Why it matters

GPU P2P avoids CPU copies. High-bandwidth intra-node pattern.

Advertisement

The architecture

Enable P2P access.

Direct memory read/write.

Over NVLink or PCIe.

GPU P2P memory accessEnable P2PcudaDeviceEnablePeerAccessDirect RWeach other's memoryOver NVLinkor PCIecudaMemcpyPeer + P2P enable = high bandwidth intra-node
GPU P2P memory access.
Advertisement

How it works end to end

cudaDeviceEnablePeerAccess.

cudaMemcpyPeer.

Unified addressing.