OTOY Forums

This could be a game changer if its implemented in octane I recon http://www.theregister.co.uk/2013/11/16 ... ory_party/

From what I understand, Cuda's "Unified Memory" has been misreported and does not mean that the GPU will treat system memory as slower form of VRAM. Instead it streamlines the process of memory management for developers. I could be wrong though.

I hope you are wrong High hopes I guess.

I have a dream......
one day, we will be able to pop additional VRAM into our graphics cards like into our mainboards....

This dream is old...
In the 90th, the Matrox Millennium had a 6MB addon module to expand the VRam

face

Actually, I think witpapier is right.
If you look at nVidia's own example here:
http://devblogs.nvidia.com/parallelfora ... in-cuda-6/

Perhaps someone with a deeper understanding of the process mentioned here can chime in?

CUDA toolkit 6 just seems (it's not released yet) to hide the complexity of shuffling data from and to the GPU. It doesn't magically solve the problem that you have to shuffle data between CPU and GPU. Data is sent/received to/from the GPU via the PCI bus which is orders of magnitudes slower than than the memory transfer directly from VRAM to the GPU.

-> So even it may look like you are using one big memory blob, spanning the host and GPU RAM, you are not and I expect the performance loss to be the same as with previous toolkits. The only difference is, it's easier to program.

abstrax wrote:CUDA toolkit 6 just seems (it's not released yet) to hide the complexity of shuffling data from and to the GPU. It doesn't magically solve the problem that you have to shuffle data between CPU and GPU. Data is sent/received to/from the GPU via the PCI bus which is orders of magnitudes slower than than the memory transfer directly from VRAM to the GPU.

-> So even it may look like you are using one big memory blob, spanning the host and GPU RAM, you are not and I expect the performance loss to be the same as with previous toolkits. The only difference is, it's easier to program.

Thanks for the insight Abstrax!
Seems that unified memory is not the silver bullet that we were perhaps hoping for.

But I think this feature would be very helpful in implementing hybrid (CPU / GPU) CG algorithms .. or algorithms that are difficult to be implemented purely on GPU