On December 27th, the National Development and Reform Commission, the National Data Administration, the Central Cyberspace Administration, the Ministry of Industry and Information Technology, and the National Energy Administration jointly issued the "Deepening the Implementation of the" East Digital and West Computing "Project
The Implementation Opinions on Accelerating the Construction of a National Integrated Computing Power Network (hereinafter referred to as the "Opinions"). It is proposed that by the end of 2025, the comprehensive computing power infrastructure system will be preliminarily formed. Various types of newly added computing power in national hub nodes account for more than 60% of the country's newly added computing power, and the utilization rate of computing power resources in national hub nodes significantly exceeds the national average level; The 1ms delay urban computing power network, 5ms delay regional computing power network, and 20ms delay cross national hub node computing power network have been preliminarily implemented in the demonstration area; The ease of use and cost reduction of users using various types of computing power have significantly improved, and the network transmission costs between national hub nodes have significantly decreased. At the end of December, the computing power sector led two markets up in the early trading session.
On December 26th, the Beijing Artificial Intelligence Public Computing Platform (Shangzhuang) was released, and the first phase of 500P (FP16 dense) computing power was officially launched. The computing power platform is constructed and operated by Jingneng Group. The next goal is to achieve 100% green power operation as soon as possible, and strive to launch the second phase of 1500P public computing power in the first quarter of 2024, while reserving more than 2000P of computing power expansion space. On December 20th, Shanghai Computing Power Trading Platform 2.0 was officially released, and 12 enterprises were granted on-site licenses to access platform nodes. On December 16th, the China Computing Power Network (Western) Dispatch Platform was released, and the China Computing Power Network Western Operations Center was officially established. Whether it is an integrated computing power network, or a public computing power platform, computing power trading platform, or computing power network scheduling platform, they are closely related to computing power optimization and scheduling. Since the beginning of this year, policies such as the National Supercomputing Internet and the Action Plan for High Quality Development of Computing Power Infrastructure have been successively introduced, all of which propose to explore collaborative scheduling mechanisms for computing power. At the same time, local governments have also issued documents to promote the process of computing power coordination and scheduling. In the current era of big models, the demand for intelligent computing power has significantly increased, and the continuous breakthroughs and applications of AI algorithms represented by ChatGPT have greatly driven the demand for computing power. However, there are problems with supply-demand imbalance, high costs, and insufficient robustness in China's computing power. Realizing cross regional and on-demand flexible scheduling is expected to continue empowering computing power to be universal, making computing power a basic resource that can be accessed and used like water and electricity.
02 | The domestic supercomputing cloud service market size may reach 70 billion yuan. In the current international situation, the localization process of AI computing power is expected to accelerate. Specifically, 2024 will be the "customer initial selection and adaptation year", 2025 will be the "customer active procurement year", and domestic AI chips are expected to become the dominant year in 2026. In terms of general computing power, it is expected that the domestic server CPU market size will reach 19.8 billion by 2024 and 33.3 billion by 2027; The domestic server market is expected to reach 59.4 billion yuan in 2024 and 100 billion yuan in 2027 (CAGR+19% in 24-27). In terms of supercomputing, China's supercomputing CPUs were restricted and embarked on a path of self-improvement 8 years ago. It is expected that the CAGR from 2021 to 2025 will be 24.1%, and the size of China's supercomputing service market will reach 46.6 billion yuan by 2025. In order to promote the listing rate of domestic supercomputing power, the Ministry of Science and Technology launched the construction of supercomputing internet in 2023. It is expected that the long-term scale of China's supercomputing cloud service market is expected to reach 70 billion yuan, of which 63 billion yuan is dominated by enterprises, The future commercial enterprise market has enormous potential. Starting from 2022, the United States has been continuously increasing its efforts in the production, manufacturing, personnel, and supply of high-end AI chips in China. China has continuously launched support policies such as new infrastructure for computing power. It is expected that the total demand for AI computing power in China will reach 211.50 EFlops by 2024, of which 98.24 EFlops will be domestically produced, with a localization ratio of 46.45%. The corresponding demand for 910 units is 307000, with a market size of 30.7 billion, 38400 AI servers, and a market size of 40.933 billion.
03 | Huawei may become NVIDIA's second level GPU in terms of computing power. The Ascend processor is the world's first AI chip that covers all scenarios. Based on a unified da Vinci architecture, it can support differentiated computing power requirements for different scenarios of edge clouds and has smooth expansion from tens of milliwatts of IP to hundreds of watts of chips, covering the ability to deploy edge clouds in all scenarios. It is reported that the Ascent 910 integrates a CPU
Core, DVPP, and Task scheduler
The scheduler can reduce the interaction with the host
CPU interaction, fully leveraging its high computing power advantages; It also integrates HCCS and PCle
4.0 and ROCE
V2 interface, for building horizontal scaling (Scale)
Out and Scale
The Up system provides a flexible and efficient method, where a single Ascend 910 can independently complete the entire AI training process, minimize interaction with the host, and fully utilize its computing power. At present, the bottleneck of Ascension iteration lies in the ecosystem, and the AI computing power ecosystem centered on NVIDIA CUDA is the mainstream of today's world's large-scale models. Huawei's GPU capability can benchmark against NVIDIA A100. Huawei, as the highest level member of Premier, has officially joined the global AI open-source framework PyTorch community, PyTorch
Version 2.1 has synchronized support for Ascend NPU, and developers can directly access it on PyTorch
2.1 is based on the development of models by the Ascend. Based on PyTorch, the Ascend has adapted to mainstream industry models such as BLOOM, GPT-3, and LLaMA. After deep optimization, its performance can be on par with the industry. It is expected that the future iteration speed of Ascension will accelerate and replicate the development process of Kunpeng's computing ecosystem, blooming everywhere in various industries.
04 Jingtai Viewpoint | Focus on Future Business Models Computing Power Scheduling The future may include three business models: 1. one-time project construction and subsequent operation and maintenance of infrastructure and computing power platforms. 2. Operating commission fee for computing power scheduling platform: The computing power scheduling operator connects with the supply and demand sides, utilizes computing power, and is expected to draw a commission from it. 3. Other ecological costs of computing power scheduling: The computing power scheduling platform is not only an operation that provides computing power, but is expected to further develop into an application store in the future. Customers can not only purchase computing power, but also purchase related tools and applications, resulting in similar apps
The ecological cost of Store. To implement specific industry chain links and objectives: the construction and operation of computing power scheduling platforms in China include Zhongke Shuguang and the three major operators; Regional level platforms include Yunsai Zhilian, Shenzhou Digital, Zhuolang Technology, Radio and Television Yuntong, etc. The construction companies of the computing power scheduling infrastructure layer include Shensanda A, Yihualu, and the three major operators; The construction companies of computing power scheduling platforms include Sitiqi, Oriental Materials, Qingyun Technology, etc. Optical connectivity companies include Zhongji Xuchuang, Xinyi Sheng, Tianfu Communication, Taichenguang, Guangxun Technology, Deke Li, and others. Risk warning: The progress of computing power construction is not as expected, the business expansion of related companies is not as expected, and downstream demand landing is not as expected.
