Passing messages from one GPU to another

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

does anyone know if there is a way to pass information from one GPU to another on the same PCI bus without the need of the CPU? I’m looking for a something that can work from a cuda kernel.

Thanks.

2

If the GPUs are in a P2P capable setting, you can directly access the memory of 1 GPU from another, if you enable P2P. The transactions to support this will flow over the P2P connection (either the PCIE fabric or the NVLink fabric). This doesn’t require any explicit CPU involvement.

3

Thanks a lot for the information. How can I enable P2P for GTX cards? Can you please point me to some documentation with more information about this technic?