Direct memory transfer from GPU to GPU?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi all

I am developing a dual-GPU system, where a data set is processed first by GPU-1 and then by GPU-2.

I opened a buffer in the host to put the processed data by GPU-1, using Device-to-Host copy, and then transfer them to GPU-2 by Host-to-Device copy.

This seems not that efficient due to the PCI-E bandwidth limit. So is there some direct way to transfer memory from GPU to GPU? How about SLI? Does CUDA support SLI?

Thanks

Ben

2

GPUDirect Technology can help you if your system satisfies its requirement.
It uses a shared-pinned memory-buffer to do the copy.

3

GPUDirect wouldn’t help here, as both targets are GPUs and thus both can already share pinned system memory via the CUDA API.

As for the original question… you can’t do a direct transfer in CUDA today.

4

Oops… Thanks for correcting. GPUDirect has more relevance in a cluster envmt…where the infiniband adapter and GPU share a pinned buffer to reduce buffer copies for GPU-GPU communication.