18263

[soa_rru]

Wow! Certain runs of these wu's are literally killing my ppd on my 5060ti

Nearly every one (current 181,5,19) has a TPF just over 8mins with a poor 3mill ppd.

Now for context my 4060ti is also running one (137,0,50) & TPF is just over 6mins with around 4.7mill ppd.

Anyone else noticed this or am i the only one?

[foldinghomealone]

Wow! Certain runs of these wu's are literally killing my ppd on my 5060ti ... has a TPF just over 8mins with a poor 3mill ppd.
Now for context my 4060ti is also running one (137,0,50) & TPF is just over 6mins with around 4.7mill ppd.

On my oc/uv 5080, 18263 has a TPF of ~1:50, giving ~28Mio ppd,
PPD is similar to all other recent projects, except 15xxx, which tend to make (only) ~23Mio ppd

Is it a typo or is a 4060ti really that much faster than a 5060ti?

'Strange' thing is because it was stated that 18263 is only released to fast GPU systems.
It seems, 5060ti and 4060ti are considered fast!

[muziqaz]

18263 is not exclusive to fastest GPUs. They are enabled on broad variety of GPUs, GPUs which are able to finish those WUs in time. The only requirement for it is amount of RAM available when requesting a new WU

[foldinghomealone]

We've made each WU in this project a bit longer than the others in the tau series to cater to higher end GPUs who were churning through these WUs especially quick.

viewtopic.php?t=43340

[muziqaz]

We've made each WU in this project a bit longer than the others in the tau series to cater to higher end GPUs who were churning through these WUs especially quick.

viewtopic.php?t=43340

This project is not exclusive to high end GPUs.
What Justin is saying there is- that in original form, individual WUs were to short for 5090s or the world, so these GPUs would have TPF of 3s or less. So the length was increased (the longer the WU, better for researcher) to not make them to be completed very quickly by high end and not to hammer the servers with constant download/upload requests.

P.S.
For clarity:
The size of the project simulation is measured by number of atoms it contains (in simple terms).
The length of the WU is measured by femtoseconds, or for our (donor) eyes number of steps we have to fold (steps != Femtoseconds).
Length of the project simulation is number of runs, clones and gens (RCG)

[AZBrandon]

18263 seems to work great on AMD though, which is a welcome change from the norm.

[soa_rru]

Some more numbers...

The 5060Ti is definitely faster than the 4060Ti BUT the 4060Ti is more consistent....

5060Ti (8gb)

RCG: 184,5,38 - TPF: 8m 08s - PPD: 3,065,769
179,7,32 - 8m 04s - 3,104,186
63,7,27 - 8m 00s - 3,139,489
19,7,37 - 4m 29s - 7,462,220
249,1,37 - 4m 43s - 6,930,546
82,4,35 - 4m 33s - 7,298,236
76,8,30 - 8m 03s - 3,109,778
75,3,23 - 4m 31s - 7,419,518
178,0,24 - 8m 03s - 3,107,012


4060Ti (8gb)

83,0,31 - 6m 10s - 4,631,647
137,0,50 - 6m 09s - 4,657,835
86,2,14 - 6m 09s - 4,658,277
203,7,0 - 6m 09s - 4,646,877

[AZBrandon]

OK so I have a couple related questions I guess. I see in the 18263 description it says:

Accordingly, we have implemented a minimum system memory requirement of 8000 MiB to run 182[51-56,58,61,62]. and 12000 MiB to run 182[57,60]

The description covers 18251 - 18262 but does not include what the memory requirement is for 18263. Is 18263 one that was also supposed to only go to 12GiB cards, which could be a cause of slower performance on the 8GiB cards mentioned above?

Separately, my Strix Halo system has 8GiB assigned in the UEFI but it's all unified memory, so technically it has as much as it wants. I only see mention of system RAM in the FAH logs though:

13:58:20:I1:WU43: CPU: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S
13:58:20:I1:WU43: CPU ID: AuthenticAMD Family 26 Model 112 Stepping 0
13:58:20:I1:WU43: CPUs: 32
13:58:20:I1:WU43: Memory: 119.65GiB
13:58:20:I1:WU43:Free Memory: 103.64GiB
...
13:58:21:I1:WU43:Folding@home GPU Core27 Folding@home Core
13:58:21:I1:WU43:Version 8.2.1
13:58:21:I1:WU43: GPU info: Platform: OpenCL: AMD Accelerated Parallel Processing
13:58:21:I1:WU43: GPU info: PlatformIndex: 0
13:58:21:I1:WU43: GPU info: Device: gfx1151
13:58:21:I1:WU43: GPU info: DeviceIndex: 0
13:58:21:I1:WU43: GPU info: Vendor: 0x1002
13:58:21:I1:WU43: GPU info: PCI: 244:00:00
13:58:21:I1:WU43: GPU info: Compute: 2.0
13:58:21:I1:WU43: GPU info: Driver: 3661.0
13:58:21:I1:WU43: GPU info: GPU: true

Where does FAH report on GPU memory, and as such does it know how to handle the fact gfx1151 (Strix Halo) has basically all available RAM, and similarly does it have the 5060Ti 8GB listed separately from the 5060Ti 16GB model?

Sorry for such basic questions; it didn't appear many people are doing folding on integrated graphics, probably because iGPUs have been pretty weak up until Strix Halo came out.

[muziqaz]

FAH doesn't care about VRAM. The description mentions system memory (RAM).
viewforum.php?f=66