Co-Fabric Paper Proposes Cross-Host Bus Interconnect for xPU Scale-Up
Co-Fabric: Breaking Host-Domain Boundaries for Unified xPU Interconnection
The result suggests scale-up fabrics can bypass Ethernet semantics entirely rather than optimizing RoCE further, shifting design pressure from network protocols to bus-like architectures. Infrastructure engineers evaluating cluster interconnects for large-model inference, or hardware teams building AI data centres, can use the 64-xPU benchmarks as a concrete baseline against RoCE-based deployments.
A paper on arXiv introduces Co-Fabric, a bus-based interconnect that removes host-domain boundaries to unify xPU connections across scale-up superpods.
It uses a four-layer protocol stack with nanosecond-scale processing latency, port-identifier-based P2P routing, and a unified address space via shadow-device auto-enumeration.
On a 64-xPU 3D-Mesh testbed, it cuts inter-node latency by over 50% and improves bandwidth 2-5x versus RoCE, accelerating DeepSeek R1 inference by 30%-80%.
The design also reduces interconnect cost by up to 80% and power by 5% compared with Ethernet-based RoCE.
Co-Fabric: Breaking Host-Domain Boundaries for Unified xPU Interconnection
Large-model parameters have grown beyond the capacity of a single xPU, dispersing across multiple xPUs spanning distinct host domains, where xPU-to-xPU communication dominates overall system efficiency. Existing scale-up interconnect remains inadequate: network-based solutions built on Ethernet--such as RoCE (RDMA over Converged Ethernet)--introduce specific message-semantics and protocol-stack characteristics, and rely on fragmented per-host addressing, while conventional host-based fabrics are confined to a single host domain and lack cross-host unified addressing. This paper presents Co-Fabric, a bus-based interconnect that, unlike conventional bus designs, breaks host-domain boundaries to deliver unified xPU interconnection for scale-up superpods. Co-Fabric makes three contributions: a streamlined four-layer protocol stack achieving nanosecond-scale processing latency with native reliability; a cross-domain scaling and P2P mechanism that routes using port identifiers embedded in the packet header; and a unified address space built on shadow-device auto-enumeration. On a 64-xPU 3D-Mesh system, Co-Fabric cuts inter-node communication latency by over 50% and improves bandwidth by 2-5x over RoCE, accelerating DeepSeek R1 inference by 30%-80%. Moreover, since its streamlined four-layer protocol stack and higher data-communication efficiency reduce protocol and processing overhead relative to the Ethernet-based RoCE stack, Co-Fabric cuts the cost and power of the interconnect itself by up to 80% and 5%, respectively. These results demonstrate Co-Fabric's advantage for AI computing centers.
This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Co-Fabric, RoCE, DeepSeek, inference optimization.
Primary source verification: published directly by arXiv: DeepSeek with first-party authority.