GPUDirect API?

As the press release came out last Thursday from NVIDIA and Mellanox, we finally have so-called “GPUDirect” technology to solve the incompatibility issue between infiniband memory region and CUDA pinned memory. So I was wondering if this would introduce a new set of API to make a special kind of pinned memory or just use existing cudaMallocHost() API? (please surprise me that cuda 3.1b has already shipped with that feature…)

I’d also like to now when this technology will be available to CUDA developers on Windows, and whether it requires a Fermi GPU or other GPUs will be compatible. I can find no mention of GPUDirect on NVIDIA’s site yet.

I doubt this functionality will be included in the public Cuda release. There was talk here on the forums over a year ago about the addition of APIs for supplying your owned pinned memory to Cuda, but nothing since. If you are a Linux user, you can modify the Nvidia driver, since they include the Linux-specific code for recompiling against your kernel. Tough luck in Windows, those drivers are totally closed-source.