prev
next
Why H3C AI Intelligent Computing?
Purpose-built for research. Designed to scale.
-
Heterogeneous Comouting
-
Server-Network Synergy
-
Cloud-Native Flexibility
-
Energy-Efficient Design
Heterogeneous Comouting
Connect different computing power resource pools to support one-stop application deployment from models to infrastructure.
Server-Network Synergy
Lossless RoCE network solution with high throughput and low latency, helping to unleash the high computing power of the AI era.
Cloud-Native Flexibility
Integrating computing, storage, and network resources on the cloud to provide users with elastic, flexible, high-performance, and self-service computing services.
Energy-Efficient Design
A zero carbon and green computing power center that aligns with sustainable development goals.
Solution Architecture
From hardware to applications, a full-stack AI computing architecture built
for performance at every layer.
for performance at every layer.
Key Scenarios
Full-stack Intelligent Computing Center empowers scientific research.
Computing Power Area
We build a server portfolio for all AI scenarios, delivering robust underlying computing power to fuel massive data and algorithms, ensuring the full availability of AI resources.
- 1. Unify the computing power base to build a diverse heterogeneous computing power resource pool
- 2. High-performance computing power and high-speed network work together to improve the overall performance of the cluster.
- 3. Liquid cooling technology effectively reduces data center energy consumption and operating costs
- 2. High-performance computing power and high-speed network work together to improve the overall performance of the cluster.
Lossless ROCE Network
Intelligent lossless RoCE network creates a high-performance network with high bandwidth and zero packet loss to support scientific research and innovation.
- 1. 400G TOR switch, symmetrical uplink and downlink bandwidth, no bottleneck in performance
- 2. Standard configuration supports RoCE, end-to-end lossless transmission
- 3. Automatic deployment of the entire network based on graphical interface to improve efficiency
- 2. Standard configuration supports RoCE, end-to-end lossless transmission
Key Solution Components
Every component engineered for AI workloads at campus scale.
Let's Build Something Smarter.
Your research can't wait. Neither should your infrastructure.
FAQ
Q: What specific business value and research enablement does H3C’s Education AI Computing Solution bring to universities?
A: H3C’s solution drives paradigm shifts in academic research and industry-academy integration. It enables breakthroughs such as Peking University’s biomedical imaging facility and Tongji University’s autonomous systems. It supports interdisciplinary fields like ICs, robotics, and autonomous driving. With full-stack resource monitoring (100+ object types, 300+ metrics) and intelligent O&M, it ensures month-long training stability and improves end-network troubleshooting efficiency by over 90%. It also integrates open-source LLMs (LLaMA2, ChatGLM2.0, Baichuan) to lower AIGC adoption barriers.
Q: How can communication efficiency of GPU clusters be ensured during large model training to avoid congestion-induced compute idling?
A: H3C employs the AD-DC intelligent lossless control platform with RoCE. Using a standard Fat-Tree multi-rail architecture, each GPU server’s 8x400G/200G NICs connect to the same Leaf switch, enabling high-bandwidth, low-latency transport. Compute-storage-network collaboration and RDMA achieve zero packet loss. Tests show 20% higher training efficiency and 50% better collective communication performance, ensuring month-long training without interruption.
Q: How can AI computing centers shift from reactive fault handling to proactive O&M?
A: H3C provides full-stack visibility and intelligent O&M, managing switches, servers, lossless NICs, and GPUs—unlike peers focused only on network devices. Features include flow, latency, throughput, and congestion visibility to pinpoint issues in network, servers, or GPU parameters. At Zhejiang Lab, this cut fault location time drastically and raised troubleshooting efficiency by over 90%, enabling fast failure domain isolation and recovery.