CUDA 8 Features Revealed

Hi Mark, how do I use the new patch for cuda 8. I have a gtx 1080 and ubuntu 16.04 so i have gcc 5.4

I have tried to use the run file, but it complains about the compile as it is not backwards compatible. But the patch claims to solve that problem, but I dont know how to use it :)

Clang support in 8?

Can you clarify your question?

Cuda 7.5 introduced Clang 3.5/3.6 support for the host compiler. Does Cuda 8 update clang compatibility to a newer version?

It's almost end of August already and CUDA 8 hasn't been officially released yet. Any update on when it will be?

Any Cuda 8 release updates?

It's September now and where's CUDA 8 Toolkit release?

It's almost end of August already and CUDA 8 hasn't been officially released yet. Any update on when it will be?

Any Cuda 8 release updates?

It's September now and where's CUDA 8 Toolkit release?

It is important, so let me say it three times.

What about openCL 2.x? I'd love to use that instead of CUDA.

Noticed some comments about GCC 5. How about GCC 6? Does CUDA 8 work with it?

What OS supports Pascal unified memory? I am on Pascal Titan X and Ubuntu 16.04 but Pascal unified memory doesn't work. I still need to use cudaMallocManaged. Another weird issue I am having is my program runs a lot slower with the Pascal Titan X compare to my old Titan X. What would be possible reasons to this?

It's here. It took a lot to get this huge release out, so thank you all for waiting. https://developer.nvidia.co...

Linux. Yes, you need to use cudaMallocManaged to allocate Unified Memory for now, but you should get the page faulting behavior and you will have access to cudaMemAdvise and cudaMemPrefetchAsync, etc. We are working with the Linux community to allow Pascal to page fault on regularly allocated system memory, but that is not part of the CUDA Toolkit (read about HMM, e.g. https://www.phoronix.com/sc...

GCC 6 support will be included in the next release of the CUDA Toolkit.

Does the sentence "PASCAL UNIFIED MEMORY" means this feature not suit for previous GPU architecture, even with CUDA8?

Pascal is the first GPU architecture with the page faulting capability. Unified Memory works on Kepler and Maxwell GPUs but it has limitations because the GPU is not able to page fault.

Thanks for the explanation. Could you look into http://stackoverflow.com/qu..., which describes the problem of Pascal Titan X running slower than Maxwell Titan X. Is the behavior expected because of the new unified memory framework?

Currently NVIDIA supports OpenCL 1.2.

Will games be using unified memory soon?

Is CUDA 8 going to support Pascal-based cards inside a Mac Pro tower running Mac OS X 10.11?