Zhengzhou Unicom Intelligent Computing Project "Compute as a Service" model builds a secure and controllable Intelligent Computing foundation
Adopting a computing power service procurement model, using domestic training servers and intelligent computing platform software to build a 100P training and promotion integrated cloud computing power service capability, providing continuous and secure computing power support for three years, and promoting the high-quality development of large model training and AI applications.
Background and Challenges
With the rapid development of artificial intelligence and large model technology, Zhengzhou Hanbo Chuangxin Technology Co., Ltd. is facing increasingly growing computing power demands in intelligent computing and model training. However, the traditional hardware procurement model has high investment, rapid depreciation, and heavy maintenance, making it difficult to flexibly match the business rhythm. At the same time, the security of the computing power supply chain has become a key concern - relying on imported chips has uncertainty, and the substitution of domestic products is imperative. How to quickly obtain high-specification, sustainable, and controllable training and promotion integrated computing power capabilities under the premise of controlling costs has become a core issue that needs to be solved urgently.
-
Large gap in computing power demand
High specification requirements for large model training and inference, enterprises' own computing power is difficult to meet peak demand
-
Pressure from heavy asset investment
The traditional hardware procurement model has high fixed asset investment, depreciation risk, and heavy cost of building operation and maintenance teams
-
Supply chain security risk
Uncertainty in the supply of imported chips, urgent demand for domestic substitution of computing infrastructure
-
Difficult to guarantee service quality
Lack of clear service level agreements and compensation mechanisms, it is difficult to constrain the stability and availability of computing services
Solution
Adopt the "Computing Power as a Service" model, with domestic training servers and domestic computing power intelligent platform software as the core, to build a 100P training and promotion integrated cloud computing power service capability. The ownership of the equipment belongs to the service provider, and users can use computing power services as needed, and continue to obtain safe and reliable high-performance computing power support within a three-year service cycle, achieving dual goals of light asset operation and supply chain security.
HARDWARE — Domestic computing power hardware layer
32 domestic training servers; 3 business management servers; 400G RoCE computing network; hot storage and general storage system
PLATFORM — Intelligent computing platform software layer
National computing power intelligent computing platform; model adaptation and optimization services; cloud integration and deployment services; multi-level network switching management
SERVICE — Computing power service delivery layer
100P training and promotion integrated computing power; three-year continuous service guarantee; quarterly service delivery; SLA level agreement constraints
Core characteristics
Nationalization-oriented, full-stack self-controllable
Adopting national training servers and national computing power intelligent computing platform software, fully realizing nationalization replacement from hardware to platform level, effectively ensuring the safety and controllability of the computing power supply chain, laying a technical foundation and practical experience for the subsequent larger-scale computing power construction.
Service-oriented delivery, light asset operation
Delivered as a computing power service, users can use computing power on demand without bearing the cost of hardware procurement and O&M management pressure. The ownership of the equipment belongs to the service provider, effectively reducing the investment risk and operating burden of fixed assets.
High specification configuration to meet large-scale model training
Adopting a 400G RoCE high-speed computing network and high-performance GPU training server to build a 100P training and inference integrated computing power. Multi-level network switching equipment and hot storage system work together to fully meet the high-performance computing needs of large language model training and inference.
Perfect SLA, service quality is guaranteed
Establish a clear service level agreement and compensation mechanism, provide a complete set of security protection equipment and software, and form a complete closed loop from the availability of computing power, service response to security assurance, to ensure the stable and reliable operation of computing power services within a three-year service cycle.
Application Scenario Display
Profit and value
Build a secure and controllable domestic intelligent computing infrastructure through a service-oriented model and promote high-quality development in the field of AI and large models.