In the AI era, the demand for computing power is surging, and high speed optical modules are moving

Mondo Technology Updated on 2024-02-27

2.Overview of the optical transceiver industry

From 2022 to 2028, the global optical module market size CAGR will reach 12%, and the global scale is expected to exceed $20 billion by 2028. According to Yole Intelligence, the global optical module market size was valued at $11 billion in 2022 and is expected to grow to $22.3 billion in 2028, growing at a compound annualized growth rate of 12% from 2022 to 2028, driven by the high demand for 800G high data rate modules from large cloud service operators and the requirements of national telecom to increase fiber network capacity.

The domestic market demand for optical modules is 2 billion to 3 billion US dollars, accounting for about 1 4 in the world. According to lightcounting**China's optical module deployment accounted for 25%-35% of the world's total from 2018 to 2023, and 20%-25% from 2024 to 2029, a slight decline. North American cloud service providers have aggressive plans to deploy 800G in AI clusters, which will be a major factor in China's declining share over the next 2-3 years.

2.1.The capital expenditure of operators in the telecom market is stable and long, and the proportion of expenditure is tilted towards computing power networks

In terms of the telecom market, since the announcement of commercial 5G services in 2019, the capital expenditure of the three major telecom operators has shown a steady and growing trend, with a total capital expenditure of about 299.9 billion yuan in 2019 and is expected to be about 359.1 billion yuan by 2023, with an annualized compound growth rate of 461%。2019 was the first year of 5G network construction, followed by the peak of 5G development from 2020 to 2022.

On June 6, 2019, the Ministry of Industry and Information Technology issued 5G commercial licenses to the three major operators and China Radio and Television, and the first batch of 150,000 5G base stations was built in the same year, with an average of 67 new 5G base stations per year from 2019 to 2023540,000 units. In the field of 5G, optical modules are used in fronthaul, midhaul and backhaul, of which fronthaul mainly corresponds to 25G 50G optical modules, midhaul mainly corresponds to 50G 100G optical modules, and backhaul mainly corresponds to 100G 200G 400G optical modules.

The three major telecom operators are gradually increasing their investment in industrial Internet and computing power networks. In 2023, China Mobile expects capital expenditure to reach 183.2 billion yuan, of which 45.2 billion yuan will be spent on computing power networks, a year-on-year increase of nearly 35%, with more than 240,000 new cloud servers and more than 40,000 newly available IDC racks. China Telecom expects capital expenditure to reach 99 billion yuan in 2023, and industrial digitalization investment is expected to increase by 40%, including 19.5 billion yuan in computing power and 9.5 billion yuan in IDC. China Unicom's computing power network accounted for 16 percent of capital expenditure in 20228%, the company expects that the capital expenditure of the computing power network will reach 14.9 billion yuan in 2023, accounting for 19 percent of the total capital expenditure4%, a year-on-year increase of more than 20%.

2.2.The high capital expenditure and the upgrade of the network architecture have become the core driving forces to enhance the prosperity of the data communication market

The four largest cloud service providers in North America account for about 85% of the global capital expenditure, and the CAPEC annualized compound growth rate of CAPEX from 2019 to 2022 is as high as 3077%。According to Counterpoint Research**, the top four cloud service providers in North America, Microsoft, Amazon, Google, and Meta, will account for 85% of total global capital expenditure in 2023. Among them, the overall capital expenditure slowed down in 2022, which may be attributed to the global economic downturn caused by the epidemic, but it has gradually recovered since 23Q2.

The expansion of the global data scale has accelerated the investment in the construction of data centers, and the internal equipment of data centers needs to be interconnected by the network, which corresponds to the large demand for optical modules. With the continuous development of communication and Internet applications, the demand for computing power and data storage of individuals and enterprises has gradually migrated to the "cloud". According to the latest data from IDC, the global data size will increase from 103. in 202266ZB, up to 284 in 20273ZB, the annualized compound growth rate from 2022 to 2027 will reach 2236%, we believe that the increase in data scale requires more servers and switches, and these devices need information exchange and network interconnection, which corresponds to a large demand for optical modules.

Statista estimates that the number of hyperscales worldwide has increased from 259 in 2015 to 700 in 2021, and the latest data from Synergy Research Group shows that the number will be close to 900 in 2023, accounting for 37% of all data centers worldwide. According to Precedence Research, the global hyperscale market size is 801600 million US dollars, which is expected to increase to 935.3 billion US dollars in 2032, with a compound annualized growth rate of 279%。

The leaf-spine network architecture greatly increases the demand for the number of optical modules, helping optical modules to be iteratively upgraded to higher speeds. There are two types of traffic in the data center: 1) North-south traffic: the traffic that communicates between clients outside the data center and servers in the data center; 2) East-West Traffic: The traffic of communication between different servers in a data center. Cisco's 2021 data shows that east-west traffic is the dominant source of data center traffic, accounting for 85% of the traffic.

East-west traffic in a traditional Layer 3 topology network architecture must pass through devices at the aggregation and core layers, passing through many non-essential nodes, resulting in slower response times for end users. In contrast, the flat design of the spine-leaf architecture can distribute the east-west traffic across multiple paths, which is more suitable for use in scenarios with large east-west traffic. Furthermore, the interconnection between the spine switch and the leaf switch needs to be matched with a reasonable bandwidth ratio, and the uplink should always run faster than the downlink to avoid port link blockage.

2.3.AIGC drives the demand for computing power, and the "barrel effect" accelerates the iterative upgrade of high-speed optical modules

On November 30, 2022, the release of ChatGPT, a model owned by OpenAI, kicked off a global arms race. On November 30, 2022, OpenAI released ChatGPT, a large language model, and CEO Sam Altman announced that ChatGPT users had exceeded 1 million in just 5 days, and by the end of January 2023, the number of ChatGPT users exceeded 100 million, making it the fastest-growing consumer app in history. The rise of ChatGPT has led many AI companies around the world to release their own models one after another, with overseas Google launching the Palm2 model, Meta releasing Llama-13B, and Microsoft building New Bing based on ChatGPT; It was the first in China to release Wenxin Yiyan.

Taking ChatGPT as an example, behind the model upgrade and iteration is the training and inference of massive data, and the global demand for computing power has ushered in rapid growth since 2012. GPT-1 has 1 model parameter and 1 pre-trained data1.7 billion and 5GB, to GPT-3 has surged to 175 billion and 45TB, and it is expected that in the future, GTP-5 will have 100 times the number of parameters and 200 400 times the amount of computation required for GTP-3. According to the "AI and Compute" analysis report released by OpenAI, since 2012, the computing power demand of AI training applications has doubled every 3-4 months, and since 2012, AI computing power has increased by more than 300,000 times.

There is a "barrel effect" in the application and development of AI, and computing power clusters need to be matched synchronously with network interconnection, which accelerates the development of optical modules and optical chips in the direction of high speed. The core of computing power is AI chips, in which GPUs are good at parallel and large number of repeated calculations, which are widely used in the AI field.

Common AI chips include three types: general-purpose (GPU, DPU), semi-customized (FPGA) and fully customized (ASIC), among which GPU commercialization is more mature; However, there are many types of AI chips developed with FPGA and ASIC architectures, which are still in the development and exploration stage. According to IDC data, GPUs were the most widely used in China's AI chip market in 2022, accounting for 89% of the market share and reaching $6 billion. The data exchange between computing server clusters requires the network interconnection within the data center, which stimulates a large demand for high-speed optical modules. The huge training task of the model requires a computing power cluster composed of a large number of GPU servers to provide computing power, and these servers need to exchange massive amounts of data through network connections, which also requires a higher rate of optical modules to match.

In March 2022, NVIDIA launched the DGX H100, and the demand for 400G and 800G optical modules has increased. Due to the different networking methods, it is difficult to accurately calculate the relationship between GPU and optical module usage, so we take the NVIDIA DGX H100 server cluster as an example: 1) In the computing network, each single H100 corresponds to 15 x 800G transceivers and 09 x 400G transceivers; 2) In the storage network, each monolithic h100 corresponds to 125 x 400G transceivers. Assuming that only the requirements for optical modules in the above two network scenarios are considered, each single h100 corresponds to 15 x 800G transceivers and 2Fifteen 400G optical modules are required, and each H100 corresponds to about 20 100G optical chips.

3.The innovation of packaging and materials has become the future development direction of optical transceivers

In the era of computing power, data centers have become major energy consumers, and the upgrade of optical module technology is not only a simple doubling of speed, but also needs to solve the power consumption and cost problems caused by high speed. In 2021, the power consumption of China's data centers was 216.6 billion kWh, which is about 1,036 of the annual power generation of the Three Gorges Power Station in the same period2 times 4.9 billion kWh; In 2022, the power consumption of data centers in China will reach 270 billion kWh, accounting for about 3% of the electricity consumption of the whole society. According to China Energy News, the proportion is expected to be close to 5% in 2025.

The energy consumption of optical modules accounts for 40%-50% of the energy consumption of the data center switching network. According to FibalMall data, the energy consumption of 400G optical modules in data center applications is 10-12W, and the energy consumption of 800G is 15-18W, and the future is 16T will consume 2 times more than 400G and is expected to be as high as 20-24W; At the same time, Cisco's data shows that from 2010 to 2022, the energy consumption of the optical engine increased by about 26 times. Obviously, the surge in the energy consumption of optical modules has brought great pressure to the cost side of data centers, and solving their energy consumption problems has become the key to the current update of optical module technology.

3.1. lpo

LPO (Linear-Drive Pluggable Optics) uses linear drive technology to replace the traditional DSP (Digital Signal Processing) CDR (Clock Data Recovery) chip, which can reduce power consumption and reduce costs, but at the cost of removing the DSP, it will lead to an increase in the system bit error rate and a shortening of the communication distance, so LPO technology is only suitable for short-distance application scenarios, such as data center cabinet-to-switch connection.

Traditional DSP can fix the distortion problem of high-speed signal after conversion between optical-electrical and electrical-optical, thereby reducing the impact of distortion on the system bit error rate, but the power consumption is high and the cost is high: 1) In the 400G optical module, the power consumption of the 7nm DSP is about 4W, accounting for 50% of the power consumption of the entire module; 2) In 400G optical modules, the cost of DSP BOM accounts for about 20%-40%. LPO technology removes the DSP and integrates its related functions into the switch chip on the device side, leaving only the driver (driver chip) and TIA (trans-impedance amplifier) with high linearity for a certain degree of compensation for high-speed signals.

3.2. cpo

Co-packaged optics (CPO) refers to the co-packaging of network switching chips and optical modules on the same slot to form a co-packaging of chips and modules. Compared with traditional hot-swappable technology, the advantages of CPO technology include: 1) Low latency and low power consumption: Since the optical module and switch chip are in the same package, the signal transmission path is shorter, which can achieve lower latency. In addition, the photoelectric co-packaging technology can reduce the power consumption of signal transmission and improve the energy efficiency of the overall system. 2) High bandwidth: Optoelectronic co-packaging technology supports high-speed optical communication, which can provide greater data transmission bandwidth. 3) Small size: Compared with the traditional way of separating optical modules and electronic chips, optoelectronic co-packaging technology can achieve a more compact size, which is conducive to the application of high-density integrated circuits.

3.3.Silicon photonics technology

Silicon photonics technology is a new technology for the development and integration of optical devices based on silicon and silicon-based substrate materials, using existing CMOS processes. The core concept of silicon photonics technology is "light instead of electricity", that is, the use of laser beams instead of electronic signals to transmit data, the integration of optics and electronic components into a separate microchip, and the speed of connection between chips.

Compared with traditional discrete device solutions, the advantages of silicon photonics technology include: 1) High integration: It uses a semiconductor manufacturing process to integrate silicon photonic materials and devices on the same silicon-based substrate to form an integrated photonic device composed of optical modulators, detectors, passive waveguide devices, etc. Compared with traditional discrete devices made of active materials such as indium phosphide (INP), silicon photonics modules do not need ROSA (optical receiving components) and TOSA (optical transmitting components) packaging, so the size and number of silicon photonics devices are smaller and more integrated. 2) Low cost: Compared with traditional discrete devices, silicon photonic modules have a higher degree of integration, and lower packaging and labor costs; In addition, the low cost of silicon-based materials and the ability to manufacture large sizes mean that the cost of silicon-based chips can be significantly reduced. 3) Compatible with mature CMOS processes: Silicon photonics technology can take advantage of the mature processes of semiconductors in ultra-large-scale, micro-manufacturing and integration.

From 2022 to 2028, the annualized compound growth rate of silicon-based optoelectronic chips is expected to reach 44%. According to Yole data, the size of silicon-based optoelectronic chips in 2022 was about $68 million, and the market size is expected to grow to more than $600 million in 2028, with an annualized compound growth rate of 44% from 2022 to 2028.

4.Focus on the target

1) Optical modules and upstream optical devices: Zhongji Innolight, Tianfu Communication, Xinyisheng; 2) Optical chips and silicon photonics: Yuanjie Technology (jointly covered with the electronics team), Borche Technology, Shijia Photonics; 3) Switches: ZTE, Unisplendour (jointly covered with the computer team), Hudian (jointly covered with the electronics team); 4) Computing power scheduling: Runjian Co., Ltd., Runze Technology (jointly covered with the mechanical team).

This is an abridged excerpt from the report, the original PDF of the report

In-depth research on information technology-communication industry: the demand for computing power in the AI era is surging, and high-speed optical modules are traveling with the tide-Tianfeng**[Tang Haiqing, Wang Yihong]-20240224[Page 25]".

Report**: Value Catalog

Related Pages