Summary

  • The IETF’s current text for draft-xu-idr-fare-in-mpson-01, dated 16 August 2026, extends FARE-BGP’s bandwidth-weighted ECMP idea to RoCE NICs in multi-plane scale-out networks.
  • Its operational problem is twofold: an RNIC must select between isolated CLOS planes, while host-route state can become too large to retain everywhere.

The IETF Datatracker record identifies the document as Fully Adaptive Routing Ethernet in Multi-Plane Scale-Out Networks. It is an Internet-Draft with intended Standards Track status. That is an important limit on the signal: the IETF describes Internet-Drafts as working documents that may change, be replaced or expire; the status is not a standards decision.

The version -01 text describes a topology in which every RNIC is attached to an independent CLOS fabric for each plane. The planes have no links between them, so the RNIC is asked to decide which plane carries a packet or a flow. The draft extends the FARE-BGP concept—weighted ECMP informed by a Path Bandwidth Extended Community—from switches to those interfaces. It calls for a separate BGP session to attached leaf switches in each plane, through which an RNIC can learn remote reachability and the path-bandwidth information used for weighted choice.

That puts a deceptively familiar routing primitive in a different control location. The policy question is no longer only how fabric switches split equal-cost traffic. It is also whether the values supplied to a host-facing interface are complete, comparable and promptly withdrawn when one plane loses reachability. The draft’s aim is not proof that an AI cluster will perform better; it is a proposed mechanism for exposing a routing choice that an isolated-plane architecture already forces the interface to make.

Scale is the other constraint. The text uses an illustrative network of 100,000 GPUs and four planes, in which each plane can propagate up to 100,000 host routes. It says the resulting 400,000 routes are impractical to store on every RNIC and undesirable to retain in full on switches. Its answer is route-table suppression. But aggregation can leave a stale aggregate pointing into a plane after a remote RNIC has become unreachable. To avoid that blackhole, the proposal pairs aggregation with an explicit unreachable-host advertisement; one stated option is a Path Bandwidth value of zero.

This is design-stage work, not evidence that the scheme has been deployed or interoperates across products.