| Internet-Draft | Abbreviated Title | July 2026 |
| Zhang, et al. | Expires 5 January 2027 | [Page] |
Multicast has the potential to be applied in Artificial Intelligence Data Centers (AIDCs) to improve the efficiency of point-to-multipoint data transmission during large language model training and inference. This document identifies key requirements of multicast in AIDCs, and analyzes the gaps between these requirements and the capabilities of existing multicast technologies.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 5 January 2027.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Artificial Intelligence (AI) Data Centers (AIDCs) serve as the key infrastructure for AI large language model (LLM) training and inference, where point-to-multipoint (P2MP) communication patterns are common and critical to overall system efficiency. Network multicast leverages in-network data replication to achieve efficient distribution of identical data, reducing processing overhead and network bandwidth consumption of the sender, thereby enhancing the efficiency of P2MP data transmission. Multicast is a promising technology for deployment in AIDCs.¶
Despite the potential opportunities, existing multicast technologies are not originally designed to address the specific characteristics of AIDC networks. AIDC networks are defined by ultra-high bandwidth (often 400 Gbps or greater), microsecond-level latency, and high reliability that demands near-zero packet loss. These core performance characteristics necessitate corresponding qualities in multicast technologies, including interactivity, reliability, and simplicity. Furthermore, emerging multicast use cases in AIDCs, such as token dispatch, also introduce specific requirements, including high dynamics and membership sparseness.¶
This document identifies the typical multicast use cases and key requirements for multicast in AIDCs, and analyzes the limitations of existing multicast technologies in meeting these requirements.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
Tasks in AIDC include model loading and distribution, model training and inference, model saving, and other key operations. These tasks generate various traffic patterns, including communication between computing devices (e.g., GPUs), traffic among storage nodes, and data transmission between computing devices and storage nodes. Among these, many typical communication patterns exhibit P2MP characteristics, making multicast a critical enabling technology. The typical multicast use cases in AIDCs are as follows:¶
AI workloads are highly sensitive to packet loss. In LLM training, packet loss without a reliability acknowledgment mechanism can corrupt model parameters, leading to degraded model quality or even training failures. Moreover, congestion control is required to actively avoid congestion and packet loss. Therefore, networks in AIDCs are required to support closed-loop control, such acknowledgment and congestion control, to meet the high-performance and high-reliability requirements of AI workloads.¶
Traditional IP multicast only supports best-effort P2MP data delivery, while multicast in AIDCs should support bidirectional interaction, including both efficient P2MP data forwarding and multipoint-to-point (MP2P) feedback forwarding, as shown in Figure 1.¶
XPU1 Switch XPU2 XPU3 | | | | | <P2MP> | | | Data | Data | | |--------> |--------->| | | | Data | | |----------------->| | | | | | | | | | <MP2P> | | | ACK | ACK | | |<---------|<---------| | | | | ACK | | |<-----------------|
The core interactivity demands are as follows:¶
AI workloads exhibit near-zero tolerance for packet loss. Even with acknowledgments and retransmissions mechanisms with the support of bidirectional interactivity, extremely low packet loss rates can trigger massive retransmissions. In multicast scenarios, loss at any receiver can force the sender to retransmit data to all receivers, causing significant redundant traffic and efficiency degradation. Furthermore, maintaining uninterrupted tasks for long periods is crucial for LLM training. However, hardware is prone to failures, and as the scale of training networks increases, the likelihood of network failures rises due to an increasing number of switches, network interface cards, and optical modules [I-D.cheng-rtgwg-ai-network-reliability-problem]. Therefore, multicast in AIDCs should provide high reliability to ensure service performance and continuity. The specific requirements are as follows:¶
AI workloads, especially those using sparse architectures like MoE, have highly dynamic communication patterns. MoE-based AI training and inference uses token dispatch, where gating networks select expert nodes per token at microsecond timescales, dynamically determining real-time multicast receiver sets with no fixed groups. This ultra-fast selection leaves no time for traditional multicast to establish, update, or tear down trees, leading to delays, packet loss, or AI task failure [I-D.zhang-rtgwg-llmmoe-multicast]. Therefore, multicast in AIDC should meet high dynamics requirements, and the key points are as follows:¶
Multicast in AIDCs frequently involves multicast groups where only a small fraction of the total nodes in the cluster are multicast members, a characteristic closely tied to the sparse activation mechanism of modern AI models such as MoE. For example, DeepSeekV3 uses 256 experts and activates 9 experts at a time. Multicast technologies that are designed for dense groups are inefficient for this sparse mode. The multicast should be efficient when the group size is small relative to the network size, and meet the following sparseness requirements:¶
Simplicity is a foundational architectural principle for multicast in AIDCs, directly enabling the microsecond-timescale low-latency transmission in large-scale AIDC networks. Complexity in the control or data plane manifests as variable latency, unpredictable jitter, and an inability to meet the strict performance bounds of AI workloads. Therefore, multicast in AIDCs should be governed by the following overarching simplicity requirements:¶
To address the gaps between multicast requirements in AIDCs and existing technologies, typical multicast technologies are first introduced, followed by an analysis of their capabilities against key requirements.¶
Protocol Independent Multicast (PIM) is a widely deployed multicast routing protocol that operates independently of underlying unicast routing protocols. It supports dense mode (PIM-DM) [RFC3973] and sparse mode (PIM-SM) [RFC7761]. PIM-SM builds unidirectional shared trees rooted at a Rendezvous Point per group and it optionally creates shortest-path trees per source.¶
Multipoint extensions for Label Distribution Protocol (mLDP) [RFC6388] constructs the P2MP or multipoint-to-multipoint (MP2MP) Label Switched Paths (LSPs) in Multiprotocol Label Switching (MPLS) networks without interacting with or relying upon any other multicast tree construction protocol.¶
Segment Routing Point-to-Multipoint (SR-P2MP) [RFC9960] enables the creation of P2MP trees for efficient multi-point packet delivery in a Segment Routing (SR) domain. It requires the routing module of the controller or ingress node to calculate and determine the path of the multicast traffic, and the data plane can reuse existing SR unicast forwarding mechanisms.¶
Bit Indexed Explicit Replication (BIER) [RFC8279] is a stateless multicast technology that eliminates the need for explicit tree construction. Instead, the set of intended receivers is encoded as a BitString within the packet header. Intermediate BIER Forwarding Routers (BFRs) replicate packets based on the BitString, without maintaining any per-flow or per-tree state.¶
The support of typical multicast technologies for multicast requirements in AIDCs is summarized in Table 1.¶
| Technology | Interactivity | Reliability | Dynamics | Sparseness | Simplicity |
|---|---|---|---|---|---|
| PIM | Poor | Poor | Poor | Good | Poor |
| mLDP | Poor | Poor | Poor | Good | Poor |
| SR-P2MP | Poor | Moderate | Moderate | Good | Moderate |
| BIER | Poor | Moderate | Good | Moderate | Good |
Interactivity: These multicast technologies can support best-effort P2MP data delivery, but none of them can natively support the reverse MP2P forwarding or even aggregation to achieve bidirectional interactivity.¶
Reliability: These multicast technologies fail to meet the lossless requirement of AIDC networks. The reliability of PIM and mLDP basically relies on routing convergence and multicast tree reconstruction. Although some fast detection and recovery mechanisms [RFC9186][RFC9860][RFC7715] can be adopted to accelerate failure recovery, their tree-based architectures often keep the failure impact domain tree-level. In contrast, BIER and SR-P2MP can effectively reuse unicast’s reliability capabilities such as Fast ReRouting, and control the failure domain within the damaged receivers, demonstrating good reliability.¶
Dynamics: PIM and mLDP adjust multicast trees via control signals, leading to slow convergence that struggles to handle high-frequency member changes. SR-P2MP dynamically recalculates forwarding trees via a controller, which needs global recalculating and result distribution. BIER only requires updating the BitString in packets, enabling faster responses to member changes and exhibiting good dynamics.¶
Sparseness: PIM, mLDP, and SR-P2MP natively support sparse multicast scenarios. These protocols maintain mappings between multicast group identifiers and multicast members. They identify individual members using discrete IP addresses or dedicated identifiers, and data packets carry only group identifiers instead of full member information. By contrast, standard BIER does not pre-establish bindings between multicast groups and member sets. It encodes all nodes in a BIER sub-domain into a fixed-length BitString, where each bit corresponds to one BFR. Even with only a small number of sparse members, the full BitString sized for the entire sub-domain must be transmitted. This causes excessive bandwidth overhead and degraded forwarding performance. To improve BIER’s adaptability for sparse multicast member scenarios, optimization solutions have been proposed, such as Unmasked BIER [I-D.zzhang-bier-unmasked-bier].¶
Simplicity: PIM and mLDP require the maintenance of complex multicast tree states and signaling mechanisms, resulting in high operational complexity and poor simplicity. SR-P2MP reuses the SR unicast forwarding plane, with the control plane relying on a controller, leading to moderate complexity but still requiring additional tree management logic. BIER, on the other hand, eliminates the need for explicit multicast tree construction, with no per-flow state at intermediate nodes, resulting in better simplicity. Moreover, simplicity still needs further optimization to meet the ultra-high performance requirements of AI networks.¶
In summary, the most critical common gap is the lack of native support for efficient, scalable bidirectional interactivity, which is the cornerstone for implementing closed-loop acknowledgement and congestion control. Furthermore, no single multicast technology excels in all dimensions: some lack reliability, dynamics or simplicity (PIM, mLDP, SR-P2MP), others are inefficient for sparse groups (BIER). Consequently, merely deploying or combining these existing technologies is insufficient to meet the stringent demands of AIDC workloads. This gap analysis underscores the need for either a new architecture designed from the ground up for AIDCs or significant extensions to existing technologies.¶
TBD.¶