GSE architecture for AIDC#

Rui Zhuang (China Mobile), Sandy Zhang (ZTE).

• Why: A fundamentally different answer to AI-cluster congestion: instead of reactive ECN/PFC, schedule everything before sending. China Mobile/ZTE are fielding this, so it will surface in AIDC requirements even if it doesn't win in IETF.
• What: Global Scheduling Ethernet: credit-based authorization (a sender must get a bandwidth grant from the receiver-side last hop before transmitting), packet aggregation into uniform-size segments for forwarding, improved ECMP that sprays packets while preserving receiver-side ordering, and interworking with PFC/ECN. The draft presents intra-PoD (GPU-to-GPU credit negotiation) and inter-PoD (Leaf-to-Super-Spine) scenarios, with port IDs advertised via BGP (draft-zhang-idr-portid-ec).
• How: New Ethernet packet types carry a GSE header for uniform-size segments and for credit request/grant messages; switches recognize and forward GSE-framed traffic. PoDs from different vendors can interoperate by restoring original packets at the Core.
• Where applicable: AI data centers. Implemented in prototype hardware/software with "early performance data [showing] significant gains over traditional RoCE" — no numbers published yet. China Mobile-driven; treat as one contender for the post-RoCE AIDC architecture rather than a settled standard.