Stuck on Nvidia and Microsoft's necks? What is the charm of electromagnetic shie

Electromagnetic shielding is not a new technology in the PC hardware field, but why has it recently been highlighted by industry giants such as NVIDIA and Microsoft?

01

"Electromagnetic Shielding" Choking the Throat of NVIDIA's GB200

Not long ago, the Blackwell architecture and GB200 made their debut at NVIDIA's GTC conference, quickly sparking a "copper cable high-speed connection craze" in the secondary market. Compared to optical communication, copper cable high-speed connections have the advantages of low cost, low power consumption, easy wiring, and maintenance, but they also have a weak point — copper cables are highly susceptible to electromagnetic interference.

Compared to traditional optical communication, copper cable high-speed connections have the advantages of low cost, low power consumption, easy wiring, and maintenance, but they also have a weak point — copper cables are highly susceptible to electromagnetic interference. AI servers based on copper cable high-speed connections therefore have a demand for electromagnetic shielding.

At NVIDIA's latest quarterly earnings call, CEO Jen-Hsun Huang predicted that the Blackwell architecture chips would generate significant revenue this year. It is reported that the intended customers for the Blackwell architecture chips include Amazon, Google, Meta, Microsoft, OpenAI, Oracle, Tesla, and others.

Advertisement

The market demand and progress of GB200 have far exceeded expectations, reigniting the market's fervent anticipation that copper cable high-speed connections will lead AI servers.

In addition to NVIDIA, Microsoft also recently announced its first Copilot+PC, which is their latest Surface series, to be launched on June 18th. This laptop will be equipped with Qualcomm's Arm architecture processors, including the Snapdragon X Elite and Snapdragon X Plus, two major self-developed processors recently released by Qualcomm. Versions with Intel and AMD chips will be released later.From an external perspective, terminal AI hardware requires more powerful computational support to run AI-related functions, which not only poses higher demands on chips such as CPUs, GPUs, and NPUs but also drives the demand for storage, cooling, and electromagnetic shielding solutions. AIPCs and AI smartphones are becoming development trends, which will set higher standards for cooling and electromagnetic shielding solutions.

According to a forecast by BCC Research, the global market size for electromagnetic shielding materials will reach $9.25 billion in 2023. Market research firm QYResearch estimates that the global market size for electromagnetic shielding films will reach $230 million by 2029, with a compound annual growth rate of 5.3% over the next few years.

02

Copper Interconnect Maximum Expectation Gap

NVIDIA's GB200 demand and progress have exceeded expectations, with the core increment lying in copper interconnects, which means using high-frequency, high-speed copper cables for data communication between GPUs within the rack to facilitate large model training and inference. The extensive use of high-speed copper wires inside servers is a paradigm shift in intra-machine connectivity promoted by NVIDIA, and it is currently in the adoption phase.

Electromagnetic shielding materials refer to functional materials that can shield electromagnetic waves, and their working principle is to block or attenuate electromagnetic waves by reflecting and absorbing them. Currently, its downstream industries cover terminal application areas such as communication equipment, computers, mobile terminals, automotive electronics, household appliances, and national defense and military.

In the fierce competition of AI servers, if electromagnetic interference cannot be effectively resisted, it will lead to shortened product lifespan, additional costs, and even affect research and development progress and corporate competitiveness. Therefore, the application of electromagnetic shielding materials in copper interconnect AI servers has become crucial.

Copper interconnects have the advantages of low cost, low power consumption, and ease of wiring and maintenance compared to optical communication, but they also have a rather fatal weakness: copper cables are highly susceptible to electromagnetic interference (while optical fibers are hardly affected). AI integrated server systems represented by NVIDIA's GB200 are highly complex electrical systems, but they will face the threat of electromagnetic interference in the future. The main source of electromagnetic interference is the server itself. To enhance training efficiency, current AI servers generally consume a large amount of power, using multiple high-power power supplies. The DGX H100 is equipped with six 3300W high-power power supplies, and the GB200 is expected to have a design power consumption of 120,000 watts. The high-power power supplies inside will inevitably generate strong electromagnetic interference, and the copper wires inside the GB200 will struggle to resist.

Currently, major manufacturers are in a tense AI arms race and large model development competition. If AI servers cannot effectively resist electromagnetic interference, it will not only shorten their lifespan and cause additional costs but also delay research and development progress, affecting corporate competition.Therefore, with the wave of copper interconnects being introduced into AI servers, electromagnetic shielding materials are the core expected difference. Past market discussions on electromagnetic shielding materials have mostly focused on the demand pull from AIPC, and the pull from AI servers on electromagnetic shielding materials will bring a huge increment, which is the core expected difference in the GB200 track that is born alongside copper interconnects.

03

The contention between copper interconnects and optical modules

Similarly applied to AI servers and data transmission, the rise of copper interconnects naturally leads people to think of the rise of optical modules over the years. What is the connection between the two?

First, we must start with the acceleration of data center network architectures towards high-speed bandwidth iteration. Large models and applications have high requirements for computational infrastructure, requiring stable, efficient, and secure digital infrastructure to support the entire interaction process of generation, storage, and transmission, which brings an unprecedented demand for computing power. According to data from the China Intelligent Computing Industry Alliance, driven by AIGC, the demand for AI computing power is growing exponentially, roughly doubling every 1-2 months.

Interconnects between data center devices have become a bottleneck in improving the overall system computing power of data centers, and it is imperative to increase the bandwidth rate of communication networks. Taking the generations of PCIe protocols as an example, PCIe1.0 was released in 2003, with a single-channel bidirectional bandwidth of 500MB/s and a transfer rate of 2.5 GT/s; the latest PCIe 6.0 specification was released in 2021, with the transfer rate increased to 64 GT/s. According to the technical roadmap released by PCI-SIG, the PCIe 7.0 expected to be released in 2025 will have a maximum data rate of 128.0 GT/s, providing up to 512GB/s of bidirectional transfer bandwidth through 16 channels.

Copper connection products have always played an important role in high-speed interconnects in data centers, especially in short-distance transmission scenarios inside servers and Server-Top of Rack. Copper connections have significant advantages in terms of heat dissipation efficiency, signal transmission, and cost. Therefore, copper interconnects remain the most cost-effective solution in many applications today and in the future.According to Light Counting, the market size for global passive direct-attach cable DAC and active cable AEC is expected to grow at a compound annual growth rate of 25% and 45%, respectively. Globally, large artificial intelligence clusters and long-distance model transmission mainly rely on optical communication, while smaller data center systems will continue to primarily use cables.

Optical module technology is primarily applied in the high-speed transmission field of data centers, especially in scenarios requiring high data transfer rates and low latency, such as the training and application of AI large models. There are also significant differences in cost and efficiency between the two.

Copper interconnect technology has a significant cost advantage. For instance, NVIDIA's GB200 server can achieve substantial cost savings by replacing optical cables with copper cables, with a cost reduction of up to six times. This makes copper interconnect an important short-term solution for cost reduction, while although optical module technology may be more expensive in some cases, it can provide higher bandwidth and better reliability in the long term. For example, as AI large models become more widespread, the demand for computing power and network equipment continues to grow, which drives the upgrade and development of optical module technology.

In fact, Wang Jun, Vice President and Secretary of the Board of Directors of InnoLight Technology, responded to the copper cable replacing optical modules at the 2023 annual online performance briefing, stating that copper connections and optical connections have their own distinct application scenarios. Copper connections are mainly used for short-distance connections between chips, within server racks, and between racks, while optical modules are primarily used for interconnections between switches and between top-of-rack switches and server network cards.

Optical modules can continue to iterate with bandwidth demands and may further penetrate into the application scenarios originally occupied by copper interconnects through methods such as silicon photonics, which is also the trend and direction for the development of future AI data centers.