Bonus point system kind of balanced unfairly?

[Theodore]

Hypothetically speaking,

If 2 slower cards on a PC, would have been assigned the same work unit (split to both cards), as a PC running a faster card,
Is there a possibility that both cards can get the same amount of work done, in the same amount of time, than the faster card?
If so, will they receive exactly the same bonus points?

I believe, with current FAH setup, that there's a chance that both cards (eg: 2x GTX 1070) can get more work done than the faster card (eg: RTX 2080) , in the same or even a shorter amount of time; and yet they still are rewarded a lower bonus and overall PPD.

This in part, because both cards are not contributing to the same WU, but are contributing to different projects.

Likewise,
If an RTX 2080, instead of running 100% to 1 work unit, was split in to running 2 work units simultaneously, it's bonus points as well as overall PPD would drop dramatically!

For this I wondered if bonus points are awarded per Graphics card, or per PC?

If they were awarded per card, the current bonus system could seem unfair towards those that are running multiple GPU systems; as they could be contributing more, and get rewarded less.

[JimboPalmer]

A given project contains many generations of sequential WUs, each sequence completes when the slowest WU completes. So F@H rewards you for not being slow.

If each WU stood alone, I am sure F@H would be rewarding other ways, but given that the science is held up by the slow WUs, it is not surprising they use a Quick Return Bonus, favoring the quick.

[bruce]

The FAHCore processes the WU using the resources of a single GPU plus one CPU thread. While, in theory, the processing using two identical GPUs (like a pair of GTX 1070s) might be possible, the OpenMM software doesn't contain the the logic to manage distribution between those two distinct devices.

I don't think adding that logic makes sense. First the percentage of donor systems that have a pair of identical GPUs is small enough to warrant spending development resources elsewhere. Second, if we take the simplest of cases, and assume that half of the protein should be allocated to each one, that adds a step of coordinating those two halves every time a step is processed, reducing efficiency. Moreover, if other processes are using your GPU, then one of them will be slower than the other so the faster one has to wait for the slower one to finish. Third, when a large number of shaders are assigned to a small protein, the GPU(s) become less efficient. The RTX20xx GPUs really don't do as well on small proteins as they do on large proteins, so you'd artificially be making that problem worse.

RTX 2080Ti has 4352 shaders
Two GTX 1070s have 1920+1920
so if small or medium sized proteins are assigned, the pair of 1070s would become more efficient by running two WUs that if they somehow shared the same WU.