For example using two graphic cards with e.g. 3GB VRAM, will lead to usable 6GB of VRAM, since the copy of data will have not to be kept in both VRAMS ? or am I completely wrong ?
Sadly, yes. It just means you do not have to code all copy commands from host memory to the device memory on your own. Basically the compiler makes it just easier to programm.
That’s a pity ;( it would be nice to share the available RAM between different devices and add all available ram together. I wish it will be possible once in the future.
I’ve read someplace, dont remember what tech site it was, that the newest Batch of AMD apu’s are already capable of much of this. I think some of them already capable of using GDDR memory, if a motherboard manufacturer decides to make one.
Although technically true, this sounds like too much of an understatement of what Unified Memory brings. The effort of creating your own virtual memory solution is immense, which is evidenced by the fact that none of the major GPU-rendering applications currently offer a fully out-of-core solution. Future hardware will also bring hardware support to outperform any such software solutions (like the current solution, which is also in software).
With Unified Memory, we will first see the hard memory limit gone, then specific optimizations can be applied to reduce the penalty of data spilling into system memory. For raytracing, ray-reordering and mip-mapping are two methods that can dramatically reduce the amount of memory required in-flight.
Let’s also not forget that Unified Memory allows you to directly use all of your host code data structures, with all the pointers intact. This can save you some data conversion steps and is once again a large productivity booster.
Here is a more depth look into the unified memory, Mark Harris, from NVidia talking
The CUDA 6 SDK is now available to registered developers. Cycles does compile in it, there aren’t any significant performance differences.
Unfortunately, I have to clear up a misconception on my side. The current implementation of unified memory does require allocating a fixed chunk on device (GPU) memory and therefore the hard memory limit persists. True unified memory is not going to be available until the next CUDA releases and will (to my understanding) require the “full” Maxwell hardware, to be released later this year.
Once again, my apologies for spreading misinformation.
OS X Cuda 6 is out
CUDA toolkit 6.0 now official versionhttp://lists.blender.org/pipermail/bf-committers/2014-April/043431.html
Shall I download Cuda 5.5 or Cuda 6.0 for Mac OSX ?
If you are wanting to build the latest version, it will need to be cuda 6.0