ROCEv2 vs OSI Layer

Key Points:

• Traditional OSI Stack (right):

  • Applications communicate through the full OSI stack (Application → Presentation → Session → Transport → Network → Data Link → Physical).

  • The CPU handles most of the processing overhead.

• ROCEv2-Optimised Stack (left):

  • Uses an RDMA NIC (RNIC) that implements parts of the transport (UDP), network, and data link layers in hardware.

  • Allows kernel bypass, letting applications (like GPUs) directly access the network stack without involving the CPU.

  • Data flows directly between GPU memory and the NIC, reducing latency and CPU usage.

• Comparison:

  • Traditional stack = CPU-intensive with full OSI traversal.

  • ROCEv2 stack = hardware offload + kernel bypass, optimized for low-latency GPU-to-GPU or GPU-to-server communication.

⚡ In short: The image shows how RoCEv2 removes CPU involvement by offloading networking functions to the NIC and enabling direct GPU-to-GPU data transfers, bypassing much of the traditional OSI stack.