AI Server PCB

AI server PCBs serve as the fundamental electronic platform, connecting and supporting the processors, memory, accelerators, and power management systems required for artificial intelligence computational workloads. Within AI servers, PCB design addresses high-speed signal transmission, high power density, and thermal management to meet the demands of GPU/TPU clusters and large-scale data processing.

AI servers primarily involve three key PCB product categories: GPU substrates commonly employ high-layer boards exceeding 20 layers; compact AI accelerator modules predominantly rely on 4–5-layer HDI for high-density interconnections; while traditional CPU motherboards form the foundational support structure. As AI servers undergo iterative upgrades, GPU motherboards are progressively transitioning towards HDI structures. This trend positions advanced HDI as the fastest-growing segment within the AI server PCB market over the next five years, with particularly urgent demand for 4th-layer and higher products.

Within servers, PCBs are primarily deployed in components such as accelerator boards, motherboards, power supply backplanes, hard drive backplanes, network interface cards, and riser cards. Their core characteristics manifest as high layer counts, high aspect ratios, high density, and high transmission rates. As server platforms undergo continuous iteration and upgrades, PCB layer counts continually increase, imposing higher demands on materials, design, and manufacturing processes.

Three supply relationships govern AI server PCBs:

① GPU board assemblies are entirely designed by GPU manufacturers, who consequently dictate PCB supply arrangements.

② CPU board assemblies adhere to established server manufacturer supply chain relationships. CPU carrier boards are determined by the CPU designer, while CPU templates and expansion card boards for the complete system are specified by the end customer. For most other chip-equipped PCBs, the customer presents design requirements to functional component manufacturers, who then independently determine PCB procurement.

③ Accessories: Accessories are typically purchased directly by customers from established module manufacturers. In some scenarios, customers may propose specific design requirements to accessory module suppliers, though this does not affect the module supplier's autonomy in PCB procurement decisions.

In AI server power supply upgrades, PCBs undergo enhancements in materials, processes, and other aspects. As the carrier for electronic components, PCBs within server power supplies are utilised in modules such as power switches, power filters, voltage regulators, and heat sinks. Compared to general-purpose servers, PCBs in AI server power supplies feature upgrades in materials, manufacturing techniques, and technology.

1) Increased copper thickness to accommodate higher currents: PCB interconnections rely on pure copper traces on the substrate. Thicker copper foil enables higher current carrying capacity. Copper foil accounts for approximately 9% of PCB upstream raw materials, while copper-clad laminate materials constitute over 30%. Copper foil is also a primary raw material for copper-clad laminates, representing about 40% of their cost. Increasing copper thickness simultaneously imposes higher demands on processes such as laminating prepreg between layers, drilling, and electroplating, significantly elevating the value of PCB products.

2) Embedding power modules to enhance power density: PCB embedded power module technology holds immense performance potential. Compared to traditional packaging methods, PCB-embedded power modules can increase the current-carrying capacity per semiconductor by approximately 40%, or reduce semiconductor usage by one-third for equivalent current output. Under identical power output conditions, the material cost of power modules is expected to decrease by 20%. The overall switching losses of the inverter are reduced to one-third of conventional inverter products. Consequently, the increase in switching losses resulting from higher switching frequencies is reduced by two-thirds compared to traditional inverters.

3) Thermal management employs materials with superior thermal conductivity: Higher thermal conductivity in PCB substrate materials enhances heat dissipation. Generally, resins exhibit poor thermal conductivity, whereas copper foil traces and vias act as excellent thermal conductors. Consequently, key thermal management strategies include increasing copper residue rates, boosting the number of thermal vias and enhancing copper thickness within them, and embedding copper blocks or ceramic plates. Concurrently, rational routing design prevents hotspot concentration on the PCB.

Key obstacles facing AI server PCBs:

Challenges in High-Speed Signal Integrity: AI servers require high-speed interconnect capabilities (such as PCIe 5.0/6.0, CXL, HBM interfaces, etc.). During high-speed differential signal transmission on PCBs, issues like crosstalk, reflections, delays, and losses readily emerge. Moreover, as PCB layer counts increase and routing density intensifies, maintaining signal integrity and ensuring low latency becomes progressively more challenging.

Exorbitant Material and Manufacturing Costs: AI server PCBs typically require high-performance substrates, such as high-speed materials with low dielectric constant (Dk) and low loss factor (Df), or even hybrid materials. They also demand ultra-high layer counts (20 layers and above) and employ precision HDI and blind/buried via processes. This not only substantially increases manufacturing costs but also makes yield rates difficult to guarantee, thereby limiting cost-effectiveness during large-scale deployment.

Opportunities for Industry Development:

Opportunities driven by high-speed interconnect demand: To achieve high-speed computing and large-bandwidth data transmission, AI servers impose heightened requirements on PCBs. These include multi-layer designs, superior high-speed signal integrity, and the use of low-loss materials such as high-frequency, high-speed substrates featuring ultra-low dielectric constant (Dk) and low loss factor (Df). This presents growth opportunities for advanced PCBs such as HDI, Substrate-Like Package (SLP), modified semi-additive process (mSAP), and arbitrary-layer interconnect PCBs.

Despite facing signal integrity and cost challenges, the development of AI server PCBs is being propelled by high-speed interconnect demands and power supply technology upgrades. This is accelerating the penetration of advanced products, with prospects for expanding market share by balancing performance and cost in the future.