OTOY Forums

Hi all,
just wondering if the upcoming release of pci-e 3.0 cards will open up any doors as far as far as the ability to share vram between gpus goes. I know there are issues with bandwidth limitations and the amount of traffic that would be generated from sharing vram on pci-e 2.0/2.1, but pci-e 3.0 effectively doubles that bandwidth and increases the efficiency of handling data packets significantly. Is this improvement enough to implement shared vram? I am going to be building a new render machine soon for fairly complex scenes, I would like to know this before I commit fully to either waiting for pci-e 3.0 and making a machine that is more gpu dependent, or sticking with the cpu solution for now.

Thanks!

yamanash wrote:Hi all,
just wondering if the upcoming release of pci-e 3.0 cards will open up any doors as far as far as the ability to share vram between gpus goes. I know there are issues with bandwidth limitations and the amount of traffic that would be generated from sharing vram on pci-e 2.0/2.1, but pci-e 3.0 effectively doubles that bandwidth and increases the efficiency of handling data packets significantly. Is this improvement enough to implement shared vram? I am going to be building a new render machine soon for fairly complex scenes, I would like to know this before I commit fully to either waiting for pci-e 3.0 and making a machine that is more gpu dependent, or sticking with the cpu solution for now.

Thanks!

You should ask to Nvidia.
It depends by Cuda programming language.
For example iray and arion supports yet cpu processing and I think that this feature will be implemented soon in Octane.
I'm sure that soon we will see Teslas with 12 or 16GB and GTXs with 4GB.
Now I'm working with 3GB GTXs only and I can render very complex scenes. At the moment I can't render forests but with the implementation of instances this will be possible.

I just emailed them, I will post back with their reply as soon as possible in case anyone else is also curious about pci-e 3.0.
I really don't see quadro or tesla cards as a very good solution especially with the massive price hike bound to accompany such cards lol.
Ntm good luck finding full cover blocks or any kind of decent water cooling/overclocking solution for them.
Reading over the Cuda 4.0 pdf it says they now support direct gpu to gpu memory reading and writing between cards, which sounds like a step in the right direction..
I still don't know if the bandwidth exists with pci-e 3.0 to efficiently constantly read and write between 4 or more gpus though.