Alder Lake results

[zotric]

I've just run a quick test with folding for Alzheimer's.
CPU 12900K, Memory 4X8Gb DDR4 3200 MHz, On die graphics only, MSI Z690 Edge WiFi DDR4, 850W PSU (OK this is overkill), Arctic Liquid Freezer II 240 cooler.

Out of the box, the 12900K with Windows 10 seemed unwilling to use p cores for folding. This kind of behaviour is not unexpected for Windows 10 because the scheduler has not been updated for Alder Lake.

Process Lasso is free scheduling utility suggested by PaulTV. With Lasso, by setting CPU Affinities to p-core only for fahcore_a8.exe, I got it folding with all p-cores (no e cores).
It then returned about 400,000 ppd. and I think it can do better because it was throttling at Tj. Max (100 degrees).
Power was around 200W measured by Core Temp 1.17.1.

Unfortunately folding seemed to hang when just under 100% complete. When I restarted the PC it went back to zero % with no project selected so I do not really know if it finished or not.

With the Arctic Liquid Freezer II 240 cooler, which is a good AIO cooler, the CPU easily reaches 100 degrees with core clocks around 4.8-4.9GHz. There it throttles, so I would tend to turn down the power a bit.
I currently think the Liquid Freezer II 360 is the way to go for maximum performance at a reasonable price without killing the CPU.

For comparison an RTX 3070 gives about 4.5M ppd at about 200W (running on a separate i7-10700K PC).

[toTOW]

Unfortunately folding seemed to hang when just under 100% complete. When I restarted the PC it went back to zero % with no project selected so I do not really know if it finished or not.

What error do you have in the log ?

[zotric]

Apologies. I think we can assume all is well.
I just took the web client as being definitive and probably indicating failure.

According to the logs both of the runs runs I have done started (searched all of them for DOWNLOAD error:NO_ERROR") and completed successfully (see below).
The logged upload percentage did not reach 100% on either so perhaps that is why the web client did not show 100%.

19:16:43:WU00:FS00:0xa8:Completed 125000 out of 125000 steps (100%)
19:16:45:WU00:FS00:0xa8:Saving result file ..\logfile_01.txt
19:16:45:WU00:FS00:0xa8:Saving result file frame53.gro
19:16:45:WU00:FS00:0xa8:Saving result file frame53.xtc
19:16:45:WU00:FS00:0xa8:Saving result file md.log
19:16:45:WU00:FS00:0xa8:Saving result file science.log
19:16:45:WU00:FS00:0xa8:Saving result file state.cpt
19:16:45:WU00:FS00:0xa8:Folding@home Core Shutdown: FINISHED_UNIT
19:16:45:WU00:FS00:FahCore returned: FINISHED_UNIT (100 = 0x64)
19:16:45:WU00:FS00:Sending unit results: id:00 state:SEND error:NO_ERROR project:18206 run:8904 clone:0 gen:53 core:0xa8 unit:0x00000000000000350000471e000022c8
19:16:45:WU00:FS00:Uploading 27.87MiB to 128.252.203.11
19:16:45:WU00:FS00:Connecting to 128.252.203.11:8080
19:16:51:WU00:FS00:Upload 9.64%
19:16:57:WU00:FS00:Upload 21.31%
19:17:03:WU00:FS00:Upload 32.30%
19:17:09:WU00:FS00:Upload 43.74%
19:17:15:WU00:FS00:Upload 55.18%
19:17:21:WU00:FS00:Upload 66.84%
19:17:27:WU00:FS00:Upload 78.28%
19:17:33:WU00:FS00:Upload 89.94%
19:17:38:WU00:FS00:Upload complete
19:17:38:WU00:FS00:Server responded WORK_ACK (400)
19:17:38:WU00:FS00:Final credit estimate, 16794.00 points
19:17:38:WU00:FS00:Cleaning up
20:16:20:55:127.0.0.1:New Web session

[JimF]

For a bit of comparison, my Ryzen 5900X (Ubuntu 20.04.3) is doing 478 kPPD on P18419.
That is a bit on the high side, it usually runs a little over 400 kPPD.

I can't measure the power directly, but it is running at 78C according to Psensor, with an Arctic II 120 mm (single fan) cooler.

[FaaR]

This is interesting, my 9700X folding at 17 threads and a nosehair under 4GHz all cores/165W package TDP (any more and it hits upwards of 95C on air cooling and likely phantom throttles - it runs 88-90C as it is typically), and manages between 220k and 300k roughly depending on the work unit.

One would really like to see a Win11 result too for comparison, you don't feel like "upgrading" maybe perhaps...? In the name of science, pretty please? lol If maybe the E-cores could be used for folding as well then, they should be able to as I believe they support AVX2.

[VRoman]

Here is my F@H setting. 12900K running at 0.900 volts. Using P-Cores only @ 3.8GHz this CPU generates about 370-375k PPD with a 16-thread task and consumes 78 Watts. P-Cores are very-very efficient at low voltage!

Comparing to my old 8600K @ 4.3GHz on a 6-thread task (same task on both CPUs), both CPUs were consuming 65W. 12900K with P-Cores only was exactly twice faster than Skylake CPU. Wow! Impressive!

P-Cores @ 3.8GHz vs E-cores @ 3.2GHz with 0.900 volts. P-cores are 2.4 times faster and consume 1.55 times more power. Surprisingly, P-Cores are way faster and more power efficient at the same time in F@H. I was using same 6-thread task and assigning different cores in task manager.

My 5950X runs at 3.6GHz with 0.900v and generates a little over 500k PPD using 30-threads. Power consumption is 95 watts.

The fun part of being a beta-tester for Intel: assigning a 22-thread task to all P-Cores and 1 E-Core results in F@H running on just ONE SINGE E-Core and all P-Cores idling.

[FaaR]

The fun part of being a beta-tester for Intel: assigning a 22-thread task to all P-Cores and 1 E-Core results in F@H running on just ONE SINGE E-Core and all P-Cores idling.

Dayum.
I was hoping it would have handled that situation better!

Alder Lake's probably not ideal for F@H due to the way the software synchronizes work threads with each other combined with the CPU's asymmetry, with some cores supporting multithreading and some not supporting it, and some cores being inherently much faster than other cores.

I'm thinking that when utilizing the whole CPU, I'm thinking folding should ideally be split into one WU using only P-cores, and another WU using only E-cores. Something which the current client cannot automate, which makes it a total drag. (Not that you can blame the client of course... Alder Lake is such a wonky concept as a CPU in the first place, and is also the first time we see something like this in the PC marketspace.)

That, or one should probably disable multithreading, but even that wouldn't be ideal; P-cores would likely periodically sit idle while waiting for E-cores to finish up.

[Frontiers]

This is interesting, my 9700X folding at 17 threads and a nosehair under 4GHz all cores/165W package TDP (any more and it hits upwards of 95C on air cooling and likely phantom throttles - it runs 88-90C as it is typically)

7900х have thermal paste under his heatspreader. Deliding it and replacing thermal paste with some good liquid metal will improve thermal conductivity in heat pathway and will drop temperatures by 10 or more C depending on aircooler capacity.
You can check google or youtube for deliding guides.

[jchang6]

my Alder Lake 12900K seem to do 90K PPD running on all P & E cores, come with 270K PPD for Comet Lake 10850K 10-core, I think my Tiger Lake system is 180K. this is very much at odds with what is on folding.lar.systems. I may be running at less than full power

[Joe_H]

my Alder Lake 12900K seem to do 90K PPD running on all P & E cores, come with 270K PPD for Comet Lake 10850K 10-core, I think my Tiger Lake system is 180K. this is very much at odds with what is on folding.lar.systems. I may be running at less than full power

What OS are you using with your 12900K? Utilization of the cores may depend on whether the OS has scheduling support for the Alder Lake P&E cores. Characterizing usage may be difficult with 8 P cores with 16 threads and 8 more threads available through the 8 E cores. You may see near the highest throughput using 8 threads for a WU with each assigned to a P core. Adding more threads may get a bit more performance using HT threads or E core threads, but those will be slower and delay processing of the faster threads.

[jchang6]

Windows 11. I will look into this later. right now, the 12900 machine is off. I have just upgraded to power supply on this system to 1000W and will wait to get a 4090. My intent is to run just the GPU and the CPU contribution is too low to be of interest. the 10850 has a 3080Ti. frequently on reboot, it kicks off the CPU even though it is set to idle. This does not occur on the Skylake system with a 2080 Super.
I may upgrade the power supply on the 11700 system and then upgrade the GPU, currently RX 5600. Not sure which GPU to get for this, 4090 or 4080.
I hope there will be a new FaH package properly optimized for the P/E cores. I will want to get a Raptor Lake 8P/16E along with the latest Zen + 3D sram. I am thinking the FaH memory footprint might largely fit in SRAM.
I should point out the 12900K system has G.SKill memory base 2133MT/s, and XMP 3600 16-19-19. the 90K PPD was in 2133MT/s mode. PPD is actually much lower in XMP 3600MT/s, so there must be something going on

[Joe_H]

There was an attempt during the coding of Core_A8 to handle a mix of P and E cores as it included being used for ARM processors. The settings to compile the code to handle that mismatch in performance were turned off in early testing when they ran into problems doing that. Hopefully a later codebase will handle this better.

[toTOW]

my Alder Lake 12900K seem to do 90K PPD running on all P & E cores, come with 270K PPD for Comet Lake 10850K 10-core, I think my Tiger Lake system is 180K. this is very much at odds with what is on folding.lar.systems. I may be running at less than full power

If your Alder Lake is running a single CPU slot on all cores, it will fold at the pace of the slower ones (E cores) ... manually set the number of threads on the CPU slot to the number of P cores.

[jchang6]

running with 8 threads, CPU/OS seems to put the active threads one on each of 8 P cores, good for 140-150K PPD. at 16 threads, PPD seems to be higher but not guaranteed to run one thread per each of the P logical processors,
with FaH at 16 threads, and Fah affinity set to the 16 P logical, currently showing 400K PPD, but this involves changing during run, and may not be true PPD. unforunately, FaH does not respect thread affinity between runs?

[PaulTV]

FAH can't set affinity itself (hopefully that'll come with client version 8). People seem to get good results with process lasso though. There are some posts about that on the forum.