Is GPUDirect RDMA supported under Windows?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I would like to transfer data directly from a frame grabber to a GPU and therefore use RDMA. It seems to be possible to use the nVidia Kernel API (nv-p2p.h) to get the physical addresses from GPU virtual addresses using function “nvidia_p2p_get_pages()”.
This seems to be made possible by nVidia via the GPUDirect API ([url]GPUDirect RDMA :: CUDA Toolkit Documentation).
This section [url]GPUDirect RDMA :: CUDA Toolkit Documentation shows how to check if a given Virtual Address belongs to GPU memory or not and how to convert it to a list of physical addresses if it is GPU memory.

The question is: “Is this available under Windows?”

2

I too am interested in transferring data from a frame grabber to GPU using RDMA.

3

Hello,
I am also trying to transfer image data directly from a frame grabber to a Tesla (K20) using RDMA. But is seems that there is not enough documentation on this. Has anybody has successfully transfer data to the GPU through RDMA??

4

What information would you need that is not covered by this document:

[url]GPUDirect RDMA :: CUDA Toolkit Documentation

5

Although the document is a great springboard, there is some additional information that would make my life much easier.

The document specifies developing for a Linux system. In other documentation, it appears that Red Hat is the preferred (or most common) distribution. Is there support for RDMA in Windows 7? Windows 8?

Are all Kepler-class GPUs supported? What about those in laptops?

Thanks in advance for any help. I would hate to spend time and money on paths that have no chance of success.

6

Thanks for the feedback. I will forward this internally so clarifications to the documentation can be made going forward.

7

Thanks.

8

But what is the answer? Is the GPUDirect API available for the Windows platform?

9

I too am interested in support for GPUDirect RDMA for Microsoft Windows. Can someone please give me a confirmation that it is either only supported for Linux or supported across multiple platforms?

The developer guide i found seems to be focused on Linux in particular.
(http://docs.nvidia.com/cuda/gpudirect-rdma/index.html)

Thanks in advance.

10

Is there any support for this one Windows on recent GPUs ?
I want to move away from AMD GPU to NVIDIA but must have some solution for DMA transfer, without CUDA, on Windows O/S.