SRv6 for Deterministic Path Placement in AI Backends#

Ahmed Abdelsalam (Cisco), on behalf of co-authors from Cisco, Microsoft, Oracle Cloud, IBM Cloud, NVIDIA and Arrcus.

• Why: This is the hyperscaler reference design for AI backends — the draft the SRv6/srv6ops community is rallying around, and it is requesting WG adoption at this meeting.
• What: Describes how SRv6 NEXT-CSID enables deterministic path placement: the NIC's transport stack encodes each path as an ordered segment list in the packet header, and the fabric forwards statelessly. It documents the MRC (Multipath Reliable Connection) design: packet spraying, ECN-based adaptive load balancing, out-of-order memory placement, selective retransmission, and packet trimming for incast — reaching 131,000 GPUs with only two tiers of switches (8×100Gb/s per host instead of 1×800Gb/s).
• How: Switches carry no per-flow state, no resilient-ECMP hashing, no BGP convergence logic; the end host steers deterministically and can recover from failures within one RTT. Loopback SRv6 probes allow path probing without a peer.
• Where applicable: Hyperscale production deployments (the deck cites OpenAI, Microsoft, and Oracle MRC+SRv6 deployments; rev04 adds Microsoft and Oracle as co-authors). Not directly transplantable to smaller estates, but it defines the pattern NICs and switches will be asked to support.