| 27 | 0 | 15 |
| 下载次数 | 被引频次 | 阅读次数 |
针对图形处理器（GPU）Scale-Up卡间互联，系统梳理了其发展背景、核心需求与现实挑战，总结了从通用总线、私有协议到开放标准的3阶段演进历程及技术趋势。以中国移动提出的全向智感互联架构（OISA）为分析对象，详细阐述了其架构设计、核心技术特征与创新优势。在此基础上，结合大模型推理阶段键值缓存（KV Cache）的存储通信需求，以及超节点场景下Scale-Up互联面临的新挑战，对其未来发展方向进行了展望与探讨，为智算网络体系构建及超节点演进提供了参考。研究结果表明，各协议技术正逐渐趋同，开放互联已成为重要发展态势。OISA依托分层协议栈设计与关键技术创新，构建了高效、可靠、灵活、开放的GPU高速互联体系，为智算服务器向超节点迭代升级提供了关键支撑。
Abstract：A systematic review is conducted on the background, core requirements, and practical challenges of graphics processing unit(GPU) Scale-Up interconnection, and its three-stage evolution from general-purpose buses and proprietary protocols to open standards, along with corresponding technical trends, is summarized. With the omnidirectional intelligent sensing express architecture(OISA) proposed by China Mobile taken as the analytical object, its architecture design, key technical features, and innovative advantages are elaborated in detail. On this basis, combined with the storage and communication demands of Key-Value(KV) Cache in large-model inference and the emerging challenges faced by Scale-Up interconnection in SuperPod scenarios, its future development directions are discussed and prospected, providing references for intelligent computing network construction and SuperPod evolution. The results indicate that interconnection protocols and technologies are gradually converging, and open interconnection has become a key development trend. Based on hierarchical protocol stack design and key technological innovations, OISA establishes an efficient, reliable, flexible, and open high-speed GPU interconnection system, providing critical support for the iterative upgrade of intelligent computing servers toward SuperPods.
[1]段晓东,程伟强,张昊.智算网络发展综述[J].中兴通讯技术, 2025,31(2):53-62. DOI:10.12142/ZTETJ.202502008
[2]Li A, Song S L, Chen J Y, et al. Tartan:evaluating modern GPU interconnect via a multi-GPU benchmark suite[C]//Proceedings of IEEE International Symposium on Workload Characterization(IISWC). IEEE, 2018:191-202. DOI:10.1109/IISWC.2018.8573483
[3]崔晨,吴迪,陶业荣,等.多GPU系统的高速互联技术与拓扑发展现状研究[J].航空兵器, 2024, 31(1):23-31. DOI:10.12132/ISSN.1673-5048.2023.0138
[4]华为技术有限公司,中国信息通信研究院.超节点发展报告[R]. 2025
[5]Song X Y, Zhou D Y, Li K, et al. Survey of intra-node GPU interconnection in scale-up network:challenges, status, insights,and future directions[J]. Future Internet, 2025, 17(12):537. DOI:10.3390/fi17120537
[6]Das Sharma D. PCI-express:evolution of a ubiquitous load-store interconnect over two decades and the path forward for the next two decades[J]. IEEE circuits and systems magazine, 2024, 24(2):47-61. DOI:10.1109/MCAS.2024.3373556
[7]Foley D, Danskin J. Ultra-performance pascal GPU and NVLink interconnect[J]. IEEE micro, 2017, 37(2):7-17. DOI:10.1109/MM.2017.37
[8]Schieffer G, Shi R M, Markidis S, et al. Understanding data movement in AMD multi-GPU systems with infinity fabric[C]//Proceedings of SC24-W:Workshops of the International Conference for High Performance Computing, Networking,Storage and Analysis. IEEE, 2024:567-576. DOI:10.1109/scw63240.2024.00079
[9]UALink Consortium. UALink consortium releases the ultra accelerator Link 200G 1.0 specification[EB/OL].(2025-04)[2026-04-09]. https://ualinkconsortium. org/wp-content/uploads/2025/04/UALink-1.0-Specification-PR＿FINAL.pdf
[10]华为.灵衢基础规范[R]. 2025
[11]Open Compute Project. The OCP ESUN 1.0 specification has been released[EB/OL].(2026-03-10)[2026-04-09]. https://www. opencompute. org/blog/the-ocp-esun-10-specificationhas-been-released
[12]中国移动.全向智感互联OISA 2.0技术规范[R]. 2025
[13]Liu F X, Zhang Q H, Shen H J, et al. HyperOffload:graph-driven hierarchical memory management for large language models on SuperNode architectures[PP/OL]. arXiv[2026-04-09]. https://arxiv.org/abs/2602.00748
[14]Saber M G, Jiang Z P. Physical layer standardization for AI data centers:challenges, progress, and perspectives[J]. IEEE network, 2026, 40(2):147-155. DOI:10.1109/MNET.2025.3557812
[15]Liu Y H, Cai Z Z, Chen Y X, et al. InfiniteHBD:building datacenter-scale high-bandwidth domain for LLM with optical circuit switching transceivers[EB/OL].(2025-02-07)[2026-04-09]. https://arxiv.org/abs/2502.03885
基本信息:
中图分类号:TP332
引用信息:
[1]宋晓勇,李锴,陈佳媛.Scale-Up网络内GPU卡间互联及OISA技术体系[J].中兴通讯技术,2026,32(03):15-21.
2026-06-26
2026-06-26
2026-06-26