Requirements and Gap Analysis of Multicast in AI Data Centers#

Junye Zhang, Weiqiang Cheng, Kefei Liu (China Mobile)

• Why: MoE token dispatch (AlltoAll), AllReduce broadcast phases, model distribution, and checkpoint saving are all point-to-multipoint; unicast replication burns source GPU/NIC capacity and bandwidth. If network multicast could serve AIDCs, it would offload replication to switches — but no existing multicast tech fits.
• What: A requirements-and-gap document. Requirements: interactivity (backward MP2P forwarding and ACK/congestion-feedback aggregation), reliability (lossless — RDMA throughput collapses to 0 at ~2% loss), dynamics (µs-scale member-set adaptation for MoE gating, no pre-defined groups), sparseness (e.g., 9 of 256 experts active in DeepSeekV3 — no BIER BitString bloat), and simplicity.
• How: Gap analysis: PIM/MLDP convergence too slow and too complex; SR-P2MP needs global recalculation; BIER relatively good on dynamics/simplicity but BitString overhead hurts sparse membership; none handle MP2P feedback or losslessness. Conclusion: a new architecture or deep enhancement of existing multicast is needed for AIDCs.
• Where applicable: AIDC scale-out fabrics. Individual gap-analysis draft (-00) seeking feedback — this is a problem statement that could charter new work rather than a solution; useful for the PM to anticipate where multicast requirements are heading.