RoCEv2 (RDMA over Converged Ethernet)
Zero-copy network protocol bypassing CPU for direct GPU-to-GPU memory access
Hardware
In plain English
A VIP direct pipeline between graphics cards in different computers so they can share memory across a network cable without asking the CPU for permission.
Technically
RDMA over Converged Ethernet v2. An internet layer protocol using UDP/IP that allows GPUs on separate networked nodes to write directly to each other's memory buffers via ConnectX NICs, achieving sub-2µs latencies required for distributed NCCL tensor parallelism.