AI servers can be disassembled into five main hardware sections: front fan module, drives and front control board, GPU board assembly, CPU motherboard assembly, and power modules, each with specialized PCBs and design considerations.

Core Hardware Sections

1. Front Fan Module The front of an AI server typically houses a fan module with multiple fans (e.g., eight in an 8U NVIDIA DGX A100), designed for high airflow to cool densely packed components .
2. Drives and Front Control Board Below the fan module are storage drives and a front control board. For example, the DGX A100 integrates eight 3.84 TB drives totaling 30 TB, managed by a control board responsible for system oversight and external device signaling .
3. GPU Board Assembly (GPU Board Tray) This is the defining section of an AI server. It contains the GPUs, module boards, and NVSwitch units. Each GPU uses advanced FCBGA substrates (14–16 layers, 70x70 mm to 100x100 mm) for 2.5D/3D packaging, and NVSwitch modules enable high-bandwidth GPU communication . The OCP Accelerator Module (OAM) hosts the GPU chip, acting as the GPU accelerator card itself. This section is the most computationally intensive and valuable in terms of PCB content .
4. CPU Motherboard Assembly (CPU Motherboard Tray) This section includes the CPU motherboard, system memory, network adapters, and PCIe switches. High-speed signal transmission and power distribution are critical here, with thick copper layers and specialized dielectric materials used to reduce signal loss and ensure stable operation .
5. Power Modules AI servers like the DGX A100 include multiple power supply units (e.g., six units) at the rear lower section. Modern AI servers increasingly adopt high-voltage systems (48V or 800V) and resonant circuits for efficient power conversion, requiring large numbers of MLCCs and other passive components .

PCB Design Considerations

• Heat Dissipation: Multi-layer PCBs use embedded copper blocks to conduct heat from chips to edge heat sinks, improving efficiency over surface-only dissipation .
• Signal Integrity: High-frequency boards use specialized dielectric materials to minimize signal loss, ensuring reliable high-speed data transfer .
• Power Distribution: Thick copper foils and carefully designed power layers handle high instantaneous currents from AI chips, maintaining system stability .
• Component Value: GPU substrates and NVSwitch carriers are high-value PCBs, with GPU substrates estimated at around $100 each and NVSwitch carriers around $30 each .

Safety and Best Practices

Disassembling AI servers requires careful handling of sensitive components, including GPUs, memory modules, and high-voltage power supplies. Anti-static precautions, proper tools, and documentation of connections are essential to avoid damage. Understanding the layout and function of each section helps in maintenance, upgrades, or educational teardown exercises. By following this structured approach, one can safely explore the internal architecture of AI servers and gain insight into the specialized PCB designs that enable high-performance AI workloads.