In-Depth Analysis of H3C's Next-Generation Lossless Network Solution Based on DDC Architecture

H3C's next-generation lossless network solution based on the DDC (Diversified dynamic-Connectivity) architecture and its computing cluster switch, the H3C S12500AI, directly addresses the demands of AI data centers for large-scale, 10,000-cards-level interconnect scenarios.

Testing Data shows that the networking solution based on the DDC architecture effectively improves bandwidth utilization compared to traditional solutions. Its bandwidth utilization is comparable to that of InfiniBand networks, and it can flexibly adapt to diverse cluster deployment requirements ranging from 1K to 70K nodes.

As large model parameters scale to trillions, the global evolution of AI technology is extending from algorithmic innovation to comprehensive infrastructure upgrades. Data indicates that network performance fluctuations can reduce training efficiency by 30% and increase training duration by 25%. Despite its relatively small investment share, the network's impact on overall computing efficiency and economic returns is critical, making it a key variable in driving intelligent computing center upgrades.

High-Performance Network Elements Support Ultra-Large-Scale Networking

Flexible Adaptation to Thousand-to-Ten-Thousand-Card Cluster Scales

The next-generation computing cluster switch, H3C S12500AI, targets ultra-large-scale intelligent computing scenarios, fully supporting the DDC architecture with Excellent scalability and adaptability.

Its NCF utilizes a box design and supports up to 128 800G OSFP cell switching ports. The NCP downlink ports are compatible with 36 400G Q112 Ethernet ports and 18 800G OSFP Ethernet ports, flexibly adapting to mainstream network interface card (NIC) form factors and ensuring an efficient and stable access experience.

Based on the multi-cluster networking model of NCF + NCP, the DDC architecture enables large-scale cluster interconnection of over 70,000 cards. Furthermore, the open networking approach of independent network elements eliminates the need for a centralized network control unit, effectively mitigating management risks associated with single points of failure. Test data shows that the DDC architecture effectively increases bandwidth utilization compared to traditional networking solutions, significantly enhancing the throughput and operational stability of intelligent computing networks.

Advanced Cell Switching Mechanism

Building a 100% Non-Blocking Intelligent Computing Network

The DDC architecture utilizes a cell switching mechanism to ensure congestion-free, efficient transmission, and completely eliminate traffic imbalances. By employing byte-level, equal-length cell slicing, the cell switching mechanism ensures traffic balance across the entire link, effectively addressing all-to-all, full-line-speed traffic surges. Combined with cell switching, packet slicing eliminates sensitivity to traffic patterns and packet characteristics, achieving 100% load balancing and completely decoupling GPUs and NICs.

According to a Tolly test report, under an all-to-all traffic model, the RoCE network bus bandwidth based on the DDC architecture is comparable to that of an InfiniBand network. In bandwidths above 1G, it demonstrates superior performance, with an average improvement of 2.5%. This result fully demonstrates the leading performance of the DDC architecture in AI training scenarios.

Establishing DDC Core Framework Standards

Building an Open Intelligent Computing Ecosystem

With the accelerated development of intelligent computing, the overall computing infrastructure ecosystem is becoming increasingly mature, and the challenges of adapting diverse computing components are becoming increasingly prominent.

Adhering to the principle of open and collaborative development, H3C has collaborated with industry ecosystem partners to develop the DDC core framework standard based on OSF. This standard utilizes BGP to advertise Tunnel End Points (TEPs) to optimize the traffic scheduling capabilities of AI networks, focusing on addressing three core issues: load balancing, congestion control, and network reliability. This promotes flexible heterogeneity and interoperability at the network element level. The development of this standard system not only demonstrates H3C's deep expertise in lossless Ethernet technology, but also fosters industry collaboration, breaks down vendor barriers, and builds an open intelligent computing ecosystem.

Simplified O&M

Life Cycle Management for Intelligent Computing Networks

Intelligent computing networks connect massive numbers of devices and heterogeneous computing resources, and face complex and diverse failure scenarios, posing significant challenges to O&M. To address this, H3C has launched AD-AIDC, enabling full-lifecycle operations and maintenance management for intelligent computing networks.

The platform features one-click automatic onboarding, enabling plug-and-play device deployment without complex tuning, significantly simplifying the deployment process.

Furthermore, leveraging end-to-end interconnect visualization and end-to-end intelligent monitoring, it provides real-time insight into network operational status and ensures efficient O&M.

The platform supports cross-domain problem location through device-network collaboration, ensuring business continuity and providing a more stable, intelligent, and efficient management experience for intelligent computing networks.

The continuous exploration and upgrade of the DDC architecture not only represents technological innovation but also represents H3C's redefinition of the intelligent computing network paradigm. Its open, decoupled architecture transcends the limitations of traditional closed systems, enabling seamless integration of a diverse computing ecosystem and providing a sustainable connectivity foundation for the AI industry. Going forward, H3C will continue to promote technological innovation in intelligent computing networks and accelerate the empowerment of AI across all industries. In the AGI era, this transformation initiated from the network architecture layer will surely usher in a new stage of intelligent computing.