Context and Need:
With the explosive growth of AI models scaling into trillions of parameters, the demand on data center networking is unprecedented. AI training requires extremely high-performance, low-latency, and lossless networks to ensure efficient data flow among thousands to tens of thousands of GPUs or network cards. Network bottlenecks or inefficiencies can directly reduce training throughput by up to 30% and increase training times by 25%, severely impacting operational efficiency and economics.
H3C’s DDC (Diversified Dynamic-Connectivity) architecture, paired with its flagship computing cluster switch S12500AI, aims to meet these exacting needs, enabling ultra-large-scale, lossless, high-throughput networking tailored for AI data centers.
Key Highlights of the DDC-Based Solution:
1. Massive Scalability and Flexibility
-
Supports cluster interconnections from 1,000 up to 70,000+ network cards.
-
NCF (Network Control Fabric) box supports up to 128 ports at 800G OSFP speeds.
-
NCP (Network Control Ports) support a mix of 400G Q112 Ethernet and 800G OSFP Ethernet, allowing flexible adaptation to the most common NIC form factors.
-
Open networking design avoids single points of failure by eliminating centralized control units, improving network stability and manageability.
2. Advanced Cell Switching for 100% Non-Blocking Performance
-
Uses byte-level equal-length cell slicing, breaking packets into uniform cells.
-
This cell-switching approach achieves perfect load balancing and eliminates traffic congestion and imbalances.
-
Fully decouples GPUs and NICs traffic patterns, meaning network performance is consistent regardless of workload or packet types.
-
Independent lab tests (e.g., Tolly report) show RoCE networks with DDC outperform traditional solutions, surpassing InfiniBand in bandwidth usage efficiency by 2.5% or more at bandwidths above 1G.
3. Open Core Framework Standards & Ecosystem
-
H3C has developed an open DDC core framework standard based on OSF (Open Switching Framework).
-
Uses BGP to advertise Tunnel End Points (TEPs), enhancing traffic scheduling through improved load balancing, congestion control, and reliability.
-
This open standard encourages interoperability, breaking vendor lock-in and promoting a diverse, multi-vendor AI data center network ecosystem.
4. Simplified Operations and Maintenance (O&M) with AD-AIDC
-
The AD-AIDC platform supports full lifecycle O&M for intelligent computing networks.
-
Offers one-click automatic onboarding of devices, enabling plug-and-play deployment without complex manual tuning.
-
Provides end-to-end network visualization and intelligent real-time monitoring for operational insight.
-
Facilitates cross-domain fault location via device-network collaboration, enhancing reliability and reducing downtime.
Why This Matters:
-
Network efficiency and stability are critical to fully unlocking the performance potential of massive AI training clusters.
-
H3C’s DDC solution pushes network throughput from typical 30% utilization up to nearly 100%, meaning AI workloads can run faster, more reliably, and more cost-effectively.
-
By embracing open standards and flexible design, H3C fosters an ecosystem that allows diverse hardware and software components to work together seamlessly.
-
Simplified O&M reduces operational complexity and cost, crucial for managing the sprawling, heterogeneous infrastructure of AI data centers.
Looking Ahead:
H3C’s DDC architecture redefines intelligent computing networks with its open, scalable, and lossless design. This evolution aligns perfectly with the AGI (Artificial General Intelligence) era’s infrastructure demands, promising to accelerate AI innovation across industries by providing a robust, future-proof network foundation.
Jika Anda ingin menggali lebih dalam mengenai solusi H3C Workspace Cloud Desktop dan bagaimana produk ini dapat mengoptimalkan infrastruktur TI di perusahaan Anda, jangan ragu untuk menghubungi H3C Indonesia atau PT. iLogo Infralogy Indonesia. Tim kami siap memberikan informasi lengkap serta konsultasi yang disesuaikan dengan kebutuhan spesifik bisnis Anda, sehingga Anda dapat mengambil keputusan terbaik untuk kemajuan organisasi Anda.
