When I refer to poly count or texture limits I refer to RAM limits of course :)
We had a project that we rendered in iRay back in the iRay 2.0 times (a bit before the release, we used a beta release for it) and we were able to render that project thanks to iRay because the quality required was prety high and with Maxwell it was a non sense, there was no other option at that time, Octane was too inmature, Arion was a bit slower at that time, and that was all.
We managed to finish the project (it's in our site) but we had to lower the polycount because it could not fit in the 3Gb of the 580 so wehad to lower the subdivisions and you can notice that in some shot.
We are talking about a video of an industrial machine, nothing more on the stage, so it wasn't a complex interior scene or anything, the project we are doing right now (and some of the latest ones) could not fit in 6 or 8 Gb of ram at all, that is for sure, we have several STP models, we have several OpenSubDiv models at pretyt high res, the scene is around 20 million polygons, sometimes a bit more, we can't fit this in a GPU, and maybe those 2 computers with 4 GPU each are great for that project, and you may use them for that one project, but what happens when you need to work in more complex projects?
We have a pipeline, at first GPU seemed to be pretty great because of speed, but in the end, in the majority of projects 6Gb's of GPU ram are not enough :P at least for us, plus you have to add the lack of TONS of features in GPU render engines, several AOV's, depending on the engine different features like Volumetric rendering, etc... we don't like to be constrained to a sub set of features, right now in Corona we are more or less contrained in features, but not so much, we can deal with almost anything, we can't say the same with GPU render engines, at least I speak specially for iRay, Arion has a pretty big and good feature set, but it's more aimed towards other markets and it's evolution is not as fast as I would like.
So that is my storey and why we abandoned GPU render engines (amongst other things) we need speed but we also need reliability to realize our projects, no matter wich type of project is it, and the investment for a proper GPU farm is too high, at least for us, and of course the power consumption is massive, just thinking in having 10 computers draining power with 2 GPU's each one, the energy invoice the month we did the industrial machine video was around 700€... I've never received such invoice using CPU, we've been all month rendering 24/7 with 10 computers and the invoice for 2 months is going to be 170€ ... it's a pretty big difference.
Cheers!