AI Server PCB Hardware Breakdown

Introduction

This article explains the internal PCB composition of an AI server by disassembling the server hardware, so readers can gain a clearer understanding of the PCB types and their relative value within a system. The analysis focuses on representative NVIDIA DGX systems to illustrate the basic architecture of an AI server.

1. Five main hardware sections

Based on a teardown of NVIDIA DGX A100 and consideration of DGX H100, the DGX A100 is used as the primary reference due to greater available detail. The system can be divided into five hardware sections:

Front fan module: The front contains a fan module made of eight fans, consistent with common 8U server arrangements.
Drives and front control board: Below the front fan module are drives and a front control board responsible for control and external device signaling. The DGX A100 is configured with eight 3.84 TB drives, totaling 30 TB internal storage.
GPU board assembly (GPU Board Tray): The rear area houses the critical components of an AI server. The GPU board assembly is the defining section that differentiates AI servers from conventional servers. In the DGX A100 architecture the GPU assembly mainly includes GPU components, module boards, and NVSwitch units, each involving different PCB types.
CPU motherboard assembly (CPU Motherboard Tray): This section contains the CPU motherboard, system memory, network adapters, PCIe switches, and other elements common to servers. The CPU motherboard, system memory modules, and NICs are the primary PCB consumers here.
Power modules: The DGX A100 includes six power supply units located at the rear lower section. Power supplies typically use thick-copper PCBs.

From a functional perspective, PCB value in an AI server can be grouped into three parts: the GPU board assembly, the CPU motherboard assembly, and accessory modules such as fans, drives, and power. The following sections analyze these parts in detail.

2. GPU board assembly: per-system PCB value about RMB 12,250

The GPU board assembly PCBs are composed of four main parts: GPU substrate, NVSwitch carrier, OAM, and UBB.

GPU substrate

The NVIDIA A100 GPUs and DRAM use advanced 2.5D/3D packaging. The supporting substrate is an FCBGA carrier approximately 70×70 mm to 100×100 mm with 14 to 16 layers, one substrate per GPU. For a DGX A100 equipped with eight GPUs, eight substrates are required. Supply-chain research indicates a per-substrate value of about $100, i.e., RMB 650 per piece, yielding a per-system GPU substrate value of RMB 5,200.

NVSwitch carrier

NVSwitch modules implement NVLink-based communication between GPUs. The NVSwitch carrier is similar to a substrate with relatively simple processing requirements but must support high-volume, high-speed data transmission. Industry research estimates about $30 per unit, i.e., RMB 195 per piece. With six NVSwitch units on an A100-based system, the per-system value is RMB 1,170.

OAM (GPU accelerator card)

The OCP Accelerator Module (OAM) is the GPU accelerator card that carries the GPU chip. OAMs correspond one-to-one with GPUs, so a DGX A100 with eight GPUs uses eight OAMs. Using PCIe-based form-factor reference dimensions of 267.7 mm × 111.15 mm, an OAM PCB area is about 0.03 square meters. For the SXM OAM version used in high-end configurations, the PCB typically requires 20 layers, Ultra Low Loss CCL material, and multi-stage HDI, with an estimated price of RMB 12,000 per square meter. The PCIe OAM version has lower specifications, requiring about 14 layers, Ultra Low Loss mixed with high Tg FR4 CCL, and single-stage HDI, at around RMB 7,000 per square meter. Based on DGX A100 configuration, the estimated per-OAM value for a high-end system is RMB 2,880.

UBB (Unit Baseboard)

The Unit Baseboard is the PCB that hosts the entire GPU platform; one UBB per system. Based on chassis dimensions and supply-chain research, the UBB area is estimated at 0.30 square meters and uses a 26-layer through-hole PCB with Ultra Low Loss CCL. The estimated unit price is RMB 10,000 per square meter, giving a per-system UBB value of RMB 3,000.

Summary: The DGX A100 GPU assembly comprises GPU substrates, NVSwitch carriers, GPU accelerator cards, and the GPU module baseboard. Total GPU-related PCB area is about 0.624 square meters, with an estimated per-system PCB value of RMB 12,250. Substrate-level products account for RMB 6,370 (52%), while PCB-level products account for RMB 5,880 (48%).

3. CPU motherboard assembly: per-system PCB value about RMB 2,845

The CPU motherboard assembly includes CPU substrates, the CPU motherboard itself, and functional daughter boards such as system memory modules, NICs, expansion cards, and system drive PCB.

CPU substrate

CPU substrates are similar in specification to GPU substrates. Assuming a per-CPU substrate value of $100 and two CPUs in the DGX configuration, the per-system value is about RMB 1,300.

CPU motherboard

The CPU motherboard hosts the CPU, PCIe switch chips, TPM module, and various functional daughter cards. For the DGX A100 design using 64-core AMD processors and PCIe 4.0, the motherboard typically uses 10 to 12 layers, Low Loss CCL, and through-hole design. Industry research indicates a price of about RMB 3,000 per square meter. With an estimated motherboard area of 0.38 square meters, the per-system value is about RMB 1,140.

Functional daughter boards

Functional daughter boards include system memory cards, NICs, riser/expansion cards, and the system drive PCB. These boards typically use 8 to 10 layers with Mid Loss CCL at about RMB 1,500 per square meter. Using DGX A100 component counts and estimated areas:

CPU memory cards: DGX A100 has 32 CPU memory modules totaling 2 TB RAM. A typical server memory card area is estimated at 0.004 square meters per module.
NICs: DGX A100 uses Mellanox ConnectX series NICs (options include X-7 and X-6). The standard configuration has 10 NICs (eight single-port 200 Gb/s InfiniBand and two dual-port 200 Gb/s Ethernet). Using disclosed ConnectX-7 dimensions of 68.90 mm × 167.65 mm, each NIC PCB area is about 0.012 square meters.
Riser/expansion card: One riser card is used to accommodate certain board placements; estimated area is 0.01 square meters.
System drive PCB: The DGX A100 uses two 1.92 TB M.2 NVMe system drives mounted on both sides of a single PCB, so one system drive PCB is used with an estimated area of 0.01 square meters.

Combined, the functional daughter boards total about 0.27 square meters, corresponding to an estimated per-system value of RMB 405.

Summary: The DGX A100 CPU motherboard assembly PCBs total about 0.662 square meters, with an estimated per-system value of RMB 2,845. Substrate-level products account for 46%, the main motherboard accounts for 40%, and functional daughter boards account for 14%.

4. Other accessories: per-system PCB value about RMB 226

Other accessories include power supplies, drives, and the front control board. These boards typically use 6 to 10 layers with FR4/Mid Loss CCL specifications and unit prices around RMB 1,000 to RMB 1,500 per square meter. Using DGX A100 component counts and areas:

Power supplies: DGX A100 uses six power supply units. Using a reference 2,200 W server PSU with PCB dimensions 73.5 mm × 265.0 mm, a single PSU PCB area is estimated at 0.019 square meters.
Hard drives: DGX A100 uses eight drives. For a 3.5" drive, the PCB area per drive is estimated at 0.008 square meters.
Front control board: A single PCB located between the eight drives, estimated area 0.01 square meters.

Total PCB usage and value

Combining the GPU assembly, CPU motherboard assembly, and accessories, the estimated total PCB area for a DGX A100 system is 1.474 square meters, with an estimated per-system PCB value of RMB 15,321. The GPU assembly contributes about RMB 12,250 (80%), the CPU motherboard assembly about RMB 2,845 (19%), and other accessories about RMB 226 (1%). By board classification, substrate-level products represent RMB 7,670 (50.1%), and PCB-level products represent RMB 7,651 (49.9%).