What are the limitations of using cudaMemcpyPeerAsync? Such as GPU model, motherboard type and so on? Where can I find the restrictions? Is there an indication on the parameters of the graphics card that this feature is available? Or more directly, where can I find a list of hardware devices that support this feature?
There isn’t any collated information provided by NVIDIA to make this determination.
The runtime method to determine this is cudaDeviceCanAccessPeer().
For casual usage, you should use that method. For acquisition purposes, either confirm with an actual configuration tested, or make it a condition of purchase.
It’s possible to make various generalizations, which you can find with a bit of searching, but nearly all the generalizations I’m aware of change over time, or stated another way have exceptions.