Production process of AI server assembly line

The AI server assembly production line is not a single standard process, but a "dual-track system" that integrates electronic assembly and liquid cooling system manufacturing. The core links include: component incoming inspection, motherboard subsystem assembly (CPU/GPU/memory/storage), liquid cooling module integration (if applicable), complete machine assembly, firmware programming, multiple rounds of automated testing (power-on, pressure, thermal, liquid tightness), and factory calibration.

- Incoming materials and pre-inspection: GPU (such as NVIDIA H100/B100), CPU, DDR5 memory, NVMe SSD, power supply (typically 2000W+ with titanium redundancy), motherboard (high-layer PCB), chassis, and liquid cooling components (cold plate, connectors, CDU) are inspected and accepted for storage according to the BOM. Strict electrostatic protection and moisture control measures are implemented.

- Assembly of mainboard subsystem: automatic placement/manual insertion and extraction (depending on the level of customization) → CPU/GPU installation (aligning notches, tightening diagonally, precise application of TIM thermal interface material) → memory/SSD insertion → RAID/NIC/expansion card fixation → cold plate base and GPU/CPU fitting (torque control ±0.1N·m, flatness ≤30μm).

- Liquid cooling module integration (if configured): Assembly of split cold plate and heat exchanger core (quick connector blind insertion force ≤35N, residual liquid <0.05mL) → Initial inspection of pipeline connection and leak tightness (air pressure maintained at 0.8MPa for 60s, pressure drop <0.5%) → This step is skipped for non-liquid cooled models.

- Complete machine assembly and wiring: placing the motherboard into the chassis (positioning with copper pillars, locking diagonally with 8-12 screws), power supply connection (24ATX + 8CPU + multiple PCIe), arranging internal cables in a "horizontal and vertical straight" manner (to avoid blocking the air duct), and connecting the IPMI/BMC module.

Firmware and BIOS programming: Writing BMC firmware, BIOS configuration (RAID/PCIe lane allocation/power consumption limit), serial number, and security key injection through automated testing stations.

- Power-on and functional testing: Power-on inspection (voltage/fan/LED), POST self-test, memory stress test (MemTest86), NVLink bandwidth verification (>900GB/s/link), PCIe Gen5 link training.

- Thermal and fluid tightness testing (key differences): Run at full load (GPU/CPU 100%) for 48–72 hours. For liquid-cooled models, measure thermal resistance (<0.08°C/W@1L/min), flow resistance, and leakage (helium mass spectrometer leak detection <1×10⁻⁹ atm·mL/s) in a 45°C ambient water tank. For air-cooled models, measure the temperature rise in the air duct and noise.

- System pre-installation and calibration: Deploy OS (such as Ubuntu Server), drivers, monitoring agents (Prometheus), preset AI inference frameworks (such as Triton) → Generate unique "digital twin" traceability codes (binding all component SNs, test data, and cold plate soldering temperature curves).

- Packaging and Shipping: Shockproof packaging (especially for GPUs and liquid cooling connectors), anti-static bags, custom labels (such as rack positions, network topology diagrams), and high-value full containers (such as NVL72) require calibration of the cold plate mating surface before being sealed and transported in their entirety.

Currently, the leading ODMs (Foxconn, Wistron, Quanta) have achieved 27-35 standard processes. Liquid cooling integration has upgraded the traditional server production line to a composite production line encompassing "electronics + fluid + thermal control". The L1-L10 level NVIDIA certification system (from Die to full container) dominates the manufacturing of high-end AI servers. The production line generally adopts anti-static flexible conveyor lines + robotic arm coordination (motherboard/GPU installation) + AOI visual inspection + fully automated aging warehouses. Each server's serial number is traceable throughout the entire process, and a CPK of ≥1.67 is the threshold for mass production.