Acesso ponto a ponto à memória

Moving data from one GPU to another usually bounces through the CPU's memory first. That round trip is slow and wastes the host's bandwidth.

Modern GPUs can skip the CPU entirely. Peer-to-peer access lets one GPU read and write another's memory over a direct link.

The fast link between cards is often NVLink, far quicker than the shared PCIe bus. P2P transfers ride this highway when it is available.

Not every pair of GPUs can do P2P. Ask first with cudaDeviceCanAccessPeer, which reports whether one device may reach another.

Permission is off by default. While GPU 0 is current, call cudaDeviceEnablePeerAccess to let it reach into GPU 1's memory.

Enabling peer access grants only the direction you ask for. For GPU 1 to read GPU 0, you must enable that direction too, from device 1.

To move a buffer directly between cards, use cudaMemcpyPeer. You name the destination pointer and device plus the source pointer and device.

With access enabled, a kernel on GPU 0 can dereference a pointer that lives on GPU 1. The hardware fetches it over the peer link transparently.

If two cards cannot peer, the runtime quietly falls back to staging through host memory. You still get a copy, just at the slower PCIe rate.

Peer transfers can overlap other work. Pair cudaMemcpyPeerAsync with a stream so the copy runs while kernels keep computing.

Peer access uses resources, so release it when finished. cudaDeviceDisablePeerAccess tears down the link you opened earlier.

Recall the benefit of peer-to-peer access between two GPUs.

P2P lets GPUs share memory directly, ideally over NVLink. Check support, enable access per direction, then use cudaMemcpyPeer. Next: scaling with NCCL. ✨