Some projects work better on some hardware

[HaloJones]

I'm sure this is not new for some but I recently tried moving all my clients to using Advanced and noticed an anomaly.

Two 1070 clients, one 1060 and one 970 all increased their PPD significantly

but my TitanX reduced its PPD by around 20%.

I have a particular machine with a 1070 and a TitanX and historically they've produced very similar numbers - usually just north of 700K each

With both running a 14158 at the moment, the 1070 is claiming 790K but the TitanX is showing only 613K.

Is there something about the current crop of Advanced units that doesn't scale across more GPU cores?

[bruce]

FAH has never scaled across the number of GPU cores. That's why FAH was originally benchmarked with specific hardware with a disclaimer attached that performance of other hardware wouldn't necessarily be proportional.

For a small number of shaders and a large number of atoms, all of the shaders will spend most of their time working in parallel. For a large number of shaders and a relatively large number of atoms, the percentage of time spend getting the data to/from the GPU becomes more significant and the percentage of time when all the shaders are working in parallel gets smaller because the time spend preparing a block of work to use the shaders doesn't go down when you add more shaders.

Looking at it another way, (for all you gamers) the newer/bigger/more-powerful GPUs are great when you have multiple monitors with high resolution screens and with higher numbers of colors but that capability is wasted if you have a single low-resolution screen running with a reduced number of colors.

[Theodore]

I presume the client does some sort of selection, based on your hardware, before releasing a work unit.
There are projects that are released recently, and still aren't calibrated correctly, which could also account for lower or higher performance numbers.

[bruce]

The most important step in WU selection is
1) What WUs are available at the time your client requests a WU?
and
2) Can your hardware process this WU?

Many of the new projects happen to be better suited for mid-range GPUs than for the top-of-the-line GPUs. That changes based mostly on scientific need so it's not expected to be that way forever.

Also, we've recently had a problem with lots of WUs being drained -- but that has been circumvented.

Please tell us your GPU and which specific projects you're concerned about?

[HaloJones]

I've recently put the clients to Advanced and found that P14163, for example, loves my 1070s (840K ppd) but hates my TitanX (625K) (both under Linux on the same machine. The 1070 is driving the screen.) Now, my TitanX just got a P14167 and is hitting 1:37 TPF for an estimated 960K ppd.

So there you can see how a TitanX can go from 625K on a P14163 all the way to 960K ppd on a P14167.

These are just examples obviously.

[MeeLee]

I've recently put the clients to Advanced and found that P14163, for example, loves my 1070s (840K ppd) but hates my TitanX (625K) (both under Linux on the same machine. The 1070 is driving the screen.) Now, my TitanX just got a P14167 and is hitting 1:37 TPF for an estimated 960K ppd.

So there you can see how a TitanX can go from 625K on a P14163 all the way to 960K ppd on a P14167.

These are just examples obviously.

I'm not sure what I'm about to say, but it appears like a lot of the 14k projects have higher rated scores.
They seem to me to be some sort of 'verifying' rather than full on emulation.
Some of my mid-ranged cards (2060) will show lower PPD, and consume considerably lower power (150Watt instead of 175Watts).

[Mstenholm]

So there you can see how a TitanX can go from 625K on a P14163 all the way to 960K ppd on a P14167.

I have 1070s, 1070ti, 2060s and 2070s and the one thing they all have in common is that 14167 is the best and 14158 the worst in terms of PPD. The 14158 is around 80% of average of WUs for that particular GPU and the 14167 as high as 25% above. The new 14163-14165 is 130 W on a 2070 and the 14167 is 176 W.

[toTOW]

Look at the number of atoms in those two projects, and you'll quickly understand why they perform in two opposite ways on you big GPUs ...

[MeeLee]

Why would they even get assigned to big GPUs?

[bruce]

Why would they even get assigned to big GPUs?

Assignments are made based on the needs of FAH's RESEARCH. Projects that carry a high value for the science being studied isn't necessarily related to the number of atoms in the protein being studied. As long as a big GPU can benefit the research, then it's eligible for the assignment.

[toTOW]

And there's no perfect classification of GPUs that would work in every cases ...