(Report Producer/Analyst: Chen Haijin, Debang Securities)
01 Company Overview: A leading enterprise of domestic CPU and DCU
The company is a leading enterprise of domestic general-purpose processors and coprocessors. Founded in 2014, the company is one of the few integrated circuit design enterprises in China with both high-end general-purpose processor (CPU) and co-processor (DCU) R&D capabilities, focusing on R&D, design and sales of high-end processors used in computing and storage devices such as servers and workstations.
General-purpose processor (CPU) products have strong performance, rich product lines, and are compatible with x86 instruction set, with significant ecological advantages. The company's CPU products have built-in multiple processor cores, integrate common high-performance peripheral interfaces, and have a complete software and hardware ecological environment and a complete system security mechanism.
According to the requirements of different application scenarios for high-end processor computing performance, function, power consumption and other technical indicators, Haiguang 7000 series products, 5000 series products and 3000 series products are provided respectively, among which Haiguang 7000 and 5000 series are mainly used in servers, and Haiguang 3000 series are mainly used in workstations and edge computing servers.
At present, the company's CPU products have been iterated to the third generation (Haiguang No. 3), and the product performance has been significantly improved, Haiguang No. 3 is the company's main sales product in 2023, and as of 23H1, Haiguang No. 4 and Haiguang No. 5 are in the research and development stage. At the same time, the company's CPU series products are compatible with x86 instruction set and international mainstream operating systems and application software, with rich software and hardware ecology, and have been widely used in important industries or fields such as telecommunications, finance, Internet, education, and transportation.
Coprocessor (DCU) products are compatible with "CUDA-like" environments and have a wide range of application scenarios. The company's DCU products are based on the GPGPU architecture, with a large number of built-in computing cores, strong parallel computing capabilities and high energy efficiency ratio, and are suitable for computing-intensive applications such as vector computing and matrix computing.
At present, the company's DCU products are 8000 series, as of 23H1, the first generation of the 8000 series product Shensuan No. 1 has achieved mass production and sales, Shensuan No. 2 will be released soon, and the research and development of Shensuan No. 3 is progressing smoothly. At the same time, the company's DCU products are compatible with ROCm GPU computing ecology, can be better adapted to international mainstream commercial computing software and artificial intelligence software, and can be widely used in big data processing, AI, business computing and other computing-intensive application fields, mainly deployed in server clusters or data centers.
1.1. Shareholding structure: Jointly established a subsidiary with AMD and licensed x86 technology
The company established subsidiaries with AMD, Haiguang Microelectronics and Haiguang, to integrate x86 technology licensing.
Since 2018, the company's top two shareholders have been Sugon and Chengdu State-owned Assets (including Chengdu Industrial Investment Co., Ltd., Chengdu High-tech Investment Co., Ltd. and Chengdu Jitri Co., Ltd., which are acting in concert), with shareholding ratios of 27.96% and 17.00% respectively, and the shareholding structure is relatively stable and clear.
In 2016 and 2017, Haiguang Microelectronics and Haiguang Integration signed the "Technology License Agreement" with AMD respectively, agreeing that AMD will license high-end processor-related technologies and software to two joint ventures.
AMD delivered the relevant intellectual property rights to Haiguang Microelectronics and Haiguang Integration in accordance with the agreement, and Haiguang Microelectronics and Haiguang Integration paid relevant fees to AMD in accordance with the agreement, and there are no intellectual property disputes or potential disputes.
Many of the company's directors, executives and core technical personnel have worked in Sugon, AMD, and Intel, and their R&D and management experience can provide stable support for the company's development.
High-quality R&D team is an important part of the company's core competitiveness, as of 1H23, the company has 1,382 R&D technicians, accounting for 90.33% of the total number of employees, most of the core R&D personnel have more than 20 years of experience in high-end processor R&D, and the backbone R&D personnel have the experience of successfully developing x86 processors or ARM processors, providing strong technical support for the company's self-developed chips.
1.2. Financial analysis: The scale of revenue has grown rapidly, and profitability has continued to improve
The company's operating income has grown rapidly. In 2022, the company's operating income will be 5.125 billion yuan, a year-on-year growth rate of 121.83%, and its growth is mainly due to 1) industrial development and the substantial increase in the demand for domestic processors in many industries, which has promoted the company to occupy a larger market share;
In 2021, the company turned losses into profits, and the net profit attributable to the parent company increased significantly in 2022. The company's continuous losses from 2018 to 2020 are mainly due to the relatively small scale of operating income in the early stage of the company's product listing, the large investment in R&D funds since the establishment of the company, the implementation of multiple equity incentives for key employees and the corresponding confirmation of a large amount of share payment.
In 2021, the company achieved profitability for the first time since its establishment, mainly due to the rapid increase in market demand and the large-scale sales of the company's DCU products.
In 2022, the net profit attributable to the parent company increased by 145.65% year-on-year, mainly due to the increase in income, interest income and government subsidies.
The company's revenue structure is gradually diversified, and the downstream application fields are broad.
The company's initial products were mainly high-end server products 7000 series, and in recent years, the 5000 series for the mid-range server market and the 3000 series for the workstation market have also begun to contribute revenue rapidly. In 2021, the company's 8000 series Shensuan No. 1 will achieve sales, and the company's revenue structure will be further diversified.
In terms of downstream application fields, in 2019, the company's Internet and telecommunications industry market promotion has achieved initial results, based on the testing and recognition of the company's products in the terminal fields such as telecommunications, finance, and Internet in the early stage, in 2020, the company has achieved market breakthroughs such as centralized procurement of telecom operators and shortlisting in the financial industry, with strong market demand, and the products are mainly used for telecom operators' cloud service resource pool systems to support cloud business, banking and securities company query and transaction systems, Internet search, computing services, storage and other applications. In the same year, the company entered the special catalog, and the market demand in the terminal fields such as party and government grew rapidly.
The company's gross profit margin rose steadily, and 1H23 achieved a high level of 62.87%.
From 2019 to 2021, the company's gross profit margin increased from 37.31% to 55.95%, showing a gradual growth trend overall, mainly due to the small business scale of the company's early stage and the high proportion of amortization of self-developed intangible assets in operating income, which increased the unit cost of products. In 2022, the company's gross profit margin will be 52.42%, a decrease of 3.5pcts from 2021, mainly due to the increase in packaging and testing costs in 22 years.
Compared with peer companies, the company's gross profit margin in 22 years and before is basically the same as that of the domestic CPU leader Loongson Zhongke, and the gross profit margin in 1H23 greatly exceeds that of Loongson Zhongke and moves closer to Cambrian, the domestic AI chip leader.
The company's expense ratio showed a downward trend, profitability continued to improve, and R&D investment continued to strengthen.
Due to the rapid growth of the company's operating income and the improvement of operational efficiency, the company's sales expense ratio decreased to 1.58% and the management expense ratio decreased to 2.63% in 2022. In 2022, the company's total R&D expenditure will be 2.067 billion yuan, with a R&D investment rate of 40.33%, and the total R&D investment is significantly higher than that of peer companies.
As of 1H23, the company has 1,382 R&D technicians, accounting for 90.33% of the total number of employees, and has obtained 550 invention patents, 81 utility model patents, 3 design patents, 200 integrated circuit layout design registration certificates and 206 software copyrights.
02 CPU: The information and innovation industry continues to promote domestic substitution, and the company is expected to fully benefit
2.1. Global CPU market: Intel and AMD duopoly pattern, x86 ecosystem occupies the right to speak
The CPU (Central Processing Unit) is the computing and control core of the computer, the final execution unit of information processing and program operation, and the core component of computers, servers and other equipment.
Due to weak demand for consumer electronics due to macroeconomic pressures, the CPU market size declined in 2022, with global PC shipments reaching 286 million units in 2022, down 16.2% year-on-year, according to Gartner data.
Canalys learned through the PC inventory survey of channel partners that the number of channel dealers with less than one week of inventory increased from 33% in January '23 to 41% in June, that is, the inventory situation in 2Q23 further improved. Under the influence of continuous inventory correction and seasonal factors, the PC market is expected to gradually begin to recover from 3Q23 (according to Yole and Canalys).
2.1.1. x86 is a mainstream architecture with significant ecological advantages
Applications eventually need to be translated into "instructions" to run on the CPU, so the instruction set used is especially important for the design of the CPU. According to the instruction set adopted, CPUs can be divided into two categories: Complex Instruction Set (CISC) and Reduced Instruction Set (RISC).
The complex instruction set architecture and the reduced instruction set architecture are designed based on two different instruction set ideas, which have their own different characteristics: the complex instruction set is rich in instructions, flexible in addressing mode, with the microprogram controller as the core, the instruction length is variable, the function is powerful, and the execution efficiency of complex programs is high; the structure of the reduced instruction set is simple, easy to design, and has a high execution energy efficiency ratio.
The x86 architecture has a higher level of software and hardware maturity and has significant ecological advantages.
In the field of operating systems, Windows and Linux are compatible with the x86 architecture; in terms of application software, thanks to the openness of the instruction set to independent software developers and the consistency of the operating system of the application platform, the threshold for technical development has been significantly lowered, making the maturity of the software and hardware environment under the x86 architecture have obvious advantages compared with other architectures; in terms of instruction set expansion, the mainstream x86 CPU has extended the instruction sets such as MMX, SSE, AVX, SIMD, etc., and has achieved great success.
Whether it's in the PC or server world, x86 is the mainstream choice. According to Mercury Research data, the market share of x86 architecture in 1Q23 consumer products is about 85.2%, and according to IDC data, the market share of x86 architecture in 1Q23 servers is 90.3%.
2.1.2 The .AI server market is driving the booming demand for CPUs
CPUs can be used in different devices such as servers, workstations, personal computers (desktops, laptops), mobile terminals, and embedded devices, and there are certain differences in their architecture, functions, performance, reliability, energy efficiency ratio, and other technical indicators according to different application fields.
The server has high-speed data processing capabilities, powerful I/O data throughput capabilities, good scalability, and needs to run reliably for a long time, and its CPU chips have more demanding requirements in terms of performance, reliability, scalability, and maintainability.
Therefore, server processors are the processors with the strongest data processing capabilities, the most complex design processes, and the highest reliability. Applications for servers include real-time analytics, 5G applications, artificial intelligence, machine Xi, finance, big data, and cloud computing.
Intel and AMD occupy about 90% of the market share (shipment volume) in the server CPU market, jointly building a monopoly position in x86.
According to IDC data, Intel's shipments accounted for 71.1% of global server CPU shipments in 1Q23, AMD's shipments have grown rapidly since 2019, accounting for 17.3% in 1Q23, and the shipments of other x86 architecture CPU manufacturers accounted for only 1.8%. As enterprises become more diversified in cloud workloads, the proportion of ARM-based server CPUs has gradually increased, accounting for 9.6% of shipments in 1Q23.
Looking ahead, AI servers are booming and are expected to be a growth driver for the server CPU market.
According to The next platform, unless the growth of AI models slows down, or the price of AI training and inference computing falls, it is reasonable to believe that AI servers could account for about half of the total server market revenue by 2026 or 2027.
According to the guidance given by TSMC at the 2Q23 results meeting, at present, the demand for AI server processors (CPUs, GPUs and AI accelerators) accounts for about 6% of TSMC's total revenue, and this demand will grow at a CAGR of nearly 50% in the next five years, accounting for more than ten percentage points of TSMC's revenue. We believe that with the rapid development of AIGC applications driving the demand for AI servers, the server CPU market is worth looking forward to in the future.
2.2. The rapid development of "2+8+N" has opened up the CPU space of domestic servers
The information and innovation industry is developing rapidly at the pace of "2+8+N". The information and innovation industry, that is, the information technology application and innovation industry, is in the same vein as the "863 Plan", "973 Plan" and "Nuclear High-tech Base", aiming to realize the independence and controllability of the mainland's information technology industry.
(1) "2": 2022 is the final year of the three-year comprehensive promotion stage of party and government information innovation, and the localization transformation of the electronic document system above the municipal level has been basically completed, and in 2023 and beyond, it will further sink and expand to the district and county level. In September 2022, the state issued Document No. 79 to comprehensively guide the development and progress of the state-owned information and innovation industry, requiring 100% of central and state-owned enterprises to complete the replacement of information and innovation by 2027, of which chips, complete machines, operating systems, databases, and middleware are the most important industrial chain links.
(2) "8": Under the guidance of the party and government departments, eight key industries such as finance, telecommunications, electric power, and transportation have also begun to accelerate the pace of independence and control. In 2020 and 2021, two phases of the pilot project were launched, which put forward specific requirements for the procurement of IT software and hardware by financial institutions and the localization of office and business systems.
As an important part of the basic hardware sector of the information and innovation industry, servers and CPUs are expected to accelerate the process of domestic substitution.
The information and innovation industry can be divided into four major sectors: basic hardware, basic software, application software, and information security. Among them, the basic hardware includes: chips, servers/PCs, storage, switches, routers, etc., the basic software includes: databases, operating systems, middleware, etc., the application software includes: OA, ERP, office software, government applications, flow version software, etc., and the information security includes various products such as border security products and terminal security products.
CPU has been widely used in consumer electronics, industrial electronics, Internet of Things, data centers and other fields, and has become one of the most concerned categories in the information and innovation industry.
With the implementation of the information and innovation industry policy, the market size of domestic processors has increased rapidly.
Compared with other domestic processors, Haiguang processor products have excellent performance and ecological advantages, high product compatibility, and can be directly adapted to complex application scenarios of end customers, with high user acceptance. With the continuous rise of users' recognition of Haiguang products, the company has accumulated a large number of procurement needs, and domestic servers and workstations based on Haiguang processors are in strong demand.
According to the company's 2022 issuance and registration link inquiry letter reply announcement, important large-scale centralized procurement orders in the industry have been disclosed one after another, among which financial and telecom operators are the two key industries with the fastest promotion of localization, and related large-scale centralized procurement orders can be used as benchmark cases that affect market expectations.
As of February 27, 2022, China Mobile and China Telecom have won the bid list for 2021-2022 server centralized procurement, with a total of 7.347 billion yuan of domestic server centralized procurement, of which 3.784 billion yuan is used for server machines equipped with Haiguang processors, accounting for 50%.
In 2021, the demand for Haiguang from domestic end customers will be increasingly strong, and the prospects for continuing operations are good. With the launch of server and workstation products equipped with Haiguang processors by Inspur, Lenovo, Xinhua III, Tongfang, ZTE and other machine manufacturers, the number of domestic server manufacturers equipped with Haiguang processors is growing.
2.3. Strong performance + mature ecology, the company's CPU products have significant advantages in domestic substitution
2.3.1. The performance of the company's server CPU products is leading in China, and the introduction, absorption, and breakthrough are realized
The company introduced and absorbed the Zen architecture, which laid a good foundation for the performance of the company's CPU products.
Since 2016, the company has begun to use AMD's Zen architecture for the research and development of x86 processors, and AMD completed the delivery of the entire set of source code of the licensed technology in December 2016 and September 2018 respectively, and the company has completely obtained and digested the AMD licensed technology.
On the basis of AMD authorized technology, according to the actual needs of the information industry, through continuous independent innovation, the company has continuously enriched the functions of Haiguang processor, continuously improved the performance of Haiguang processor, and maintained the technical competitiveness of Haiguang processor.
On the basis of the Zen architecture, the company continues to make breakthroughs with independent research and development strength, and the product performance has been rapidly improved.
When designing Haiguang No. 1 products, the company focused on "digestion and absorption" of imported technologies, and Haiguang No. 1 CPU was mainly based on the source code delivered by AMD for research and development, and at the same time carried out some independent innovations, including security enhancement technology based on Chinese cryptographic algorithms. Haiguang No. 1 CPU basically meets the technical indicators of the same generation of AMD authorized technology.
On the basis of the successful research and development of Haiguang No. 1 CPU, Haiguang No. 2 CPU has carried out a large degree of independent innovation on the microarchitecture of the CPU processor core, and the product performance has been improved. According to the relevant data on the official website of SPEC, the relevant performance indicators of AMD's second-generation product of the same type based on licensed technology, AMD EPYC 7542, are still stronger than Haiguang 7285, but the gap is not large.
Starting from the Haiguang No. 3 processor, the company has independently carried out iterative development of CPU products, and the overall measured performance has increased by about 45% compared with the previous generation.
Haiguang-3 CPU has further innovated independently in microarchitecture: optimized the branch predictor in the finger extraction unit to achieve a more accurate branch prediction algorithm and higher prediction bandwidth, optimized the PCIe controller to upgrade the bandwidth from 8Gbps to 16Gbps, optimized the memory controller to increase the memory frequency from 2667MHz to 3200MHz, increased the retransmission mechanism when the memory read and write error was added, improved the RAS capability of the server system, optimized the on-chip network, increased the bandwidth, modified the internal communication protocol, and increased the working frequency。
The Haiguang No. 3 series chips have the highest specifications of 32 cores and 64 threads, with as many as 128 PCle4.0 channels, according to the company's investment records, compared with the previous generation of products, the overall measured performance of Haiguang No. 3 has increased by about 45%. In terms of SOC design, I/O bandwidth, finger taking unit, functional module, defense mechanism, etc., Haiguang-3 has been optimized to varying degrees, and its comprehensive performance has jumped significantly.
The performance of the company's CPU products can be equivalent to the products released by international manufacturers in the same period, and it is in a leading position among domestic manufacturers.
Since Intel's processor products are in a dominant position in the market, the company distinguishes its product positioning against Intel.
By comprehensively comparing factors such as processor market positioning, number of cores, and product price, the company selected 6 Xeon Platinum series products released by Intel in 2020 (the same period as Haiguang 7285) (which can reflect the performance level of mainstream CPU products released by Intel in 2020), and compared the performance with Haiguang 7285 (using the industry's internationally accepted test program SPEC CPU 2017 test data).
Through comparison, it can be seen that the measured performance of the SPEC CPU 2017 of Haiguang 7285 CPU is generally comparable to the measured performance of the mainstream processor products released by Intel, an international leading chip design company, at the same time.
The company attaches great importance to the security of the processor, and effectively improves the security of the Haiguang processor by expanding the instruction set of the security algorithm and natively supporting trusted computing.
Processor security technologies mainly include trusted execution environment, cryptographic acceleration, trusted computing, and vulnerability defense.
In terms of trusted execution environment, based on automatic data encryption and decryption, it effectively prevents security attacks and provides an advanced full-process security execution environment on cloud computing.
In terms of cryptographic operation acceleration, Haiguang CPU integrates cryptographic co-processors that meet the national secret standard, and supports the national secret standard SM2, SM3, and SM4. The processor has a built-in trusted computing platform and supports the Chinese standard TPCM and the international standard TPM 2.0.
The trusted computing platform not only implements the root of trust required for trusted computing, but also actively measures and monitors the system, and takes timely measures when abnormalities are detected to effectively protect the system and meet the requirements of DJCP 2.0.
In terms of CPU vulnerability defense, Haiguang CPU is immune to circuit breaker vulnerabilities, and uses effective software and hardware technologies to defend against ghost vulnerabilities and side-channel vulnerabilities. High-end processors are the "brain" and "center" of all high-end information equipment, and are the cornerstone of information security.
2.3.2. The company is a scarce x86 architecture server CPU target in China, with broad growth space
The instruction system belongs to the interface between hardware and software in the computer, and is the top priority in building the CPU ecosystem.
At present, the CPU industry is dominated by two major ecosystems: the Wintel system based on the X86 instruction system and the Windows operating system, and the AA system based on the ARM instruction system and the Android operating system.
In the 80s of the last century, Intel developed its own X86 instruction system architecture, rapidly expanding its market share and building ecological advantages by virtue of its first-mover advantage, and gradually occupied the desktop CPU market by forming a "Wintel" alliance with the Windows Alliance; ARM is based on the low-power mobile market with the efforts of Apple, Qualcomm, Samsung, Huawei, Nvidia, etc., with a series of advantages such as open source instruction system, heterogeneous computing, and customization.
At present, there are 6 domestic CPU companies, namely Loongson Zhongke, Dianke Shentai, Huawei HiSilicon, Feiteng Information, Haiguang Information, and Shanghai Zhaoxin.
According to the type of instruction system adopted, it can be roughly divided into three categories: (1) Haiguang Information and Shanghai Zhaoxin, which use the X86 command system, (2) Huawei Hisilicon and Feiteng Information, which use the ARM command system, and (3) Loongson Zhongke and Denke Shentai, which used MIPS-compatible command system and Alpha-like command system respectively in the early days, and have now independently developed their own command systems.
Haiguang CPU series products are compatible with x86 instruction set and international mainstream operating systems and application software, with rich software and hardware ecology, and are expected to gain a relatively larger market share in domestic server replacement and new demand compared with other server CPU manufacturers.
2.3.3. The company is rooted in the local market and fully benefits from the high-quality industrial chain
Haiguang products have been recognized by well-known domestic server manufacturers such as Inspur and Lenovo, and their local advantages are significant.
The company took the lead in developing high-end general-purpose processors and co-processor products in China, and realized commercial applications.
Compared with foreign manufacturers, Haiguang focuses on the local Chinese market, understands the needs of Chinese customers better, can provide more safe and controllable products, has localized competitive advantages, and has been widely recognized by domestic industry users.
According to the company's prospectus, since 2018, the products of many well-known domestic server manufacturers such as Inspur, Lenovo, Xinhua III, and Tongfang have been equipped with Haiguang CPU chips, and have been successfully applied to customers in the financial field such as Industrial and Commercial Bank of China and Bank of China, as well as customers in the energy and chemical fields such as PetroChina and Sinopec, and have been widely used in the data center business of telecom operators.
The company cooperates with the "Photosynthetic Organization" to promote the construction of localized ecology and lay out long-term development.
The full name of "PS-CCO" is "Haiguang Industrial Ecological Cooperation Organization", which is to unite upstream and downstream enterprises, universities, scientific research institutes, industry enterprises and other related innovative forces in the industrial chain around the domestic general computing platform, promote the common development of the members of the cooperation organization, and jointly build an inclusive and prosperous information technology ecosystem.
On April 28, 2020, the company initiated the establishment of the "PS-CC Organization", and by the end of 2022, the PS-CC Organization has 1000+ member units, 500+ adaptation certification manufacturers, 1000+ product adaptation certification, 10 regional branches, and 15 adaptation centers.
03 GPGPU: The AI model drives the demand for computing power and opens the company's second growth curve
3.1. GPU/GPGPU: An indispensable AI computing chip
GPGPU will become the mainstream of computing coprocessors. There are multiple technical routes based on different design ideas, including GPGPU, ASIC, FPGA, etc. Among them, the representative companies of GPGPU include NVIDIA and AMD; using ASIC technology, many large companies have developed co-processor products, including Intel's Phi and NNP, Google's TPU, Huawei's Ascend, Cambrian Siyuan, etc.; and FPGAs based on Intel and Xilinx, there have been many dedicated co-processor products.
Considering factors such as performance, energy efficiency ratio, and programming flexibility, GPGPU has very obvious advantages in the field of coprocessor applications, and is currently widely used in commercial computing, artificial intelligence, and pan-artificial intelligence.
GPUs were originally developed as ASICs specifically designed to accelerate specific 3D rendering tasks. For image processing, because each pixel needs to be processed, the amount of data is very large, and the CPU is more suitable for a single workload task with a relatively small number of cores that requires high core performance, which leads to the CPU being unable to be efficient in processing such a large amount of data, and the GPU came into being.
GPUs have become more programmable and flexible over time, and graphics processing is still one of the main functions of GPUs, but it has evolved into a more versatile parallel processor capable of handling a growing number of use cases.
3.2. Global GPU market: The AI wave is driving the demand for computing power, and GPU chips have broad prospects
3.2.1.AI large model drives the rapid increase in computing power demand
At present, AI large models have emerged one after another, and the number of large models of more than 10 billion has grown rapidly after accumulation.
After a long process of technology accumulation such as shallow machine Xi and deep machine Xi, the new wave of AI started with the pre-trained model based on Transformer in 2017 and broke through the technological singularity after the release of the GPT-3 model in 2020.
Since 2022, AIGC (AIGC) has achieved rapid development in technology and industry.
Taking the version evolution of the OpenAI GPT series as an example, from GPT-1 (2018.06) to GPT-2 (2019.02), GPT-3 (2020.05), GPT-3.5 (2022.03), and GPT-4 (2023.03), the number of parameters and corpus continue to upgrade, and the parameters of AI super-large models have reached hundreds of billions or even trillions of orders.
The demand for AIGC is rising, and the development of the industry relies on the underlying computing power support of high-process chips.
At present, applications such as AIGC, high-end servers, high-performance computing, and 5G are developing rapidly, and these applications are inseparable from the underlying support of high-process computing power chips. Taking the AIGC large model as an example, according to the overseas technology website lambdalabs, the model size of GPT-3 reaches 175 billion parameters, and it takes 355 years to train using a single NVIDIA V100GPU, which puts forward a massive demand for computing power chips.
Based on 1024 NVIDIA A100 chips, the training time of GPT-3 is more than 25 days.
According to our calculations, it takes 1024 A100 80G to complete a GPT-3 model training with a parameter of 175 billion by renting online computing power under ideal conditions, which takes about 25 days and costs about $930,000. However, the actual cost and time tend to be longer, for example, it is almost impossible for the model to complete the training smoothly in one go, and so on. In addition to the training computing power requirements, the inference computing power requirements will continue to increase with the commercial popularity of the model and the number of active users attracted.
We believe that the rapidly increasing demand for large model training and large model inference is expected to greatly promote the increase in the market size of GPU-based computing power chips.
AI servers are developing rapidly, and the AI chip market is mainly occupied by NVIDIA and AMD.
TrendForce expects AI server shipments to be about 1.2 million units in 23, +38.4% year-on-year, accounting for about 9% of total server shipments, with a CAGR of 22% in 2022~2026 and a 46% increase in AI chip shipments in 2023.
As the core of parallel data processing, GPU is the core increment of AI servers. According to IDC data, in 2021, NVIDIA accounted for 91.4% of the enterprise GPU market, and AMD accounted for 8.5%.
The scale of China's intelligent computing power has also grown rapidly, with a CAGR of 48%.
According to IDC's forecast, the scale of China's intelligent computing power will reach 268 EFLOPS in 2022, and it is expected to enter the level of 10 trillion floating point calculations per second (ZFLOPS) in 2026, reaching 1271 EFLOPS, with a CAGR of 48% from 2022 to 2026.
According to IDC data, in 2021, the market share of servers used for inference in China's data centers has accounted for more than half, reaching 57.6%, and it is expected that by 2026, the workload used for inference will reach 62.2%.
3.2.2. China's demand for computing power continues to grow, which is expected to open up the growth space of domestic computing power chips
Driven by AI, the scale of China's accelerated computing server market is growing rapidly, which is expected to promote the rapid increase of China's server GPU market.
According to IDC data, the size of China's acceleration server market will reach US$6.7 billion in 2022, a year-on-year increase of 24%, of which GPU servers are still dominant, accounting for 89% of the market share.
In the context of the deployment of AI-accelerated applications, the size of China's accelerated server market is expected to reach $16.4 billion in 2027, with a CAGR of 20% from 2022 to 2027. And since GPUs are still the mainstream AI acceleration chips in China, the server GPU market is also expected to continue to grow driven by AI.
GPUs are the largest increment in the industry chain, and the demand for single-order servers under the AI trend has increased significantly.
Ordinary servers are only equipped with single or dual SIM cards, while AI servers are generally equipped with four or more GPUs or AI chips due to the large amount of computing required. In addition, the computational load of large AI models in training and inference is huge, and low-end processors cannot meet their computing needs.
For example, in the NVIDIA DGX H100, it is equipped with 8 NVIDIA H100 GPUs with a total GPU memory of up to 640GB, and each GPU is equipped with 18 NVIDIA NVLinks, with a bidirectional bandwidth of up to 900GB/s between GPUs. Based on the unit price of $40,000 per NVIDIA H100 GPU, the value of the DGX H100 GPU is $320,000, which is the largest increment in AI servers.
3.3. The company's GPGPU products have excellent performance and are expected to benefit from the explosion of domestic demand for computing hardware
3.3.1. The performance indicators of the company's DCU products have reached the international leading level
The company's deep computing No. 1 index has reached the level of the same type of high-end products in the world, and can be used for model training and inference.
Based on the GPGPU architecture, the company's DCU products are one of the few GPGPU products in China with full-precision floating-point data and various common integer data computing capabilities, which can be applied to artificial intelligence servers and support complete large model training.
In 2022, Shenzhen Computing No. 1 has achieved commercial applications in the fields of big data processing, artificial intelligence, and business computing, and the company has reached certification with a number of domestic enterprises and research institutes that develop large models. The company's DCUs will be widely used in big data processing, artificial intelligence, business computing and other fields in the future.
3.2.2. Compatible with "CUDA-like" environment, with significant ecological advantages
NVIDIA CUDA (Compute Unified Device Architecture) 为当今主流AI编译生态，海光DCU采用“类CUDA”取得生态优势。
CUDA is a general-purpose parallel computing architecture developed by NVIDIA that includes an instruction set (ISA) applied to NVIDIA GPUs and an internal parallel computing engine for GPUs.
Due to the high similarity between ROCm and CUDA in terms of ecology and programming environment, CUDA users can quickly migrate to the ROCm platform at a low cost, so ROCm is also known as "CUDA-like".
Therefore, Haiguang DCU coprocessor can better adapt to and adapt to the international mainstream commercial computing software and artificial intelligence software, with rich software and hardware ecology, and can be widely used in computing-intensive applications such as big data processing, artificial intelligence, and commercial computing, mainly deployed in server clusters or data centers, providing high-performance and energy-efficient computing power for applications, and supporting high-complexity and high-throughput data processing tasks.
3.3.3. The company's DCU products are expected to achieve resonant growth in synergy with CPU products
With the synergy of CPU and DCU products, the company can provide customers with supporting solutions for a mature ecosystem.
As one of the few integrated circuit design enterprises in China with both high-end general-purpose processor and co-processor R&D capabilities, the company can provide Haiguang CPU and Haiguang DCU hardware product portfolio. Mutual authentication and continuous optimization, and active integration into the open source community at home and abroad, have initially formed a complete domestic software and hardware ecological chain based on Haiguang CPU and Haiguang DCU, and continue to provide complete and mature solutions for "new infrastructure" projects such as intelligent computing centers, industry users, AI manufacturers and Internet enterprises.
Baidu PaddlePaddle and Haiguang DCU have completed mutual verification, and Haiguang products can fully support training and inference applications. In September 2021, Baidu PaddlePaddle Depth Learning Xi Framework and Haiguang Artificial Intelligence Accelerator Card DCU series carried out installation and deployment tests, basic function tests and stability compatibility tests, and the joint test results showed that Baidu PaddlePaddle Depth Learning Xi framework can be successfully installed on Haiguang DCU series and Haiguang 3000, 5000, and 7000 series CPU environments, and can operate reliably, stably, and with high performance to meet the critical application needs of users.
04 Earnings Forecast & Valuation
We forecast that the company's operating income from 2023 to 2025 will be 57.13/79.92/10.932 billion yuan, with a growth rate of 11%/40%/37% respectively, and a comprehensive gross profit margin of 57.7%/56.7%/57.3%. The business assumptions are as follows:
CPU Products:
1) Revenue: Driven by the party, government and industry information innovation, the scale of the domestic server CPU market is expected to grow rapidly, considering that the company is expected to fully benefit as a leading domestic CPU enterprise, we predict that the company's CPU revenue is expected to grow rapidly from 2023 to 2025.
2) Product structure: The company adheres to the product strategy of "sales generation, verification generation, and R&D generation", with Haiguang No. 2 as the main sales product in 2022 and Haiguang No. 3 as the main sales product in 2023. We predict that Haiguang 4 will achieve sales in 2024 and become the main product in 2025, and Haiguang 5 will achieve sales in 2025. At the end of the life cycle of each generation of products, the company will reduce the gross profit margin of the product in order to achieve clean sales.
DCU Products:
The company's DCU products are one of the few computing hardware in China with both training + inference capabilities, and we expect that with the increase in computing power demand brought by the deployment of AI large and small models, the scale of the company's DCU products is expected to achieve rapid growth from 2023 to 2025.
We believe that 2021-2022 is the initial stage of the company's DCU products, the sales revenue is small, and the scale effect has not yet been formed.
We selected Loongson Zhongke, Montage Technology, and Jingjiawei as comparable companies.
As of December 14, 2023, the average P/E of comparable companies in 2023-2025 is 152/82/61x.
We predict that the company's net profit attributable to the parent company in 2023-2025 will be 10.4/14.5/2.07 billion yuan, and the closing price on December 14, 2023 corresponds to 168/121/85 times P/E in 2023-2025.
We believe that the company, as the leading target of scarce CPU+DCU in A-shares, is expected to fully benefit from the rapid development of information innovation and the rising demand for computing power, so it can enjoy a certain valuation premium.
05 Risk Warning
High risk of customer concentration. According to the company's 2022 annual report, the company's sales to the top five customers accounted for 97.36% of the current operating income, and the customer concentration was high, of which related sales accounted for 56.82%.
The progress of R&D is less than expected, and market competition is intensifying. The company has maintained a high level of R&D investment, and the R&D process has great uncertainty, if there are major changes in the external market, the existing technology is replaced by other new technologies, it is easy to cause the R&D progress to be less than expected. From the perspective of the industry, the market competition has become more and more fierce in recent years. In the field of general-purpose processors, domestic manufacturers such as Feiteng, HiSilicon, and Zhaoxin are competing with each other, and in terms of coprocessors, Jingjiawei and Biqian Technology cannot be underestimated. If the company's R&D progress slows down in the future and cannot cope with the fierce market competition, it will adversely affect the company's industry position, operating ability and market share.
Risks associated with Entity List restrictions. On June 24, 2019, the U.S. Department of Commerce's Bureau of Industry and Security added the company to the "Entity List" of the U.S. Export Control Regulations, and the company's procurement and sale of "controlled items" containing a high proportion of U.S. restricted technologies will be restricted. On October 7, 2022, the U.S. Department of Commerce's Bureau of Industry and Security issued a relevant announcement that the company was included in the list of objects subject to the proposed export control regulations.
Due to the scarcity and proprietality of the products or services provided by some suppliers such as wafer fabs, EDA vendors, and IP vendors, if Sino-US relations deteriorate further in the future, the relevant restrictions will escalate, which will easily adversely affect the company's overall costs, R&D progress, supply chain security, and ability to continue operations.
——————————————————
The report belongs to the original author, only for Xi! If there is any infringement, please delete it privately, thank you!
The report comes from [Foresight Think Tank]