12129 is cruel to small GPUs

[ViTe]

I think that the problem is more about rewarding points. It's too low for this project. I wrote down some statistics for this and other projects that we have running now and ... surprise... many projects affected by the same issue. I have statistics from spring of this year for the same machine based on 1080Ti. Here what I found:

Early spring 2025:

12296 Score: 407000 TPF: 1:33 PPD: 3 780k
12704 Score: 135500 TPF: 0:30 PPD: 3 630k
15001 Score: 270000 TPF: 1:03 PPD: 3 700k
16525 Score: 1880000 TPF: 8:00 PPD: 3 700k
16770 Score: 455000 TPF: 1:48 PPD: 3 630k
16771 Score: 437500 TPF: 1:48 PPD: 3 500k
18236 Score: 295000 TPF: 1:07 PPD: 3 800k
18228 Score: 520000 TPF: 2:00 PPD: 3 750k
18238 Score: 817000 TPF: 2:55 PPD: 4 035k

September 2025:

12129 (Core24) Score: 1250k TPF: 7:45 PPD: 2 400k
12708 (Core27) Score: 366k TPF: 1:51 PPD: 2 850k
12709 (Core27) Score: 363k TPF: 1:53 PPD: 2 760k
13010 (Core24) Score: 368k TPF: 1:55 PPD: 2 780k
15000 (Core23) Score: 480k TPF: 3:07 PPD: 2 600k
15402 (Core24) Score: 303k TPF: 1:12 PPD: 3 670k
15404 (Core24) Score: 302k TPF: 1:11 PPD: 3 670k
18242 (Core26) Score: 350k TPF: 1:50 PPD: 2 750k
18243 (Core26) Score: 250k TPF: 1:17 PPD: 2 820k
18244 (Core26) Score: 375k TPF: 1:58 PPD: 2 750k
18929 (Core22) Score: 104k TPF: 0:33 PPD: 2 720k

This statistics based on multiple runs for every project and reflects average number of points. Project numbers are different but you can see that during the springtime all projects produced approx the same amount of points for 1080Ti - around 3.7 mil PPD. Now we have different picture. In september only 2 projects are at the same level as before: only 15402 and 15404 making around 3.7 mil of points. All other projects produce significantly less. So looks like point system got some corruption or nerf.
I see the same picture for 1070Ti as well - PPD significantly lower then I had for this GPU before.

[muziqaz]

Point system has not been corrupted or nerfed
Here is what's happening:
nVidia stopped supporting and improving their CUDA libraries for older generation cards. New drivers new performance improvements to newer cards. This negatively impacts old cards. This is natural, especially with cards so old as 1000 series.
Also, new simulation types, as well as openmm (API used for GPU cores) enhancements are more suitable for latest generation cards

[ViTe]

I don't think so. 15402 and 15404 are still at the same level of points and these projects running on Core24, same as slowest 12129. Also we have people with RTX 5090 in this topic who reported the same issue at least for 12129.

[muziqaz]

I don't think so. 15402 and 15404 are still at the same level of points and these projects running on Core24, same as slowest 12129. Also we have people with 5090 in this topic who reported the same issue at least for 12129.

I do think so
Whatever 5090 owners are reporting is irrelevant. They bought a GPU which is clearly not built for FAH simulations, those GPUs have other issues

[ViTe]

nVidia stopped supporting and improving their CUDA libraries for older generation cards. New drivers new performance improvements to newer cards. This negatively impacts old cards. This is natural, especially with cards so old as 1000 series.

This happened not so long ago. I don't believe in miracles. Nothing changing in FAH code / cores that fast.

The official release notes for Nvidia's CUDA 12.9 Toolkit explicitly indicate that the next major release will no longer support Maxwell, Pascal, and Volta-based GPUs. Note that this deprecation is only limited to the compute side,

Next major is release 13.

Release Notes
Release 13.0
NVIDIA Corporation
Sep 03, 2025

Looks like current build of Core27 dated Mar 27 2025. Do we have newer release?
Anyway, I'd rather wait for comments from mods or devs.

[muziqaz]

nVidia stopped supporting and improving their CUDA libraries for older generation cards. New drivers new performance improvements to newer cards. This negatively impacts old cards. This is natural, especially with cards so old as 1000 series.

This happened not so long ago. I don't believe in miracles. Nothing changing in FAH code / cores that fast.

The official release notes for Nvidia's CUDA 12.9 Toolkit explicitly indicate that the next major release will no longer support Maxwell, Pascal, and Volta-based GPUs. Note that this deprecation is only limited to the compute side,

Next major is release 13.

Release Notes
Release 13.0
NVIDIA Corporation
Sep 03, 2025

Looks like current build of Core27 dated Mar 27 2025. Do we have newer release?
Anyway, I'd rather wait for comments from mods or devs.

Good luck

[appepi]

by muziqaz » Sun Sep 07, 2025 9:05 am
Point system has not been corrupted or nerfed
Here is what's happening:
nVidia stopped supporting and improving their CUDA libraries for older generation cards. New drivers new performance improvements to newer cards. This negatively impacts old cards. This is natural, especially with cards so old as 1000 series.
Also, new simulation types, as well as openmm (API used for GPU cores) enhancements are more suitable for latest generation cards

Ummm.. I wonder if the "Points" documentation at https://foldingathome.org/faq/points/ needs to be updated? Like, it states: "We have a single benchmark machine, its most important component is its processor: a Intel(R) Core(TM) i5 CPU 750 @ 2.67GHz." and it would seem that the run-time for a project on this device sets the various parameters that feed into the Quick Return Bonus calculation, and ..."Note that GPU projects are now being benchmarked on the same machine, but using that machine’s CPU. By using the same hardware, we want to preserve our goal of “equal pay for equal work”. Our GPU methods have advanced to the point such that, with GPU FAHCore 17, we can run any computation that we can do on the CPU on the GPU. Therefore we’ve unified the benchmarking scheme so that both GPU and CPU projects use the same “yardstick”, which is our i5 benchmark CPU."

OK: Peering back through the mists of time to 2020, I I recall thinking at the time that the specification of FaH's "yardstick" device was very like that of my 2011-vintage Hewlett-Packard HP Pro 3130 Microtower PC - PC2 (W10 1903) and its Intel Core i5-750 2665 MHz (4 cores). It died before Geekbench 6 arrived, but its GB5 results were 474 for 1 core and 1573 for 4, as against my latest 2017-vintage Xeon W-2145 (8C 16T) at 1114 single core and 7922 multicore. The old i5-750 managed about 4.7 K PPD, but the CPU's burned so much electricity for so little reward that I used my relatively new GTX 1060 (6GB) GPU for most of the PPD I earned that year.

Anyway, it is hard to believe that Fah would still be calibrating everything on the same "i5 Benchmark CPU" since I shudder to think how many iterative applications of the battered i5-750 "yardstick" would be needed to work out the length of a Project 12129 WU, if indeed it could do so at all. It seems probable that at some point a new CPU "yardstick" would have been calibrated against the old one, and that a GPU yardstick would have been similarly calibrated so that with all their faults the RTX5090's of the world could be properly rewarded.

Be that as it may, while I can see how the enhanced compatibility of driver/MM in newer cards and drivers and cores would benefit the PPD of those newer cards, I cannot see how this would negatively impact older cards in absolute PPD - at least for the same core and project, and certainly not why it should if it does. But of course, being even older than the 1000 series of NVIDIAs, there are many things I fail to see clearly, so I look forward to clarification.

[muziqaz]

That URL is old and obsolete. None of us have that CPU for testing anymore and we are definitely not using just one CPU for point evaluation.
We are using a plethora of various GPUs for GPU project PPD adjustment, and loads of different CPUs for CPU project PPD adjustment

[appepi]

We are using a plethora of various GPUs for GPU project PPD adjustment, and loads of different CPUs for CPU project PPD adjustment

So the short answer is that :

"the "Points" documentation at https://foldingathome.org/faq/points/ needs to be updated