build and you will find out that a R9 3950X makes >1.2Mio PPD with this project.

That's strange, but I'll take it I guess. 4x the PPD of normal WUs for me.

I uploaded some new data.

foldinghomealone wrote:The HFM output for TPF is (in my understanding) the time (between download and upload) divided by 100. So it is depended on interruptions or other slow downs.

I don't think that's the case, but just in case I modified the graphs for atom count vs PPD https://imgur.com/gallery/LaKQNyA

My assumption was based on what OpenMM says a frame is: http://docs.openmm.org/7.0.0/userguide/application.html,
what FAH says TPF is:
https://foldingathome.org/support/faq/r ... ldinghome/
https://foldingathome.org/faqs/statisti ... -is-worth/
viewtopic.php?f=61&t=26036,
and my own experiments where I paused WUs and observed that the PPD & credit were negatively affected, but TPF was not. Which supported my hypothesis, but could be explained by other factors. Unfortunately, I'm afraid we may not know exactly how TPF is calculated until they release the source code.

One Frame is 1/100 of the complete WU. That is for sure
But how should HFM know what the Time per Frame is when it doesn't measure it?

Edit:
Thanks for the new graphs.
At first glance it confirms statements that AMD is less performing with small WUs

new uploads for:
i7-8550U
RX Vega 64
RX 580
Xeon E5649 @18t
i5-8400
GTX 970
R7 3800X @15t
R3 1200 @3t
GTX 1050
RX 460
RX 570
i7-6700HQ @7t
i5-8265U @7t
i3-4130 @2t

Thanks to
AOD_N3URAL
Kebast
almcdsmi

foldinghomealone wrote:But how should HFM know what the Time per Frame is when it doesn't measure it?

HFM does measure it. I played around with it a bit more in HFM. Both PPD and TPF, as reported by "Bonus (Frame Time)", will be affected by the same interruptions and slow downs. Their relationship is PPD = 14.4 * "Base Credit" * sqrt(14.4 * 0.75 * Deadline / TPF) / TPF with TPF in minutes and deadline in days. Personally, I still prefer TPF because it tells me how long it takes to complete a WU (which is what I care about), while PPD is scaled by their benchmark machine which definitely has it's own performance curve. https://foldingathome.org/support/faq/points/

The problem with TPF is that different projects will be optimized differently in addition to having different atom count, which could skew the results when comparing projects of similar sizes, but different relative performance. I suspect that is why they use the benchmark machine. The issue with that is that all it really reveals is how optimized a project is for the GPU versus the benchmark machine, rather than how optimized the project is for the GPU overall. For example, lets say it takes 120 minutes for the benchmark machine to complete a WU, while on a folder's GPU it takes 30 minutes to complete. Now let's say that for a different WU it takes the benchmark machine 60 minutes while the folder's GPU completes it in 30 minutes again. Looking at just the PPD we would be unable to tell that the folder's GPU is actually performing just as well for both of the WUs and potentially assume that it performed worse on the second WU.

foldinghomealone wrote:new uploads for...

I just had a look, and I'm guessing this is me:
R7 3800 @12t 146,948 105,568 228,783 23,103 36
R7 3800X @12t 135,983 119,273 178,313 16,636 24

I likely just forgot the 'X' in one of my uploads. I don't think that a vanilla R7 3800 exists.

Kebast wrote:

foldinghomealone wrote:new uploads for...

I just had a look, and I'm guessing this is me:
R7 3800 @12t 146,948 105,568 228,783 23,103 36
R7 3800X @12t 135,983 119,273 178,313 16,636 24

I likely just forgot the 'X' in one of my uploads. I don't think that a vanilla R7 3800 exists.

Ok, changed

NoMoreQuarantine wrote:

foldinghomealone wrote:But how should HFM know what the Time per Frame is when it doesn't measure it?

HFM does measure it. I played around with it a bit more in HFM. Both PPD and TPF, as reported by "Bonus (Frame Time)", will be affected by the same interruptions and slow downs. Their relationship is PPD = 14.4 * "Base Credit" * sqrt(14.4 * 0.75 * Deadline / TPF) / TPF with TPF in minutes and deadline in days. Personally, I still prefer TPF because it tells me how long it takes to complete a WU (which is what I care about), while PPD is scaled by their benchmark machine which definitely has it's own performance curve. https://foldingathome.org/support/faq/points/

The problem with TPF is that different projects will be optimized differently in addition to having different atom count, which could skew the results when comparing projects of similar sizes, but different relative performance. I suspect that is why they use the benchmark machine. The issue with that is that all it really reveals is how optimized a project is for the GPU versus the benchmark machine, rather than how optimized the project is for the GPU overall. For example, lets say it takes 120 minutes for the benchmark machine to complete a WU, while on a folder's GPU it takes 30 minutes to complete. Now let's say that for a different WU it takes the benchmark machine 60 minutes while the folder's GPU completes it in 30 minutes again. Looking at just the PPD we would be unable to tell that the folder's GPU is actually performing just as well for both of the WUs and potentially assume that it performed worse on the second WU.

Basically I would agree to the last paragraph.
- In my understanding - during initial testing the base credit is definied according what they would expect in returns of PPD with their benchmark HW.
So, maybe the benchmark HW makes normally around 1Mio PPD, they define the base credit in a way that the benchmark HW returns 1Mio PPD.
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. After beta test, base credit is fixed and then the only thing that is variable is the TPF (for this certain WU).
If a certain HW performs good with this certain WU it will return PPD as expected, if it can't handle this certain WU, it will return less PPD.

However, using TPF will work only comparing exactly the same WU.
If you want to compare different projectss you have to use PPD as it standardized on the benchmark HW.

TPF differ totally from WU to WU and don't correlate to PPD.
See (R9 3900X) p13872 with a TPF of 0:00:11 and PPD of 1,2Mio and p13861 (0:00:12 - 0,33Mio PPD)
or
(2080 Super) p11749 (0:00:46 - 1,9Mio PPD) and p13879 (0:01:33 - 1,95Mio PPD)