Does OpenMM take full advantage of rDNA for Navi GPUs?

[foldinghomealone2]

Did I say something to offend you at some point? I think you've been insulting my effort since the first time I saw your username.

Having new ideas is generally a good thing.
You're not experienced (which is no problem) and you don't know much about folding (which is no problem, either).

But:
You don't ask how it works. You make statements. Statements that in many cases are just wrong.
And you put your statements directly into new topics so that many also new folders will read them and get and keep a wrong understanding.

This is what I really dislike.

I don't know everything about how folding works and I posted some wrong statements here as well.
But I'm here to learn.

[NoMoreQuarantine]

You don't ask how it works. You make statements. Statements that in many cases are just wrong.

I can not do anything about what I do not know, I can only improve with new information. I state my assumptions and show my work so people can review and critique it. I don't have a problem with you criticizing my work and actually appreciate your input. I just do not like being told my effort is worthless, without at least an alternative approach.

[foldinghomealone2]

Just be aware that when someone posts assumptions in a new topic someone will act as an influencer.
The headlines and the first post will be read properly and nobody will notice if someone else clarifies or corrects the assumptions a few posts later.

Now, back to business...

[MrFrizzy]

Well I can certainly chime in here!

Take a look at the spreadsheet in my sig. I have been tracking hundreds of projects completed with my 5700 XT at various frequencies. As you might expect, as the atom count goes up, so does the average PPD! So far the most I have achieved with my card overclocked to 2.1GHz is 1.4m PPD, but the average is lower at about 1.26m PPD. When operating more like stock settings (1800MHz), the PPD falls to around 960K PPD. I suspect that could average out to over 1m PPD over time as I have not gotten many higher atom count projects at that speed and the low atom count projects pull the average down quite heavily.

I don't have an RTX 2060(s) to compare results with so you'll have to find someone who can provide those numbers to you.

[foldinghomealone2]

...We'd have to look at what kind of distribution the projects have. Likely all over the place...

The smallest GPU Project (14321) right now has 13,252 atoms
The largest GPU Project (14416) right now has 307,167 atoms
https://apps.foldingathome.org/psummary

With new GPU Projects (if/when) they are released, there's a possibility that the above values may change.

Currently with the available projects, the average atoms count is 105k.
However i think that WUs with less atoms are distributed more often as those WUs are completed faster. (But I might be wrong, maybe some projects are preferred over others which I don't know).
Therefore I think that the average atoms count of distributed WUs is less than 105k and it could be somewhere around what the 'real WU' from FahBench.

[bruce]

Proteins with more atoms are always slower that proteins with fewer atoms. Go ahead: Compute all the forces between each pair of atoms and apply F=MA where the forces on each atom are a sum of forces exerted by each other atom. Divide by the mass and you'll have the acceleration on that atom. With twice as many atoms, you have 4x as many forces to add up. but only twice as many atoms to move because there are 4x as many pairs of atoms.

[MrFrizzy]

Currently with the available projects, the average atoms count is 105k.
However i think that WUs with less atoms are distributed more often as those WUs are completed faster.

In my experience, it isn't the atom count alone that determines how fast a WU will complete on a given GPU but rather a combination of atom count and the number of steps in the given project.

Small atom count:
Take for instance p14540. This is the lowest PPD project I have tracked at about 804K. This project had 28,233 atoms yet took 3h:43m:38s to complete.
Project 14549, on the other hand, has the same atom count but with an average PPD of 943K and completes in 2h:47m:30s on average, nearly an hour faster.
Even faster still is p14538 with an an atom count of 29,004 which averaged 877K PPD and completes in 1h:27m:02s on average, well under half of the first project mentioned.

Average atom count:
Comparing those to a project with an atom count closer to the average, say p11745 with 110,370 atoms, it averages 1.22m PPD and completes in 1h:29m:38s on average.

Large atom count:
Moving to the large side of things, p14415 with an atom count of 290,453 averages 1.36m PPD and completes in 3h:37m:21s on average.
Also on the large side of things, p14539 with an atom count of 227,375 averages 1.25m PPD and completes in 1h:24m:44s on average, the quickest of any project I have tracked to date.

So as you can see, it doesn't matter if the atom count is small or large, the time it takes to complete the WU can vary greatly! That isn't to say that small atom count projects don't run faster when talking the number of steps per minute, they are most certainly faster in that regard, but as you likely know, one project could have 250K steps and another could have 8m. If all projects had the same number of steps, I would wager that projects would complete faster the lower the atom count is.

Granted, this could all change for GPUs with fewer stream processors/CUDA cores. I do not know how they would behave when working with small atom count projects versus large atom count ones. I could definitely see large atom count projects being bottleneck'd and running slower PPD/TPF wise but I don't have any personal experience with anything less than a 980 Ti. It's been far too many years since I was folding with GTX 570s to remember those stats.

[foldinghomealone2]

Ok, thanks for clarifying. I was not aware of that in detail