Author: Big Brother, Editor: Xiao Shimei
Thanks to the explosive growth of computing power demand due to the rapid development of artificial intelligence, NVIDIA, the global AI chip leader, has become the world's brightest star in the past two years.
In 2023, Nvidia shares soared by 2389%, becoming the top-performing component of the S&P 500; In 2024, the company is still refreshing record highs, and the latest market capitalization is close to 17 trillion US dollars, rising to the fifth place in the US stocks.
Despite the stock price momentum, Nvidia still faces many challenges, the most important of which undoubtedly comes from China.
The Chosen One
At present, NVIDIA's high-end chips have encountered many restrictions in the world's most important Chinese market. In order to cope with possible restrictions in the future, many companies in the Chinese artificial intelligence industry chain obviously want to take the initiative to reduce their dependence on NVIDIA chips, and instead begin to look for domestic alternative products, which has given domestic AI chip companies huge development opportunities.
According to the news of Wall Street** at the beginning of the year, Huawei received orders for at least 5,000 Ascend 910B chips from major Chinese Internet companies in 2023, and China's two major tech giants, Alibaba and Tencent, plan to significantly reduce orders for Nvidia chips in 2024, with orders much lower than the original plan before the export ban took effect.
In addition to Huawei, the biggest hope for domestic AI chips is Haiguang Information, which will be listed on the Science and Technology Innovation Board in 2022.
Haiguang Information's main business is high-end processors, which are mainly used in computing and storage devices such as servers and data centers. At present, the company has two product lines, mainly high-end general-purpose processors (CPUs) in the early days, and has begun to focus on coprocessors (DCUs) in recent years.
If Haiguang CPU is benchmarked against Intel and AMD's CPU products, DCU challenges NVIDIA's current GPU products.
The Data Center Unit (DCU) is a high-performance AI chip, which is one of the very few domestic AI acceleration cards that fully supports training and inference. The GPGPU architecture adopted by it is a technical branch of GPU, which is relatively simple in composition and has obvious advantages in product performance, versatility, cost, etc., and is the mainstream solution for accelerated computing, and has realized commercial applications in the fields of big data processing, artificial intelligence, and business computing.
Compared with the NVIDIA A100 products currently used by mainstream artificial intelligence companies around the world, the first generation of Haiguang Information DCU's production process is the same as the 7nm process, and the memory frequency and memory bit width are basically the same as A100, and the memory capacity, memory bandwidth and memory frequency are about 50% of A100, and the gap is relatively large is the interaction rate of multi-card collaboration, which is only about 30% of A100. In terms of overall performance, Haiguang Shensuan No. 1 should be able to reach more than 50% of the level of Nvidia A100.
The performance of Shensuan No. 2 released in the third quarter of 2023 is more than 100% higher than that of Shensuan No. 1, according to which it is estimated that this product has reached at least 80% of A100, and when Shensuan No. 3 is under development on the market, it should not be much of a problem to catch up with or even surpass A100.
In addition to the fact that performance is getting closer, equally important is the technical compatibility.
The key factor in NVIDIA's dominance of the GPU field is CUDA, a parallel computing platform and programming model launched in 2006, after which NVIDIA transformed GPUs from the field of gaming graphics cards to more general-purpose computing tools, attracting a large number of developers and world-leading enterprises to join, forming a strong software ecosystem.
In contrast,Haiguang DCU and NVIDIA's CUDA are highly similar in terms of ecology and programming environment, and have good adaptability to most mainstream large models at home and abroad such as Wenxin Yiyan, which means that the cost of NVIDIA CUDA ecological users to migrate to Haiguang DCU is very low, which lays the foundation for replacing NVIDIA in the future.
[Hope in the Village].
High-end processors have extremely high technical thresholds, and are the most comprehensive and important products in the chip industry.
Haiguang Information's technology was originally authorized by AMD in the United States, but after 2019, Haiguang Information was included in the "Entity List" of the U.S. "Export Control Regulations", AMD no longer provides related technical services, and the iteration and research and development of Haiguang's follow-up products are all completed by the company independently.
Independent R&D means a huge investment. From 2019 to 2021, Haiguang Information has invested 35 percent in R&D3.9 billion yuan, accounting for 95% of operating income35%, R&D personnel account for as much as 90%, almost all of them are engaged in technical research, and the determination to call the giants can be seen.
In 2022, the company's R&D investment will still exceed 2 billion, accounting for more than 40% of revenue, and among A-share companies with annual revenue of more than 5 billion, Haiguang Information ranks second in terms of R&D intensity, second only to BeiGene.
The extremely high proportion of R&D personnel and high-precision technology make Haiguang Information a typical talent-intensive enterprise: the company has a total of 1,500 employees, which is equivalent to a per capita income of more than 3 million, a per capita market value of more than 100 million yuan, and an average salary of 87990,000 (including social security), such a high per capita efficiency is extremely rare in the A** field.
With the company's technological progress and the wave of domestic substitution, Haiguang Information's performance has shown explosive growth in recent years.
In 2020, the company's revenue was only about 1 billion, and it had not yet achieved profitability, and in 2022, when it was listed, the company's revenue had grown to more than 5 billion, and the net profit exceeded 800 million, and the company's high-end processors had a market share of more than 50% in the domestic server industry, which is really half of the industry.
*: Straight flush.
The 2023 performance forecast shows that the company's performance still maintains rapid growth: of which the revenue is 56800 million-62600 million yuan, a year-on-year increase of 1082%-22.14%, net profit attributable to the parent company 11800 million-13200 million yuan, a year-on-year increase of 4685%-64.27%γ
From the perspective of business structure, although CPU is still the company's main revenue**, DCU has become an important support point for rapid growth in 2023, and is expected to account for about 30% of total revenue.
It can be expected that if Nvidia can still only launch some low-performance artificial intelligence chips to China in the future, and when Haiguang's deep computing No. 3 is released, the performance may greatly exceed Nvidia's domestic competitors, and Haiguang's DCU business revenue will increase significantly, and the proportion of revenue may exceed 50%.
At that time, the Chinese version of AI chips may become the mainstream of the market.
[National Will].
Behind the rapid rise of Haiguang Information, it fully demonstrates China's national will for self-improvement in science and technology.
Shareholder data shows that at least six of the top ten shareholders of Haiguang Information are controlled by state-owned legal persons, with state-owned assets holding more than 50% of the shares. Among them, Sugon, a subsidiary of the Chinese Academy of Sciences, acquired Haiguang Information 1 in 2018With 94.8 billion shares, it has become the largest shareholder of Haiguang Information, with a current shareholding ratio of 2796%, in addition, the Chinese Academy of Sciences holds 265%γ
With the background of the Chinese Academy of Sciences and state-owned holdings, Haiguang Information is regarded as a veritable "national team player" in the chip industry.
*: Straight flush.
The national computing power industry plan and the domestic AI industry chain are also providing assistance for Haiguang's catch-up plan.
In recent years, Haiguang Information has been deeply involved in the construction of the "Eastern Data and Western Computing" advanced computing center and the "new infrastructure" intelligent computing center, combined with the in-depth layout and industrial resources of the Chinese Academy of Sciences and Sugon in the computing industry, the company has formed a good business synergy with major shareholders, becoming the core force supporting the foundation of China's computing power, and the circle of friends in the AI industry chain has also been expanding.
The company's DCU products have been certified by Internet companies such as Alibaba and Alibaba, and have launched joint solutions to build a full-stack AI infrastructure that integrates software and hardware produced in China. In addition, a large number of domestic leading AI companies such as iFLYTEK, SenseTime and Yuncong have been transplanted and run on the Haiguang DCU platform.
In order to improve the domestic ecology of the general computing platform, Haiguang Information also initiated the establishment of the Haiguang Industrial Ecological Cooperation Organization, which unites universities, scientific research institutes and upstream and downstream enterprises in the industry inside and outside the system to tackle key problems in a collaborative manner that combines production, education and research to provide assistance for breaking through key common technologies in the industry.
Historical experience shows that there are many successful precedents in China's science and technology field for latecomers to catch up and even overtake in corners.
From the panel in the early years to the current photovoltaic industry, semiconductors, new energy vehicles and operating systems, Chinese companies have achieved independent and controllable industry with their own strength.
We believe that with the strong support of the first-class system and the super industrial chain, China's AI chips, represented by Haiguang Information, will surely raise the banner of domestic computing power and become another example of China's technology industry catching up and being independent and controllable.
Nvidia's good days in China won't be long!
Disclaimer of Warranties
The content of this article related to listed companies is the author's personal analysis and judgment based on the information publicly disclosed by listed companies in accordance with their legal obligations (including but not limited to temporary announcements, periodic reports and official interactive platforms, etc.); The information or opinions contained herein do not constitute any investment or other business advice, and Market CapWatch disclaims any liability for any actions resulting from the adoption of this article.
endββ