Analysis of the Ethernet chip market structure: domestic substitution ushers in incremental space

To join the " Smart Car Expert Optoelectronics Packaging CPO Industry Exchange Group", please add 18512119620 (same as WeChat), note company-name-position to join the group

01 .

AI accelerates the iteration of Ethernet switch chips, and domestic substitution is imminent

The AI ​​wave drives a surge in data volume, and high-end switches help data centers grow

Ethernet switches and Ethernet switching chips are widely used in the entire information industry. With the development of technologies such as 5G, cloud computing, the Internet of Things, and artificial intelligence, the boundaries and capabilities of the network will be unprecedentedly expanded and improved, and its vigorous development will drive the information industry into the era of full interconnection. Since the first Ethernet switching device was launched in 1989, the industry has developed rapidly, and related products have greatly improved in forwarding performance. The port rate has developed from 10M to 800G, and the switching capacity of a single device has also increased from tens of Mbps to tens of Tbps.

Ethernet switches are network devices used for network information exchange and are key devices for interconnecting various types of network terminals. Ethernet switches provide high-speed network connection ports to the outside world, directly connected to hosts or network nodes, and can provide electrical signal paths and business processing models for any number of network nodes of access devices.

The core component chips of switches are highly dependent on imports, and the industry presents an oligopolistic competition pattern. Switches are composed of resources such as PCB, electronic components, operating systems, switching chips, CPUs, optoelectronic chips, and optical modules. Among them, chip costs account for the highest proportion, which is 32%. At present, the main chips in the upstream of switches are highly dependent on imports and have high technological content, resulting in high switch costs and an oligopolistic competition pattern in the market. According to IDC data, in 2020, the top five brand manufacturers in China's switch market were Huawei, H3C, Ruijie Networks, Cisco, and Maipu Technology, with market shares of 40.0%, 36.2%, 12.2%, 5.4% and 1.5%, respectively, and the concentration of the top five manufacturers reached 95.3%.

Participants in the switch market are divided into three categories: branded switch, white box switch and bare metal switch manufacturers. Among them, branded switches are relatively expensive to purchase due to the binding of software and hardware. In addition to Huawei and Cisco, which mostly use self-developed chips to produce Ethernet switches, companies such as H3C, Ruijie Networks, and Maipu Technology are the main end customers or direct customers of Centec Communications. International market participants mainly include Arista, Edgecore, Dell, etc. There are few companies in the Chinese market that focus on white box switches, and white box switches account for a relatively low market share of the overall switch market.

The surge in data volume drives the demand for storage and computing. In terms of data volume, the 21st century is considered the information age. Various emerging applications such as artificial intelligence, machine learning, blockchain, 5G, and the Internet of Things are generating large amounts of data every moment and applying them to various scenarios in people's daily lives. Data has become the fifth largest factor of productivity after land, labor, capital, and technology. According to IDC estimates, the total amount of global data will reach 85ZB in 2021 and will reach 175ZB in 2025, with a CAGR of about 20%. Huge amounts of data have created fertile soil for the birth of AI models. When OpenAI upgraded from the GPT-2 model to the GPT-3 model, the number of its parameters increased from 1.5 billion to 175 billion.

Generative AI helps increase the demand for high-speed switches, with data centers as the main application market. With the increasing application of generative AI, the amount of data brought by large model training and reasoning has increased dramatically, driving the continuous increase in the demand for high bandwidth and low latency in data center networks. As the core equipment of data center networks, switches have also ushered in new technological changes and growth opportunities. According to IDC forecasts, the global data center switch market continues to expand, and its growth rate exceeds the overall market. It is expected that sales will reach US$20.88 billion in 2024, accounting for more than half of the overall market share, becoming one of the main application markets for switches. According to data from the China Business Industry Research Institute, switches occupy the vast majority of the market share in China's network equipment market. The market size in 2022 will be approximately 59.1 billion yuan, a year-on-year increase of 17.96%, and approximately 68.5 billion yuan in 2023. It is expected that the size of China's switch market will increase to 74.9 billion yuan in 2024.

The capital expenditures of the four Internet giants in North America are increasing rapidly, and traditional data centers may usher in restorative growth. According to Bloomberg data, the total capital expenditures of Microsoft, meta, Amazon, and Google in 23Q4 were US$43.007 billion, up 8.87% year-on-year and 16.25% month-on-month, achieving a year-on-year positive turnaround in 23 years. We expect that capital expenditures will remain stable year-on-year when the downstream market is cyclically downturned in 24Q1Q2. In 2024, the four giants will increase spending on AI-related servers, chips, network equipment, and other digital devices. According to Bloomberg estimates, the total capital expenditures of Microsoft, meta, Amazon, and Google in 24 will reach US$181.349 billion, a year-on-year increase of 22.99%. Affected by a new round of capital expenditure growth by major global cloud computing companies and increased demand for high-end AI servers, data centers are expected to usher in restorative growth.

The development of cloud computing drives the demand for data centers. China's cloud computing is in a period of rapid growth, and the market demand for IaaS infrastructure such as data centers will gradually increase. Since 2019, domestic cloud computing giants and communication operators have continued to increase their investment in the field of cloud computing, and data centers as underlying facilities will directly benefit. The development of cloud computing business and traffic growth directly drive the growth of cloud vendors' demand for and investment in data centers.

According to statistics from the China Academy of Information and Communications Technology, the scale of China's cloud computing market will reach 455 billion yuan in 2022, a year-on-year increase of 40.9%, which is much higher than the global growth rate. It will be about 616.5 billion yuan in 2023, and the market size is expected to increase to 837.8 billion yuan in 2024 and 1,178 billion yuan in 2025. Overall, compared with developed countries in Europe and the United States, my country's cloud computing market started late, and there is huge room for market improvement. We expect it to maintain rapid growth in the next few years. The development and construction of cloud computing and large data centers require a large number of Ethernet switches, and also put forward higher requirements on the performance of Ethernet switching chips.

NVIDIA has created the world's first AI Ethernet network platform, which improves network performance by 1.6 times compared to traditional networks. NVIDIA Spectrum-X is the world's first Ethernet network platform built specifically for AI, which can improve network performance by 1.6 times compared to traditional Ethernet network platforms, and can speed up the processing, analysis and execution of AI workloads. The platform integrates NVIDIA Spectrum X Ethernet switches and NVIDIA BlueField-3 super network cards to solve the bottleneck of traditional Ethernet in AI applications.

The demand for traditional Ethernet data centers remains unabated, and Nvidia Spectrum-X helps upgrade. Since acquiring Mellanox for $6.9 billion in 2019, Nvidia has been promoting the development of its Infiniband technology. Infiniband is a computer network communication standard for high-performance computing. It has extremely high throughput and extremely low latency. It is used for data interconnection between computers and has been widely used in cloud vendors' data centers. But not every data center can handle Infiniband, and Ethernet is still the default setting for most data centers. Therefore, the demand for traditional Ethernet still exists and will not change.

Traditional Ethernet designs focus on high average throughput, but in deep learning and AI use cases, GPUs primarily communicate with each other, and the requirements are different. The Spectrum-X architecture integrates advanced network-level remote direct memory access (RDMA) and enhanced congestion control capabilities to significantly improve performance and avoid hot spots. Adaptive routing and noise isolation technologies further optimize Ethernet data transmission efficiency and reliability.

NVIDIA released the data center architecture for the next three years, and the switch industry may benefit. On June 2 , 2024, NVIDIA disclosed the technology roadmap for the next three generations of data center semiconductors, with new GPU architectures updated every year: Blackwell Ultra GPU (8S HBM3e 12H) in 2025; Rubin GPU (8S HBM4) in 2026; Rubin Ultra GPU (12S HBM4) in 2027, a new generation of Arm-based Vera CPU, and NVLink 6 Switch (3600GB/s). NVIDIA's AI computing power has increased 1,000 times in 8 years, and it can train large models at a lower cost. For example, the energy consumption required to train GPT-4 with 1.8 trillion parameters and 8 trillion tokens has been reduced from 1,000GWh of the Pascal architecture 8 years ago to 3GWh of Blackwell.

Data centers are expected to have millions of GPUs in the future. NVIDIA announced that it will launch new Spectrum-X products every year, providing higher bandwidth, more ports, more powerful software feature sets and programmability, and continuously improving AI Ethernet network performance. Spectrum-X800 is designed for tens of thousands of GPUs, X800 Ultra is designed for hundreds of thousands of GPUs, and X1600 can be expanded to millions of GPUs. AI cloud service providers such as CoreWeave and Lambda are the first to adopt Spectrum-X.

02 .

The port rate of Ethernet switching chips continues to increase, and the market concentration moves toward 100G and above

Definition and architecture of Ethernet switching chip. Ethernet switching equipment consists of Ethernet switching chip, CPU, PHY, PCB, interface/port subsystem, etc., among which Ethernet switching chip and CPU are the core components. Ethernet switching chip is a special chip used for switching and processing large amounts of data and message forwarding. It is an application-specific integrated circuit (ASIC) optimized for network applications. Its internal logical path consists of hundreds of feature sets, which maintain extremely high data processing capabilities while working in coordination. Therefore, its architecture implementation is complex; CPU is used to manage login and control protocol interaction; PHY is used to process the physical layer data of the electrical interface.

Some Ethernet switching chips integrate CPU and PHY inside the Ethernet switching chip. Ethernet switching chips comply with the OSI model (including physical layer, data link layer, network layer, transport layer, session layer, presentation layer and application layer) at the logical level, and mainly work at the physical layer, data link layer, network layer and transport layer, providing high-performance bridging technology (layer 2 forwarding) for the data link layer, high-performance routing technology (layer 3 routing) for the network layer, security policy technology (ACL) for the transport layer and below, as well as data processing capabilities such as traffic scheduling and management.

Working principle of Ethernet switching chip: After the message/data packet to be transmitted enters the Ethernet switching chip through the port, the packet header field is matched first to prepare for flow classification; then the hardware security detection is performed through the security engine; the data packets that meet the security requirements are switched at Layer 2 or routed at Layer 3, and the matching data packets are subjected to relevant actions (such as discarding, limiting the speed, modifying VLAN, etc.) through the flow classification processor; for the data packets that can be forwarded, they are placed in the buffers of different queues according to 802.1P or DSCP, and the scheduler performs queue scheduling according to priority or WRR and other algorithms, and executes the flow classification modification action before the port sends the data packet, and finally sends it out from the corresponding port.

Since each key application scenario of the network system adopts a networking architecture similar to access, aggregation and core, a series of Ethernet switching chip products are required. According to bandwidth and application classification, Ethernet switching chips can be divided into 100M, 1G/10G, 25G/40G/100G and 400G. The downstream application scenarios of Ethernet switching chips are divided into Ethernet switching equipment for enterprise networks, Ethernet switching equipment for operators, Ethernet switching equipment for data centers and Ethernet switching equipment for industrial use.

The increase in port speed drives the iteration of chip demand structure. The rapid development of the digital economy in recent years has promoted the rapid development of technology industries such as cloud computing, big data, the Internet of Things, and artificial intelligence, as well as the digital transformation of traditional industries, all of which have put forward new requirements for network bandwidth. The demand for 100G and above Ethernet switching chips is gradually increasing, and 400G ports will become the mainstream port form within the next-generation data center network.

According to data from Frost & Sullivan, in terms of China's commercial Ethernet switching chip market, in 2020, the market size of Ethernet switching chips with port rates of 10 Gigabit, 1 Gigabit and above and 100G and above accounted for the largest proportion, at 30.2%, 28.2% and 24.1% respectively; it is expected that by 2025, the market size of China's commercial Ethernet switching chips of 100G and above and 25G will increase significantly, accounting for 44.2% and 16.3% respectively, and the average annual compound growth rate from 2020 to 2025 will reach 28.4% and 30.5% respectively .

Data center switches continue to be faster, and 400G/800G will dominate in the future. According to the Arista white paper "Embracing 400G and 800G", as next-generation applications and distributed data workloads continue to grow, high-speed computing nodes are becoming increasingly mainstream. These emerging applications and the computing and storage they rely on require faster access to more data and drive the need for larger data center networks with higher bandwidth to maximize scalability and performance while minimizing work completion time.

So far, ordinary computers only require 10G/25G connections, while a limited number of high-end applications often require up to 50G or 100G connections. Currently, standard computing nodes are already able to reach 50G/100G data rates, while high-end computing and storage systems are migrating to higher speeds, such as 200G and 400G. According to Dell'Oro's January 24 forecast, in 2023, the speed of data center switches will be mostly 100G and below, accounting for more than 70%. In the future, 400G and above high-speed switches will gradually replace the market. It is expected that in 2026, the share of 800G will exceed 400G and dominate the market. In 2027, 400G and above will account for more than 60%.

03 .

The share of commercial manufacturers will be the main growth in the future, and there is a large space for domestic substitution

The main growth of the Ethernet switching chip market will come from commercial chips. According to data from Frost & Sullivan, the global Ethernet switching chip market size was 36.8 billion yuan in 2020, and is expected to reach 42.2 billion yuan in 2024 and 43.4 billion yuan in 2025. Ethernet switching chips are divided into commercial and self-use. In 2020, commercial and self-use accounted for 50.0% each. Frost & Sullivan predicts that the annual year-on-year growth rate of the global commercial Ethernet switching chip market size from 2017 to 2025 will not be lower than that of the overall market, which means that the proportion of commercial Ethernet chips will continue to increase.

In the future, the main growth in the size of the Ethernet switching chip market will come from commercial manufacturers. The main reasons are as follows: 1) The technical and financial barriers of Ethernet switching chips make it difficult for some self-use manufacturers to support high R&D investment and high-speed iteration of chips under their own scale, and it is difficult to achieve economic benefits, thus affecting the growth of the self-use market; 2) The future growth of global Ethernet switching chips will mainly come from the data center market, and commercial manufacturers in the data center market started early and gained first-mover advantage; 3) Affected by the industrial chain shocks caused by international trade frictions, self-use manufacturers have lower risk resistance to industrial chain coordination and production capacity shortages than commercial chip manufacturers, thus affecting the growth of self-use chips.

Data centers will become the largest downstream terminal market for commercial Ethernet switching chips in China. From the perspective of application scenarios, in 2020, the market size of Ethernet switching chips for data centers, enterprise networks, operators and industrial use in China's commercial Ethernet switching chip market accounted for 58.5%, 27.3%, 12.7% and 1.6% respectively; it is expected that by 2025, the market size of Ethernet switching chips for data centers, enterprise networks, operators and industrial use in China's commercial Ethernet switching chip market will account for 70.2%, 20.7%, 7.8% and 1.3% respectively, and the market size of Ethernet switching chips for commercial data centers will have an average annual compound growth rate of 18.0% from 2020 to 2025. Data centers will become the main driving force for the growth of China's commercial Ethernet switching chip market in the future.

Switching chips have a significant cost position and the market space is growing steadily. From the perspective of switch cost structure, switching chip costs account for the highest proportion, at 32%. According to IDC data, in 2023, Cisco and Arista will account for 43.7% and 11.1% of the global Ethernet switch market respectively, and a total of 54.8% of the market share. And according to Star Wealth, the gross profit margins of Cisco and Arista are close to 60%. We assume that the gross profit margin of the global Ethernet switch market is 70%.

According to IDC and Dell'Oro data, the global switch chip market will be worth $3.871 billion in 2023, a year-on-year increase of 4.33%. It is expected to reach $4.006 billion in 2024, a year-on-year increase of 3.48%, and $4.571 billion in 2027.

Commercial Ethernet chips are more in line with market development trends. Global Ethernet switch chip self-use manufacturers are mainly Cisco, Huawei, etc., and their self-developed chips are mainly used for self-developed switches rather than for supplying to their competitors. In terms of commercial use, with the expansion of the global Ethernet switch chip market, self-use manufacturers can no longer meet the growing demand of downstream. Therefore, Ethernet switch chip commercial manufacturers such as Broadcom, Marvell, Realtek, NVIDIA, Intel, and Centec Communications have emerged around the world. Some self-use manufacturers have also enriched their switch product lines by purchasing commercial chips from outside.

The Ethernet chip market presents a monopoly pattern, and there is a large space for domestic substitution in the future. According to data from Frost & Sullivan, in China's self-use Ethernet switching chip market in 2020, Huawei and Cisco had market shares of 88.0% and 11.0%, respectively, accounting for a total of 99.0% of the market share. In the commercial market, Broadcom is the leader in commercial Ethernet switching chips, and Marvell and Realtek are the main players in the industry. In China's commercial Ethernet switching chip market in 2020, the market shares of Broadcom, Marvell and Realtek were 61.7%, 20.0% and 16.1%, respectively, accounting for a total of 97.8% of the market share. At this stage, there are few manufacturers in China that have successfully entered the international market competition sequence for commercial Ethernet switching chips, and there is a large space for domestic substitution in the future.

Scarce material numbers are expensive, and domestic substitution is imminent. Since 24Q1, the prices of multiple switching chips in Broadcom's Tomahawk 4 series have risen rapidly. Most of them are out of stock on its official website and other trading platforms, and the delivery time is as long as 50 weeks. Among them, the market price of BCM56990B0KFLGG has reached about US$4,100. Even Broadcom's earlier product Tomahawk series BCM56960 (capacity 3.2Tbps, port rate 25-100G) is generally seriously overpriced on major platforms. It is priced at US$75 on eBay but out of stock, and the price on winsource is US$1416.42. The price on other platforms is almost above US$4,000.

The average sales price of the CTC8186 chip product of the TsingMa.MX series of Centec Communications, which has similar performance and better performance than the CTC8180 chip product (capacity 2.4Tbps, port rate 1-400G), was only 2.252.33 yuan during the trial production stage. If mid-to-high-end products can gradually achieve domestic substitution, domestic manufacturers are expected to achieve higher revenue growth .

04 .

Representative enterprise sorting

1. Broadcom: Global leader in communications chips, driven by both internal growth and external expansion and mergers and acquisitions

Broadcom is a global leader in the communications chip industry. It has been established through multiple mergers and acquisitions. The predecessor of Broadcom can be traced back to the semiconductor products department of HP established in 1961. In 1999, the department was spun off from HP and became part of Agilent Technologies. In 2005, private equity funds KKR and Silver Lake Partners acquired Agilent Technologies' semiconductor division for US$2.66 billion and established Avago Technologies. In 2014, Avago acquired LSI, and in 2016, it "swallowed an elephant with a snake" and acquired Broadcom for US$37 billion. After the integration, the company was named "Broadcom". After that, Broadcom continued to acquire industry-leading companies such as Brocade, CA Technologies, Symantec, and VMware, and gradually grew into an industry leader with a complete industrial chain layout in the semiconductor industry, and laid out infrastructure software business.

Operating income continued to grow, with the semiconductor business as the main growth. Broadcom's operating income in fiscal 2023 was US$35.819 billion, a year-on-year increase of 7.88%. Its main businesses include semiconductor solutions and infrastructure software, which accounted for 79% (a year-on-year increase of 1pct) and 21% (a year-on-year decrease of 1pct) of operating income in fiscal 2023, respectively. Among them, semiconductor solution products are mainly wired and wireless communication chips; infrastructure software mainly includes mainframe software, network security, FC SAN storage, etc. Broadcom products are widely used in broadband, data center and enterprise networks, mobile device connections, servers and storage systems, as well as downstream fields such as industry and automobiles.

2. Centec Communications deeply deploys mid-to-high-end product lines to build product competitive advantages

Closely follow the industry leaders to lay out product lines. The company closely follows the development trend of the world's leading Ethernet switching technology, seizes the industry trend and closely follows the market leaders (such as Broadcom and Marvell) to expand product layout. The company's current main Ethernet switching chip products cover 100Gbps-2.4Tbps switching capacity and 100M-400G port rate, and have been widely used in enterprise networks, operator networks, data center networks and industrial networks. In the data center field, the company has launched TsingMa.MX (switching capacity 2.4Tbps), GoldenGate (switching capacity 1.2Tbps) and other series, and have been introduced to mainstream domestic network equipment vendors and achieved large-scale mass production, but the Arctic series of high-performance switching products that meet the highest international standards and have a maximum switching capacity of 25.6Tbps and are aimed at ultra-large-scale data centers is planned to be launched in 2024.

Continue to invest in R&D to maintain the advanced nature of the company's products. Formula technology is the most important technology for copper clad laminate companies and the biggest technical threshold in the copper clad laminate industry. Formula development is extremely complex and requires not only advanced theoretical support and rich test accumulation, but also a large number of experiments for continuous trial and error and verification. After more than 20 years of continuous R&D and in-depth cultivation, the company has accumulated rich experience in the R&D and production of copper clad laminates, and has kept up with the pace of industry technology upgrades, continuously updated its own technology system, and has formed core technologies that match the development of downstream industries. (Source: BOC Securities)

Due to limited space, to be continued ...

The 3rd Global Digital Trade Expo

Organizer: Zhejiang Provincial People's Government Ministry of Commerce of the People's Republic of China

Organizer: Hangzhou Municipal People's Government, Zhejiang Provincial Department of Commerce, Foreign Trade Development Affairs Bureau of the Ministry of Commerce

2024 Optoelectronic Co-Packaging CPO and Heterogeneous Integration Forward-looking Technology Exhibition and Exchange Conference

Organizer: Hangzhou Convention and Exhibition Group

Organizer: Yimao Automotive Technology (Shanghai) Co., Ltd.

TEL ｜ 18512119620

E-mail｜ icv@enmore.com