WU 16904 VERY low PPD

[cine.chris]

PPD in the 700k range on a 2060s. Utilization ~75%, high interface utilization, low FB usage when compared to identical 2nd GPU. 2nd GPU is running at 1,627K PPD.

[bruce]

Proteins being analyized may be large or small. They are assigned (mostly) randomly to GPUs which also come in various configurations. A large protein necessarily has many calculations which can be done in parallel by GPUs that have a large number of shaders. On the other hand small proteins have very few calcluations whicn can be done in parallel and all the serisal protions still have to be performed. Your 2060 is great at processing large proteins and not so good working on small proteins.

Project 16904 has only 32,000 atoms.

There's an open ticket for an enhancement to optimize the assignments.
https://github.com/FoldingAtHome/fah-issues/issues/1479

[cine.chris]

great at processing large proteins and not so good working on small proteins.
Project 16904 has only 32,000 atoms.

Interesting.
Is there any advantage to the project for donors to be running a mix of hardware... GPU/CPU, # of cores? Smaller GPUs?

[bruce]

The smallest proteins work best on a CPU slot. Intermediate sized proteins work well on low to intermediate GPUs. Large prtiens work well on the latest (and widest) GPUs.

Unfortunately, you don't get to choose.

[Ichbin3]

Same issue here with 16907@2080Ti.
This low-atom-projects should not be assigned to hight end gpus.
As I'm informed there are some possibilities to assure this, no?

[bruce]

Not yet

[Frontiers]

With another project from 1690x bunch, PRCG 16908 (0, 8, 13), mine have all time antirecord - FAHCore used only 40% of board (not chip) TDP for only 50% of usual average PPD.
It may be good if next versions of OpenMM and FAH_cores would be modified for crunching such small atom count projects parallel in pairs, 2 simultaneously at each GPU. Or even 3 or 4 in parallel with powerful GPUs with very high CU counts.

[bruce]

You can suggest that enhancement here: https://github.com/FoldingAtHome/fah-issues.

Unfortunately I suspect the relatively small WUs that would warrant such a combination will probably end up finding a way to be assigned to slower systems including the soon-to-be-supported ARM systems. I bet they'll get pretty good utilization on my smart phone or somebody's Raspberry Pi, even if they're somewhat slower than on your RTX device.