And all that, assuming you know exactly what you are doing and your final render will turn out just fine the first time around. Which is not my case. I’m a Blender novice and I’m barely confident on my skills with this tool, so I’m bound to make a lot of mistakes and would hate to pay money for each one of them. I’d rather have full control of the means of production, that way I will only have to worry about electricity bills.
Arexma, I’d like to ask you again your opinion about the GTX590. If unified virtual addressing gets supported by Cycles, do you think we will be able to fit complex scenes in the GTX590, like we would on a GTX580 with 3GB? I’m probably going to get me either a GTX580 or a GTX590 by the end of this month. For me, maximun complexity of the scene would be a priority over CUDA performance if I have to make a choice. Thank you for your comments.
You really have to ask this in the blenderdev channel on IRC, or try to contact brecht, farnsy and whoever else is working on cycles directly.
The way I understand it each GPU gets a virtual address range.
You can fill the full 3GB with data.
GPU1 has it’s native address range of 0-1.5 and a virtual address range remapping from 1.51 to 3.0.
For GPU2 it should be the other way round.
However I have no idea how the bus and the memory controller on a dual GPU board looks like.
It´s possible that there is no device connection between gpu1 and vram2 and you got to do a host operation, accessing the data via the system bus which is increeeedibly slow.
It’s really only intresting for rendering huge single frame stuff.
For animations I´d take 2*580.
I am sure at one point there’ll be a script or cycles will support it that you render all even frames on one card, all odd frames on the other card. Right now you just save your file once for the first half of the animation, once for the second half, start blender twice and let each instance of blender render on one of the cards.
And as I already mentioned, I am certain nvidia will find another blunt weapon and hit your knee as hard as they can
After all in SP a GeForce is much faster than a Tesla, the only advantage for Raytracing a Tesla has is the huge amount of memory.
Before I´ll invest into further GPGPU raytracing hardware I wait for the GF700 and see what’s happening.
If I had to get something now I´d look for a used gtx470 or gtx480. Sure they are loud and hot, but they do the job for reasonable money.
