Want to upgrade to a RTX GPU for rendering and like to hear your opinions. Is it enough to purchase a 2000 series for a significant improvement boost? Or invest into 3000 series right away?
Also can you mix 2 different Nvidia cards, like mine 1070 and a 2080 for example?
Thanks for any input!
Check the vray benchmark list. It will be similar to Arnold.
Ok and how about the mixed cards example? Is that possible or not a good idea?
You cannot mix 1070 with 2080 (or 3000 series) because one has ray tracing cores and the other doesn't. Other types of mixing might work, but it's currently not recommended because arnold will be bottlenecked by the slower of the two cards since it currently divides work evenly between the two GPUs.
Suppose one GPU renders in 120s and another GPU in 50s. Working together, with each GPU doing half the work, it would take 120s/2=60s for the slow GPU's share of work and 50s/2=25s for the fast GPU's work. So that means you would have to wait 60s for a dual GPU render, while if the fast GPU had just been used it would have taken 50s. Yes, that's a shame that in this case single GPU rendering would be faster than dual GPU and it is something we'd like to one day improve upon...
Thank you.
How about the GPU RAM, does it behave the same way or does it add up, for example 2x 3090 cards with 24GB. Will combining both give you 48GB of RAM?
See https://docs.arnoldrenderer.com/display/A5ARP/Arnold+GPU+FAQ for details on memory sharing.
Im sorry, but i dont get this line in the docs:
You can NVLink two cards, so your memory limit is determined by the third card. For example, if you have three cards with 16GB, the scene and textures have to fit in 16GB. The two NVLink cards combine for 32KB, but the third card is still 16GB, so the limit is still 16GB. Three cards are still faster than two, so it's a tradeoff.-----
Can i add memory of 2 cards with each 8 GB, so that in the end i have 16 GB of a memory pool in which Arnold will fit the scenes textures?
If you connect two cards with nvlink your memory sort of doubles. It's not perfect (not everything can be shared) but it will be more memory than from a single card.
What can´t be shared? Displacement maybe?
... do you advise connecting 2 cards? Sounds like a bad idea, when reading this.
I could be wrong, but I think some of the geometry related data, such as BVHs, can't be shared. There's also program state, which can't be shared. For the most part it does give you a reasonably nice increase in total memory. If you can connect them with nvlink, it's a good idea to do so.