广域无损智算网络关键技术研究与实践

摘要：智算中心跨域互联已成为支撑规模化人工智能服务的关键网络新形态。系统分析了广域智算网络面临的传输需求与挑战，提出了构建广域无损智算网络的核心观点，以精准流量控制和队列缓存增强为关键技术路径，推动广域智算互联从尽力而为传输向无损承载演进。广域无损智算网络面向存算分离训练、云边协同推理等多类智算服务场景，实现了网络拥塞条件下的无损高效传输，为规模化智算服务构建了稳定可靠的网络底座。

关键词：智算网络；广域网；拥塞丢包；无损传输

Abstract: Cross-domain interconnection of intelligent computing centers has emerged as a critical network paradigm for large-scale artificial intelligence services. The transmission requirements and challenges of wide-area intelligent computing networks are analyzed, and the core concept of building a wide-area lossless intelligent computing network is proposed. Leveraging precise flow control and enhanced queue buffering as key technical enablers, this network drives the evolution of wide-area intelligent computing interconnection from best‑effort to lossless transport. Targeting diverse service scenarios such as storage‑compute decoupled training and cloud‑edge collaborative inference, the wide-area lossless intelligent computing network achieves lossless and efficient transmission under network congestion, thereby establishing a stable and reliable foundation for large-scale intelligent computing services.

Keywords: intelligent computing network; wide area network; congestion-induced packet drop; lossless transmission