So their out-of-core streaming enables pretty much unlimited texture amount ? Did it also bypass the texture amount limit CUDA previously had (and which still seems to be case for Octane or not?)
Basically, yes. In a simple test I just did, using a 2k by 1k HDRI resulted in a whopping 556 kB memory on the GPU used for textures. It's obviously only loading the parts it needs. It doesn't have a "max number of texture" like Octane. Presumably performance might suffer if you have thousands upon thousands of images, but there is no set limit as far as I know.
16mil. polies for 3GB vram is nice, but 16mil. is still nothing. How does it go around displacement ?
In the same simple test I mentioned earlier, I rendered 38 million (unique) triangles on a card with 1.6 gigs of free memory. Out of the 1.6 gigs available to Redshift, the texture cache used up 128 megs, and various other things accounted for a bit more. In the end, there was 1.2 gigs available for geometry, and it used 1.1 gigs for 38 million triangles. It stands to reason that if you had a 6 gigabyte card used
just for rendering (to save from Windows' overhead), you could fit about 190 million triangles before worrying about going out of core.
How does it go around displacement ?
Displacements are generated on the CPU and resulting triangles are streamed to the GPU as needed, same as with regular geometry. It doesn't do texture-space displacement rendering, as far as I know (like V-Ray's 2D displacement effect).
Octane isn't great. I'd rather use Corona. It's faster, more reliable, easier to use, and produces better results.