Does Pascal Unified Memory, mentioned in pascal whitepaper, supported now?

Hi dears.
I wonder does newest unified memory in pascal architecture now supported? The paper says this technology must be with operating system support and under construction with redhat and linux community.

what’s the news? i would be greatful for your help.

any one knows?

Pascal unified memory with oversubscription is supported. (oversubscription: on a Pascal device, you can do e.g. cudaMallocManaged for more than the physical RAM on the GPU)

The ability to use a system allocator to do the allocation is not widely available yet, AFAIK, as it depends on forthcoming developments in the linux kernel

So, Do the feature which data synchronize by page faulting in unified memory, and feature CPU and GPU access unified memory allocations without any programmer synchronization have supported yet?

(page faulting may be more performance in sparse data changing?)

I don’t know what those features are.

CUDA 8 + Pascal UM supports page faulting. Implicit in that is the idea that the pre-pascal device synchronization (cudaDeviceSynchronize()) which was necessary after a kernel launch is no longer necessary. But there aren’t any guarantees about data synchronization, that I am aware of. It’s still possible to have race conditions.

OK i see, thanks a lot. I try cudaHostAlloc in gtx1050ti(pascal) with cuda sdk8.0, and find that cudeDeviceSynchronize is necessary. May a traditional way, copy data in and out, is safety.

cudaHostAlloc isn’t unified memory

if you want to exchange data with a running kernel via pinned memory, there are a number of programming requirements

here is a worked example:

http://stackoverflow.com/questions/20345702/how-can-i-check-the-progress-of-matrix-multiplication

I wouldn’t bother trying this on windows WDDM, that makes it even harder

oooooooh. I just make a mistake, I will try cudaMallocManaged next and feedback conclusion. and I will read your refer later. thanks.

says:

feature CPU and GPU access unified memory allocations without any programmer synchronization have supported yet?

I try that, unified memory cannot supported this feature, may page faulting is also a dream.

more detail in:
http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#um-gpu-exclusive