The Key to AI Network, H3C’s New Generation Lossless Network Solution Based on DDC Architecture

H3C's new generation lossless network solution based on the DDC (Diversified Dynamic-Connectivity) architecture focuses on the core challenges faced by data centers in the AI era, such as traffic congestion and complex scheduling. Through architectural innovation, it achieves efficient collaboration and lossless transmission across the entire network, helping customers reshape the infrastructure of AIDC and fully unleash the value of computing power.

Diversified Computing and End-Network Synergy

Will Be the Key to the Next-Generation Intelligent Computing Network

As the parameters of large models have leaped to the trillion level, the evolution of AI is driving a comprehensive upgrade of infrastructure. Although the investment in networks accounts for a relatively small proportion of the total cost, its impact on training efficiency and computing power returns has become a crucial variable in the upgrade of AIDC.

With the rapid iteration of large models, intelligent computing networks are facing new challenges: on the one hand, the open-source trend is accelerating the expansion of the ecosystem, and computing power are becoming more diverse; on the other hand, traffic models are becoming increasingly complex, putting forward higher requirements for network bandwidth and scheduling capabilities.

In the future, intelligent computing networks must simultaneously possess the ability to be compatible with diversified computing and the agile scheduling ability for complex traffic, and continuously evolve towards a more open, efficient, and scalable direction.

DDC Is a Better Solution to AI Network Congestion and Scheduling

H3C has launched its new generation lossless network solution based on the DDC architecture. It adopts an advanced cell switching mechanism to optimize traffic balance, address the congestion and scheduling challenges of AI networks in high-concurrency scenarios, and simultaneously enhance the reliability and manageability of large-scale clusters.

This solution realizes cross-switch traffic scheduling through Scheduled Fabric, improves training efficiency, and is compatible with 400G/800G access, breaking through the limitations of traditional port capacities and facilitating the networking of clusters with over 70,000 cards.

According to Tolly tests, under the All-to-All traffic model, the RoCE network of H3C based on the DDC architecture has a throughput performance comparable to that of the InfiniBand (IB) network. Moreover, the larger the message size, the better the performance.

In addition, adhering to the concept of open co-construction, H3C collaborates with partners to formulate the core framework standards of DDC based on OSF (Open Scheduling Framework for AI Networks), focusing on solving the three core problems of load balancing, congestion control, and network reliability in intelligent computing network scenarios, and promoting flexible heterogeneity and interconnection at the network element level.

In the upsurge driven by AI, computing power and networks are accelerating their integration and driving each other, with collaboration becoming the core engine for the development of intelligent computing.

H3C will remain committed to being guided by market demands, practice the concept of "computing × connectivity", continuously tackle key technologies for the integration of computing and networking.

By joining hands with industry partners, H3C aims to create an open, diversified, and sustainable intelligent computing ecosystem and build a solid foundation with high-quality network for the intelligent upgrading of all industries.