Source: Chip Intelligence, Yangtze SecuritiesAs we all know, China’s high-end general-purpose chips are highly dependent on imports, especially CPU chips, where domestic manufacturers have a small presence in the PC and server fields, which are basically monopolized by international giants. The reasons behind this are, on one hand, high-end general-purpose chip technology originates overseas and continues to break through, leading domestic manufacturers significantly. On the other hand, international giants represented by Intel have built a mature commercial ecosystem and occupy strategic heights such as patent and standard formulation, forming multi-dimensional barriers that greatly increase the difficulty for domestic manufacturers to catch up.
In 2019, Huawei and other manufacturers were included in the “Entity List” by the United States, further amplifying the situation where the development of the technology industry is constrained. From the perspective of national security and business competition, there is a strong demand for the localization of high-end CPU chips in China, and under the government’s top-down emphasis, the government, industry, and research community are actively promoting it, with the capital market also contributing its efforts to jointly promote the continuous advancement of high-end CPU chip localization in China.
Today, we will focus on analyzing the costs of general servers that the market is concerned about and the trend of ARM servers replacing x86 servers.
1. A Brief Analysis of Server Cost Composition
Servers are important basic IT hardware, and the Chinese server industry is developing rapidly, with domestic first-line manufacturers expanding globally. According to IDC data, in 2018, China’s server market size reached $17.69 billion, a year-on-year increase of 57.82%; in 2018, the shipment of servers in China was about 3.3093 million units, a year-on-year increase of 29.33%. Driven by the demand for cloud computing and mobile internet, China’s server industry has developed rapidly in recent years.
Looking globally, according to IDC data, in 2018, the global server market size reached $88.305 billion, a year-on-year increase of 32.01%; in 2018, the global server shipment was about 11.79 million units, a year-on-year increase of 15.72%. Relying on cost advantages, Chinese server manufacturers have begun to enter the international market, which is expected to drive an increase in the shipment share of domestic manufacturers.
A server is a computer that provides services to other clients on the network. A server is a type of computer that, under the control of a network operating system, provides shared resources (including query, storage, calculation, etc.) to clients in a network environment. Its high performance is mainly reflected in the high-speed CPU computing power, long-term reliable operation, and strong I/O external data throughput capability. It mainly provides services such as network applications, databases, files, and printing to customers.
Hardware Composition of Servers: Servers are usually delivered to customers in hardware form, and their main hardware components include CPU, memory, hard disk (commonly referred to as the “three major components”), PCI-E expansion slots, network cards, RAID cards, power supplies, and fans. In addition to hardware, servers also require some driver management and support from the server operating system.
The Chinese Server Chip Market is Growing Rapidly. According to IDC’s forecast data, due to the strong demand for cloud computing, the server market is growing rapidly, and the server chip market size is also continuously increasing. In 2015, China’s server chip market size was about 15.5 billion yuan, and it is expected to reach 36 billion yuan by 2020, with a compound annual growth rate of about 18%.
Due to limited publicly available statistical data, we also take typical cases to break down server costs. Currently, the most popular and best-selling is the 2U rack server. We select the model with the highest sales under the discount price of the domestic typical server brand Inspur on JD.com as a representative, which is priced at 28,200 yuan, close to the average unit price of Chinese servers in 2017, taking its configuration as a basis for estimation and mutual verification with market statistical data.
CPU, memory, and hard disk account for about 60%~70% of the overall server price. Based on the configuration in the above image, we take the prices of corresponding components from public platforms as described in the table below. Among them, the price of CPU has less available sources, but the corresponding proportion calculation result is 29.70%, which is relatively close to the average value of 29.06% mentioned above and can mutually verify. The price of memory sticks is taken as the average value of the effective top 8 items under corresponding search items on JD.com, resulting in the memory accounting for about 13.44% of the total server price. The price of hard disks is taken from the corresponding component quotes on JD.com and Zhongguancun Online, estimating that the hard disk accounts for about 23.45% of the total server price.
In summary, the “three major components” account for about 66.59% of the total server price, falling within the 60%~70% range, which mutually verifies with the actual industry situation. Other components cumulatively account for about 30%~40% of the total server price. Based on the principle of prudence, it should be particularly noted that the above calculations are based only on some products and some models and are not completely accurate. At the same time, different models and configurations of servers also have different component models and quantities, and the prices of components and server products fluctuate with the market, which also has a certain impact on the proportion values.
2. The Commercial Value of Domestic CPU Replacement for Servers
The global server CPU market is mainly monopolized by overseas giants. According to overseas research statistics, AMD once launched the Opteron series, and in 2006 its server CPU market share once exceeded 20%. However, after Intel launched the Xeon series server chips, its market share quickly increased, and in recent years the server chip market has been basically monopolized by Intel.
Intel’s server-related business revenue reached $22.991 billion. Considering the nearly exclusive monopoly market share, Intel’s server chip revenue can be approximated as the global server chip market size. According to its financial report, in 2018, the revenue of this business reached $22.991 billion, a year-on-year increase of 20.60%. It can be seen that this is a rapidly growing and relatively large market.
3. What is the Development Prospect of ARM Chips in Servers?
The country strongly supports the development of the integrated circuit industry. Integrated circuits have risen to the level of national strategy in China. In recent years, driven by market demand and policy support, China’s overall strength in integrated circuits has significantly improved. However, it should also be noted that there is still a huge gap in innovation, frontier fields, and design links compared to international leading levels, facing significant challenges. Therefore, China continues to release relevant policies to support the development of integrated circuits and accelerate the pace of catching up.
Various entities in China actively participate, forming multiple domestic CPU routes. In addition to national policy support, huge potential commercial interests also drive various enterprises to actively attempt and promote the development of domestic CPUs. Based on the 2017 Chinese market as a base, and with a steady-state net profit margin of 20.55% as the normal profitability level, the annual net profit of the domestic server CPU industry is about 4.492 billion yuan. Currently, multiple CPU localization routes have formed based on x86 architecture, ARM architecture, and other architectures. Leading research institutions, national teams in informatization, excellent private enterprises, and local governments are actively participating to jointly promote the development of CPU localization.
In the field of server CPUs, due to the relatively higher performance requirements, the competition threshold is further raised, and there are fewer domestic manufacturers with the strength to participate. Currently, the domestic server manufacturers that the market is more concerned about are mainly Tianjin Haiguang and Shanghai Zhaoxin based on Intel architecture; Tianjin Feiteng and Huawei based on ARM architecture; and a few companies based on MIPS architecture, such as Loongson, and based on Alpha architecture, such as Shenwei.
However, Tianjin Haiguang has been placed on the “entity list,” and its x86 technology source AMD has also terminated cooperation with it. Fortunately, Tianjin Haiguang has obtained AMD’s first-generation Ryzen and EPYC Zen architecture technology and successfully launched its products. Additionally, the x86 technology source of Shanghai Zhaoxin—VIA’s x86 architecture technology authorization seems to have also expired, increasing resistance to future development. Overall, currently, domestic CPUs based on x86 architecture have encountered some setbacks.
In contrast, the development of domestic CPUs based on ARM architecture has not been smooth sailing either, and there are many difficulties in replacing x86 in the server chip field.
In the past, due to the performance disadvantages of ARM chips in the server field and the resistance from the ecological environment, its replacement of x86 chips has been fraught with difficulties. Applied Micro, Samsung, NVIDIA, Broadcom, Marvell, Qualcomm, AMD, and Huaxintong, among others, have all invested heavily but with unsatisfactory results. Among them, Qualcomm has even abandoned its server chip business, and the ARM server chip company Huaxintong, jointly established by Qualcomm and Guizhou, was also closed at the end of April this year. The remaining domestic ARM server chip companies are only Feiteng and Huawei, and Huawei has also been placed on the “entity list.” However, before that, it had already obtained permanent authorization for the ARMv8 instruction set architecture.
Shortcomings of ARM Architecture in the Server Field (Difficulties in Replacement):
Performance of the ARM System Continues to Improve, and the Ecological Environment is Gradually Friendly: Although it is difficult to replace in the server market, ARM has not given up. Last year, ARM released a new architecture, Neoverse, and hopes to significantly improve its performance through the new generation of processor designs, enhancing its competitiveness in the server and infrastructure fields.
According to the roadmap for Neoverse processors released by Arm, its first product is based on the 16nm Cortex-A72/A75 cores of the Cosmos platform. In 2019, it launched the new CPU core architecture of the 7nm process on the Ares platform, and in 2020, it will launch the 7nm+ process Zeus platform, and in 2021, it will launch the 5nm process Poseidon platform. Each generation of the platform will improve performance by 30% compared to the previous generation.
At the same time, Arm also stated that the Neoverse IP roadmap is specifically designed to meet unique requirements for performance, efficiency, and scalability, adapting to the continuous changes in data patterns, new workload challenges, and the growing demand for infrastructure to support a trillion connected intelligent devices.
From the perspective of commercial users, both Amazon and Huawei have successively released ARM-based server chips. Fujitsu has also shifted from the SPARC architecture system to the ARM v8 architecture system and launched chips, boosting market confidence. At the same time, with the prosperity of the mobile end, the underlying systems and upper-layer applications supporting the ARM-related ecology are gradually increasing. The three major Linux distributions (Ubuntu, Suse, and Redhat) have already released versions that directly support ARM servers.
Changes in Server Demand in the Cloud Computing Era. In the past, servers were mainly sold to enterprise customers, and individual enterprises emphasized performance, stability, and throughput of servers. However, in the cloud computing era, the demand for servers from internet and cloud service providers has changed. Due to their cluster model emphasizing overall performance and high demand for servers, they are gradually customizing servers in depth. For example, trimming unnecessary functions, strengthening according to business characteristics, saving space, reducing power consumption, and optimizing applications, etc. These changes in demand also bring new opportunities for ARM architecture servers to replace.
Cloud Service Providers Seek Alternatives to x86 Chip Solutions. Due to the huge demand for servers from cloud service providers, on one hand, chips account for about 30% of server costs, and on the other hand, power costs are also a significant expenditure item for cloud services. If chip costs and power consumption can be reduced, it will bring huge commercial benefits to cloud service providers. Based on this, ARM chips’ low price and low power consumption characteristics are very attractive. In November 2018, Amazon announced that its self-developed Graviton processor supports the new AWS EC2 A1 instance configuration, mainly targeting scalable applications such as microservices, web servers, and development environments, with costs reduced by up to 45%.
Huawei and Feiteng Lead the Charge to Replace x86 with ARM
As mentioned earlier, the main domestic ARM server chip manufacturers currently are Feiteng and Huawei.
From the perspective of specific deployment and business development based on ARM architecture, Tianjin Feiteng has developed earlier and currently has a good product line layout, commercialization, and ecological development, covering high-performance computing CPUs, server CPUs, desktop CPUs, and high-end embedded CPUs. It should be noted that both Feiteng and Huawei have obtained permanent authorization for the ARM instruction set architecture, and their cores are also independently designed, achieving self-control.
Currently, Feiteng has the FT-2000+/64 processor chip aimed at enterprise-level high-performance servers; moreover, Feiteng will launch the multi-processor server chip FT-2500/64 in the fourth quarter of this year, based on 16nm process, with a main frequency of 2.3GHz and 64 FTC663 cores.
Additionally, according to the information released by Feiteng, it has already cooperated with more than 400 domestic software and hardware manufacturers, and has initially established a relatively complete ecological system.
Similarly, Huawei has also started to build a series of products from chips to servers based on the ARM system. Currently, Huawei’s ARM architecture server-related industrial products have three main lines: Hi16 series CPU, Kunpeng series ARM CPU, and Taishan servers. Huawei began developing ARM architecture-related chips as early as 2004, covering multiple product lines, culminating in the heavy release of the Kunpeng 920 for servers this year.
The Kunpeng processor has achieved a competitive advantage. According to the release information of Kunpeng 920, it represents the highest performance ARM architecture CPU in the industry. This is mainly reflected in the following aspects: 1) High-performance processor core: Kunpeng 920 achieves independent design, with single-thread performance improved by 45% compared to the previous generation at the same frequency; 2) High integration: Utilizing the industry’s leading 7nm advanced process, integrating up to 64 cores, and supporting a maximum working frequency of 3.0GHz, achieving an extreme multi-core architecture and low latency performance to meet the computing needs of high concurrency and low latency applications. It integrates functions such as southbridge, network card, and SAS controller, saving server system energy consumption and enhancing system PCIe expansion capability; 3) High throughput capability: The industry’s first to support 8-channel memory controllers and PCIe 4.0, compared to the industry, memory bandwidth increased by 46%, and IO speed increased by 66%; 4) Hardware acceleration engine: Integrated data encryption/decryption, compression/decompression, EC (Erasure Code) error correction code, data verification and other hardware acceleration engines, capable of replacing processor instruction calculations, greatly improving data processing efficiency and enhancing data security.
Based on the advantages of ARM chips, Huawei’s Taishan servers support ARM application scenarios. Taishan servers focus on five application scenarios: big data, distributed storage, ARM native, HPC, and Web, helping to achieve efficient computing and reduce data center TCO.
In the field of industrial ecology, Huawei plans to continuously promote ARM industrial cooperation from three levels: hardware, basic software, and applications, intending to build a complete TaiShan industrial ecosystem in 3 to 5 years. Currently, the ecological cooperation of TaiShan has made considerable progress, and the ecological system has become relatively complete, with TaiShan and other systems having certain capabilities:
It is worth mentioning that in 2018, Feiteng’s ARM architecture servers, Shenwei’s Alpha architecture servers, and Loongson’s MIPS architecture servers were all included in the government procurement list, indicating the strong support of the state for the development of domestic server chips. The entry of manufacturers like Huawei is expected to further accelerate this replacement trend.
