In the widely discussed AI computing power base, whether it is "CPU + GPU and other acceleration cards" in the cloud or "CPU+GPU+NPU" on the device side, CPU is indispensable.

The CPU, or central processing unit, is often composed of a control unit, an arithmetic logic unit (ALU) and registers, which are responsible for instruction reading, decoding and execution, which has high requirements for R&D technology and ecological construction, and has an important impact on the performance and operation efficiency of the computer.

As the computing and control center of computers, under the wave of AI development, the development of China's computer CPU industry has entered the "fast lane": the construction of intelligent computing centers has stimulated the demand for servers, and the upgrading of AIPC is also driving the growth of demand.

Compared with GPUs, the development of domestic CPUs is more mature. According to the data of the Toubao Research Institute, the localization rate of CPUs in China will be 30%-40% in 2022. At the same time, the localization rate of CPUs in China is still increasing, and the proportion of domestic CPUs in the recent centralized procurement by operators is close to 70%.

The explosion of AI has driven up CPU demand

With the rapid development of large models, the demand for intelligent computing power has exploded, and the growth of AI server shipments has driven the demand for cloud CPUs. As the core component of the server, the performance and power consumption of the CPU directly affect the overall performance of the server.

According to TechInsights, the global server market will reach $273 billion by 2028, growing at a CAGR of 18%, and servers designed for AI applications will account for 59% of the market share, at a CAGR of 49%. In terms of the Chinese market, IDC predicts that the size of China's AI server market will reach US$9.1 billion in 2023, a year-on-year increase of 82.5%; It will reach $13.4 billion in 2027, with a compound annual growth rate of 21.8%.

In the PC market, 2024 is recognized as the first year of AIPC, and AIPC is expected to drive a new round of replacement. The industry believes that the current development of AIPC, similar to the development of the early Windows system, is expected to replicate the success of the early Windows system in the future.

Canalys expects global PC shipments in 2023 to be 249 million units, a year-on-year decrease of 12.4%, and PC shipments are expected to be 267 million units in 2024, a year-on-year increase of 7.6%, AIPC shipments in 2024 will be around 20 million, and 60% of computers will have AI processing power in 2027, and shipments will exceed 175 million units in 2027.

Driven by AI servers and AIPC, the demand for CPUs will also grow rapidly. According to the data of the Toubao Research Institute, the size of China's computer CPU market has grown steadily and rapidly at an average growth rate of 6.91% from 2019 to 2021, and is expected to reach 510.949 billion yuan by 2027.

The AI wave set off by large models has not only boosted the demand for CPUs, but also put forward higher requirements for the functions of CPUs.

"Many AI applications can actually be computed with different XPU processing units, the CPU programming is more flexible, and the operator acceleration can be processed through a general-purpose compiler, which is more general and flexible than the NPU." Zhou Jie, general manager of the ecological strategy of this core technology, told Yicai, "For example, our Armv9 CPU contains vector extension instruction acceleration, which can better help run large language models on the device side. The latest iPhone 16 series also gets a CPU upgrade, including matrix instruction acceleration to better support those CPU-based Apple Intelligence applications. ”

Zhou Jie said that in the future, this core technology will continue to innovate in CPU products, provide more powerful vector and matrix processing capabilities, and combine GPU and NPU heterogeneous processing capabilities to more efficiently meet the growing demand for end-side AI computing power such as large language models and multimodal large models.

The battle over the instruction set architecture

Chip architecture is the foundation for designing and building microprocessors or other integrated circuits, defining key characteristics such as the chip's organizational structure, functional units, data flow, and instruction set. The instruction set architecture is the top-level design of the chip architecture, which defines the set of all instructions that the computer hardware can execute, and is the core part of the computer system design, which has an important impact on both the upper-layer software and the lower-layer hardware.

According to different design principles, the instruction set architecture is mainly divided into two camps: Complex Instruction Set Computing (CISC) and Reduced Instruction Set Computing (RISC), and from this derives x86, ARM, RISC-V and other instruction set architectures.

Introduced by Intel, the x86 is a CISC design with a rich instruction set that supports multiple addressing modes and complex operations. At present, processors equipped with x86 architecture instruction set are mainly used in PCs and servers. Due to Intel's strict control over patent licensing, the x86 market is basically monopolized by Intel.

In the server market, x86 is the mainstream architecture of CPUs. According to IDC data, the market share of x86 architecture servers in 2023 will be 88%, and the shipment volume of China's x86 server market will be 3.62 million units, and it is expected to grow by 5.7% in 2024.

The Arm architecture is based on the RISC design, and the Arm core microarchitecture IP selection is diverse, the design is simple and reliable, and the performance is excellent in the field of low power consumption. Arm's business model is to focus on licensing processor core architectures, rather than directly designing and manufacturing chips, so its neutral position allows Arm to widely promote the ecosystem through licensing. It is reported that at present, the Arm architecture has been widely used in the field of mobile intelligence represented by mobile terminal chips (mobile phones and tablets, etc.), set-top boxes, video surveillance, etc.

On the whole, at present, cloud servers mostly use the X86 architecture, and the edge side adopts the Arm architecture, but in recent years, the pattern has also changed, and the Arm architecture has gradually penetrated into the server field. "Now many Internet manufacturers around the world have begun to use the Arm architecture to design their own server chips, such as Amazon, Google, Microsoft, etc., as well as Ali in China." Zhou Jie told Yicai. In addition, AIPC is equipped with NPU, and the Arm architecture may have more advantages. Arm CEO Rene Haas said at COMPUTEX 2024 that Arm expects to capture more than 50% of the Windows PC market within five years.

RISC-V, an open reduced instruction set architecture, has also begun to emerge in recent years. RISC-V is completely open source and adopts the loose BSD protocol, which can be used by enterprises completely freely and for free, and also allows enterprises to add their own instruction sets without having to be open and shared, so as to support differentiated development; At the same time, the architecture is simple, short and concise, and different parts can be composed together in a modular way, so as to meet various application scenarios through a unified architecture. It is understood that at present, many domestic start-ups have begun to use RISC-V for chip design. SHD Group expects RISC-V-based SoC revenue to reach $92 billion in 2030, growing at a CAGR of 47% from 2021 to 2030.

"At present, the RISC-V architecture is more used in some fixed scenarios, such as IoT devices, but it needs more resources and R&D support to support the general ecosystem, such as PCs and servers." Zhou Jie told Yicai.

In addition, some enterprises have developed their own instruction sets. Loongson Zhongke(688047. SH) announced in 2023 that it will independently develop the Loong Arch architecture with completely independent property rights on the basis of the MIPS architecture, and Shenwei will develop the SW64 architecture based on the Alpha architecture.

"Ecology is a very important issue, and the ecology of X86 is undoubtedly the most mature, and the Arm architecture ecology is becoming more and more mature." Zhou Jie said, "There will be a certain gap between the ecology of self-developed architecture and the ecology of X86 and Arm architecture. If you need to support a more general ecosystem, the upstream and downstream of the entire industry chain will need a lot of investment, for example, each software vendor needs to adapt to the self-developed architecture and optimize the relevant software. ”

Comparison of the six major mainstream manufacturers

In terms of market structure, Intel and AMD are currently leading the general-purpose CPU market.

According to a report by Mercury Research, a CPU market tracker, in the second quarter of 2024, Intel will have a 78.9% market share in the PC market, and AMD will have a market share of 21.1%; In the server CPU market, Intel's market share is as high as 75.9%, and AMD's is 24.1%.

In terms of the domestic market, domestic CPU manufacturers are divided into three major schools: first, self-developed, such as Loongson and Shenwei as mentioned above; The second is Arm, such as the Great Wall of China (000066. SZ) subsidiaries Feiteng and Huawei HiSilicon both use Arm architecture; The third is X86, Shanghai Zhaoxin and Haiguang Information (688041. SH) all use the X86 architecture.

With the continuous advancement of information innovation, domestic CPUs may gain greater room for growth. According to Canalys data, the Chinese government and education sector will purchase 2.72 million PCs in 2023, accounting for 6% of national shipments. In March 2024, the central government procurement website issued an announcement stating that when purchasing computers, CPUs and operating systems should be included in the procurement requirements that meet the requirements of safety and reliability evaluation. Canalys believes that the above six brands are all brands that meet the conditions of domestic CPUs in the notice.

"The industry is working hard to promote the development of advanced manufacturing processes and upstream EDA tools, which have a great impact on the competitiveness of chip products." Zhou Jie said, "Combined with the current domestic industrial situation, it is not a big problem to design a chip, the key is how to produce it, and how to build a good ecology to help developers and end users use it better, which is particularly important and the embodiment of the company's core capabilities." ”