Alibaba Cloud and Tsinghua Intelligent Computing Center's High-Performance Network HPN Wins CCF Science and Technology Progress First Prize

On October 26, the China Computer Federation (CCF) officially announced the results of its 2026 Science and Technology Achievement Awards. The project "Key Technologies and Applications of High-Performance Networking for Large-Scale Intelligent Computing Centers," jointly completed by Alibaba Cloud and Tsinghua University, won the First Prize for Scientific and Technological Progress in the CCF Science and Technology Achievement Awards.

The CCF evaluation committee, composed of multiple experts, believes: "Alibaba Cloud's innovative project addresses the challenges that artificial intelligence poses to data center networks and has achieved creative results in Ethernet-based intelligent computing center networks, reaching internationally leading levels in areas such as the large-scale intelligent computing two-layer flat network architecture, end-network collaborative proactive-awareness path selection, and root cause localization based on multi-dimensional trace correlation analysis."

The development of large models requires not only stronger single-card computing power but also the efficient coordination of tens of thousands of GPUs. Therefore, the Alibaba Cloud networking team developed the high-performance network for intelligent computing centers (HPN), aiming to make data transmission faster, network utilization fuller, and fault localization more timely, thereby unleashing the computing potential of large-scale intelligent computing centers. Specifically, Alibaba Cloud HPN has achieved major technological innovation and breakthroughs in the following three aspects:

In terms of network architecture, HPN designed a "two-layer flat, high-density multi-rail" network architecture, transforming the traditional "three-layer stacking, single-channel cascading" into "two-layer flat, multi-rail parallel." By combining high-speed interconnection within hosts with multi-rail networking between hosts, it reduces unnecessary cross-layer forwarding and bandwidth contention, making data exchange in ten-thousand-card clusters more efficient. Third-party testing shows that the effective bandwidth of ten-thousand-card clusters is nearly 30% higher than that of the traditional three-layer Clos architecture.

In terms of path selection, HPN enables servers and network devices to exchange task requirements and link status information in real time, shifting from traditional fixed-mapping path selection to proactive awareness and dynamic adjustment. It is like equipping data transmission with a "navigation system" that can sense road conditions, guiding traffic to appropriate paths and reducing situations where some links are congested while others sit idle. In related tests, bandwidth utilization reached 97%, about 37% higher than Microsoft's Weighted Bandwidth Selection (WBS) technology.

Alibaba Cloud HPN also has the ability to correlate and analyze communication information between GPUs with task operating status, identifying both faults and the causes of performance slowdowns, and shortening abnormal root cause localization to the second level. Compared with the SuperBench offline root cause localization system in comparative tests, the GPU idle time caused by fault diagnosis was reduced by an average of 97%, helping intelligent computing clusters reduce troubleshooting waits and resume efficient operation faster.

From technological innovation to industrial application, HPN has supported dozens of ten-thousand-card intelligent computing clusters and serves as the underlying network support for Alibaba Cloud's Lingjun intelligent computing clusters. It carries large-model training and inference workloads for Qwen, Xiaohongshu, XPeng, and others, and also provides support for the ModelScope open-source community, enabling high-performance intelligent computing capabilities to serve a broader range of enterprises and developers.

The project has also produced solid independent intellectual property rights and academic achievements, obtaining 33 authorized invention patents in China and abroad and publishing 45 papers in the field of computer network systems, including 14 SIGCOMM papers and 11 NSDI papers. The representative achievement HPN7.0 is the first intelligent computing network architecture paper in the history of ACM SIGCOMM, pioneering a network architecture solution for ten-thousand-card intelligent computing clusters based on open Ethernet. Crux received the ACM SIGCOMM 2024 Best Paper Nomination Award, reflecting the project's innovative value in communication scheduling and computing utilization efficiency.

"This award recognizes the technological innovation and industrial application results of networking in the AI era," said Cai Dezhong, head of Alibaba Cloud's Basic Network Business Unit. From network architecture to traffic scheduling and then to operational assurance, Alibaba Cloud and Tsinghua University, through industry-academia-research collaboration, have transformed network capabilities into foundational support for large-model development, providing a practical path for building high-performance, high-stability intelligent computing infrastructure.