13429 low ppd on fiji XT

[Yeroon]

Of the two 13429 wu that I witnessed running on my Fury Nano, both have been ~1/6th of its normal ppd. Current TPF ~1min04s, est ppd 127k. Normally in the 600k-800K range.
rx470 in the same comp, doing another 13429 at ~215k ppd.

Looking it up it seems to have a really low atom count, which seems odd that a fiji xt would get such a low count? If these are going to persist for a while, I can swap it out for another 470.

[bruce]

Is the Radeon RX 470 device 0x1002:0x67c0 or device 0x1002:0x67df?

Both appear to have 2048 cores and are currently classified as FAH GPUSpecies 5.
The fiji xt is device 0x1002:0x7300 and has 4096 shading units. It is also classified as FAH GPUSpecies 5 so as far as I can tell both should receive assignments from the same pool of projects.

A project is currently under way to benchmark and reclassify all GPUs with more appropriated GPUSpecies. There's no ETA as to when that project will be rolled out to the servers.

[PantherX]

FYI, Project 13429 had its base point adjusted since there was a mistake. Thus, see what your system performs with the new base points

[NormalDiffusion]

Please remove the high end AMD cards from this project!
Vega Frontier Edition LC bringing barely 300k PPD instead of 1.3M PPD and Radeon VII 260k instead of 1.6-1.8M.
It's just a waste of electricity!

Edit: barely better with the adjusted credits... 350k for the Vega FE LC... still 250k for the R VII.
And there is definitively something wrong with the project on the big AMDs: the R VII sees utilization of 96%, but the GPU is not going up in speed (staying at 1.3 MHz) and drawing 40W...

[Yeroon]

I had a very well timed WU completion to let me swap out the nano for another 470, so I just did that for now.
GPU useage is ~50% power compared to higher atom count WU, ppd seems to be in that drop as well for the 470s (with new base points)

I'm fine with doing these low ppd high priority wu on the 470s, it just seems like a waste to do them on the nano - or it seems any amd gpu higher than polaris.

[miker07]

I am having the same issue on my Radeon VII's. There is definitely a major issue with this wu and some AMD gpus.

[NormalDiffusion]

Following the announcement from JohnChodera, "13429 is the ligand in solvent, and restricted to narrow GPUs".
The Fiji xt, R VII, Vega Fe... are by no means "narrow GPUs" and shouldn't be part of 13429 but 13428.

[muziqaz]

None of the AMD GPUs should be getting those. It has been mis-configured most likely

[BobWilliams757]

They run fine on the onboard Vega 11, with PPD return 50% or more above average. There did seem to be some variance in PPD return/TPF. I ran three in a row and all varied.

[bruce]

In the past, there has always been variations between Run M and Run N within project P. The moonshot sprint evaluates/screens a variety of proposed compounds. They're divided into two groups, one of which tends to work well on GPUs with a multitude of shaders and the other of which works better on narrow GPUs (with low shader counts). Fitting every compound into those two categories is a challenge, but FAH can't really afford to assign a different project number to every compound.

[NormalDiffusion]

None of the AMD GPUs should be getting those. It has been mis-configured most likely

And now the Radeon VII, Vega are getting them again... <400k PPD...

[muziqaz]

Someone opened the floodgates so all the GPUs are getting that project

[NormalDiffusion]

Someone opened the floodgates so all the GPUs are getting that project

Instead of assigning 13428 to those AMD cards...
Funny that no "fast" Nvidia GPU get those 13429...

[muziqaz]

When I said all the GPUs, I meant even fast nVidia 13428s have been probably finished, so we are getting some leftovers 13429s

[Yeroon]

I was fine with doing them on my 470's, which I still have both in that system. My guess is there are not enough low CU gpus available for the amount of WU that need completion?