Co-Fabric: Breaking Host-Domain Boundaries for Unified xPU Interconnection

📅 2026-09-21
📈 Citations: 0
✨ Influential: 0
📄 PDF
🤖 AI Summary
为解决大模型参数超出单个xPU容量导致的跨主机域通信效率问题,提出Co-Fabric统一互连方案,通过四层协议栈、跨域扩展机制和统一地址空间,显著降低延迟并提高带宽。
📝 Abstract
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.
Problem

Research questions and friction points this paper is trying to address.

xPU interconnection
host-domain boundaries
large-model parameters
system efficiency
unified addressing
Innovation

Methods, ideas, or system contributions that make the work stand out.

unified xPU interconnection
four-layer protocol stack
cross-domain scaling
P2P mechanism
unified address space
🔎 Similar Papers
No similar papers found.
Z
Zhen Peng
IEIT SYSTEMS Co., Ltd., Beijing, China
J
Jiaming Huang
IEIT SYSTEMS Co., Ltd., Beijing, China
Chaofan Chen
Chaofan Chen
Assistant Professor, University of Maine
Machine LearningArtificial Intelligence
Z
Zhao Zhang
IEIT SYSTEMS Co., Ltd., Beijing, China
A
An Wu
IEIT SYSTEMS Co., Ltd., Beijing, China
B
Baoyang Liu
IEIT SYSTEMS Co., Ltd., JiNan, China
Xinglong Wang
Xinglong Wang
IEIT SYSTEMS Co., Ltd., JiNan, China
T
Tanlong Ci
IEIT SYSTEMS Co., Ltd., JiNan, China
J
Jinfeng Li
IEIT SYSTEMS Co., Ltd., Beijing, China
X
Xueke Duan
IEIT SYSTEMS Co., Ltd., Beijing, China
H
Hao Wang
IEIT SYSTEMS Co., Ltd., JiNan, China
X
Xi Chen
IEIT SYSTEMS Co., Ltd., JiNan, China
S
Shunshun Zhang
IEIT SYSTEMS Co., Ltd., JiNan, China
Z
Zhiyuan Su
IEIT SYSTEMS Co., Ltd., JiNan, China
Zhu Cao
Zhu Cao
Tsinghua University
Z
Zhichong Dou
IEIT SYSTEMS Co., Ltd., JiNan, China
S
Shaohua Wu
IEIT SYSTEMS Co., Ltd., Beijing, China
L
Lu Jing
IEIT SYSTEMS Co., Ltd., Beijing, China
Y
Yue Yuan
IEIT SYSTEMS Co., Ltd., Beijing, China