How Many PCBs Does an AI Server Use? Detailed Breakdown of NVIDIA DGX A100

The report analyzes the PCB composition of an AI server, using NVIDIA’s DGX A100 as a case study to illustrate the internal hardware structure and the associated PCB area and cost.

Hardware Sections : The server is divided into five main hardware groups—GPU board, CPU motherboard, fan modules, storage, and power modules. The GPU board is the most valuable, followed by the CPU motherboard, with the remaining components contributing a smaller share.

GPU Board Group (value ¥12,250, 80% of total): consists of GPU carrier boards, NVSwitch, OAM (GPU accelerator cards), and UBB (unit baseboard). The GPU carrier boards use 70‑100 mm FCBGA substrates (14‑16 layers) with a unit cost of about $100 (≈¥650) each, totaling ¥5,200 for eight GPUs. NVSwitch boards cost $30 (≈¥195) each, totaling ¥1,170. OAM cards use 20‑layer ultra‑low‑loss CCL material, valued at ¥2,880 per server. UBB boards require 26‑layer PCBs, valued at ¥3,000.

CPU Motherboard Group (value ¥2,845, 19% of total): includes CPU carrier boards, the main motherboard, and auxiliary boards (memory, NIC, expansion cards, storage drivers). Two CPU carrier boards cost about $100 each (≈¥1,300). The main motherboard uses 10‑12‑layer low‑loss CCL material, covering 0.38 m², valued at ¥1,140. Auxiliary boards (memory cards, NICs, expansion cards, storage driver boards) together occupy 0.27 m² and are valued at ¥405.

Other Components (value ¥226, 1% of total): includes six power supplies, eight 3.5" hard drives, and a front console board. Power supplies each use ~0.019 m² of PCB; hard drives each use ~0.008 m²; the console board uses ~0.010 m².

Total PCB Consumption : Summing all sections, the DGX A100 uses 1.474 m² of PCB material, with a total PCB‑related value of ¥15,321 per server. The PCB cost is split almost evenly between carrier‑level boards (≈¥7,670, 50.1%) and standard PCB‑level boards (≈¥7,651, 49.9%).

This detailed PCB accounting provides insight into the material and cost structure of high‑performance AI servers, helping stakeholders evaluate upgrade value and supply‑chain considerations.