In the Helingeer New Area of Inner Mongolia, a data center park covering 1402 acres, there is no roar of machines, but tens of thousands of AI computing tasks are running at high speed every moment. Refining a large-scale petroleum model and analyzing the complex working conditions under oil and gas wells; Incubate campus AI assistants for scientific research assistance and intelligent Q&A; Building a large-scale grass model and installing a "digital ruler" for grass grading and pricing... The China Mobile Intelligent Computing Center (Hohhot) node, with a computing power scale of tens of thousands of cards, is building a "super brain" in the era of artificial intelligence.
The so-called Wanka intelligent computing cluster, in simple terms, refers to connecting tens of thousands of AI acceleration cards together through a high-speed network, allowing them to work together as a whole. If a single AI chip is compared to a "worker", the Wanka cluster is a "super factory" with tens of thousands of people - each "worker" can work simultaneously and transmit information to each other with almost "zero latency".
During the 15th Five Year Plan period, the construction of the national integrated computing power network will accelerate comprehensively. The deepening development of large models and intelligent applications urgently requires a new generation of computing infrastructure to fully meet the stringent requirements of high concurrency, high throughput, multi precision, and multi task loads. Compared with small and medium-sized clusters, the core difference of Wanka clusters lies not in "too many cards", but in "collaboration".
At present, the China Mobile Intelligent Computing Center (Hohhot) has deployed about 20000 AI acceleration cards. When tens of thousands of AI chips train a model at the same time, any card failure or data transmission delay may render the entire training task ineffective.
We have improved the linear acceleration ratio of the cluster to over 90% through key technologies such as high-speed interconnection of 10000 cards and full chain monitoring of software and hardware, "Tang Xuwen, General Manager of Inner Mongolia Mobile's Computing Operations Department, told reporters. This means that after increasing the investment in computing cards, the overall computing efficiency is almost proportionally improved, rather than the efficiency loss of" too many people but too many hands ". At present, the longest uninterrupted training time for cluster models is 22 days, which is at the forefront of the industry.
Heterogeneous computing power management is another challenge. Among the thousands of AI chips, chips from different manufacturers often have their own independent driver and programming ecosystems, lacking a unified "communication standard" between each other, which makes it difficult to effectively improve the utilization rate of computing power. To this end, the center has achieved unified management and scheduling of chips from different brands by managing heterogeneous clusters at the bottom through the Nine Day Artificial Intelligence Platform.
Wanka High Speed Interconnection uses GPU pooling, card locking modules, and other technologies to make heterogeneous computing power like a well-trained 'joint fleet', each performing its own duties while working together. ”Tang Xuwen said that from thousands of cards to tens of thousands of cards, deployment tests not only the number of computing cards and theoretical peak performance, but also system architecture, network interconnection, storage access, energy efficiency control, resource scheduling, and engineering delivery capabilities. Only by breaking down these barriers between software and hardware can dispersed computing power operate efficiently like a well-trained symphony orchestra.
The Wanka intelligent computing cluster has high computing power density and usually consumes a large amount of electricity and energy. The intelligent computing center has deployed a large-scale cold plate liquid cooling system and implemented a "source grid load" green power direct supply project, providing new solutions for large-scale computing cluster heat dissipation, green power applications, etc.
Li Chenggui, Deputy General Manager of Cloud Computing Data Center of Inner Mongolia Mobile Computing Operations Department, introduced that the Intelligent Computing Center is innovatively exploring the integrated mode of "source network load storage computing", which will coordinate the management of new energy generation, power grid, servers, energy storage facilities and intelligent computing center. By deploying cold plate liquid cooling technology on a large scale, the PUE of the intelligent computing center is as low as 1.15.
When photovoltaic power generation is sufficient, the intelligent computing center can appropriately increase tasks such as model training and real-time data processing to improve the efficiency of new energy consumption. When wind power resources are abundant at night, offline inference, data storage and other business operations can be arranged for off peak operation. When new energy supply is insufficient, stable power supply can be guaranteed by energy storage and the power grid, "said Li Chenggui.
Computing power, like water and electricity, is dispatched on demand and transformed into a continuous stream of "digital kinetic energy", which will assist artificial intelligence in precise drip irrigation and deeply empower thousands of industries. Based on the 400G OTN backbone network, the one-way transmission latency from Hohhot to Beijing is currently as low as 3.5 milliseconds. The millisecond level network response meets the demand of AI large model training for massive data throughput, and enables real-time businesses that are extremely sensitive to time delay, such as financial high-frequency transactions, industrial Internet, and automatic driving, to deploy the "brain" on this intelligent computing base with ease.
The national integrated computing power network is not simply a pile of computing power, but connects computing power, electricity, data, models, and scenarios into an intelligent economic infrastructure network. The "15th Five Year Plan" for the development of the information and communication industry proposes to accelerate the construction of a national integrated computing power network, build a multi-level computing power facility system of "hub region edge", and orderly deploy intelligent computing clusters with tens of thousands of cards or more. Looking ahead to the long term, the weaving of intelligent computing nodes into a powerful network will continue to solidify the foundation of China's artificial intelligence development, assist in the deep integration of digital technology and the real economy, and seize development opportunities in the new wave of technological change. (Outlook New Era)