1. The five main hardware components can be summarized as follows: GPU motherboard, CPU motherboard, and accessories.
For our AI server analysis, we focused on two benchmark products : the NVIDIA DGX A100 and DGX H100. Given the H100's relatively recent release and limited availability of detailed information , we began by examining the DGX A100 to understand the basic architecture of a competitive AI server. The NVIDIA DGX A100 resembles a typical home desktop computer. Through in-depth component analysis, we believe the DGX A100 can be broadly divided into five hardware modules:
1) Fan module: Starting from the front, the first thing you see is the fan module panel. The DGX A100's fan module consists of 8 fans, which is basically the same as the traditional 8U server specification.
2) Hard drive: The hard drive and front control panel (for controlling signal transmission to external devices) are located below the front fan module. The DGX A100 is equipped with eight 3.84TB hard drives, for a total internal storage of 30TB.
3) GPU Board Tray: The rear section is the assembly area for key components of the entire AI server. The most crucial component is the GPU board tray, which is the key difference between AI servers and ordinary servers. From the architecture of the DGXA100, the GPU board tray mainly consists of three parts: GPU components, module boards, and NVSwitch. These three parts involve different types of PCB products.
4) CPU Motherboard Tray: This part is the core component of all servers (including ordinary servers and AI servers), which includes the CPU motherboard, system memory, network card, PCIe switch, etc. The CPU motherboard, system memory, and network card are the parts that mainly involve the amount of PCB usage.
5) Power supply module: The DGX A100 is equipped with 6 power supplies at the bottom rear, and the power supply involves the use of thick copper PCB boards.
From a functional perspective, we believe the PCB value of an AI server can be categorized into three parts: firstly, the core GPU board assembly; secondly, the CPU motherboard assembly, essential for all servers; and finally, the component assembly including fans, hard drives, and power supplies. This article will break down each of these three parts in detail.
2. GPU board assembly: Value per unit: 12,000, with carrier board accounting for 52% and PCB board accounting for 48%.
The PCB of a GPU board mainly consists of four parts: GPU carrier board, NVSwitch, OAM, and UBB.
1) GPU Carrier Board: The GPU and DRAM of the NVIDIA A100 use advanced 2.5/3D packaging technology. The carrier board used is a 70*70mm~100*100mm, 14~16 layer FCBGA carrier board. The number of carrier boards corresponds one-to-one with the number of GPUs. According to the DGX A100, which has 8 GPUs, one AI server requires 8 GPU carrier boards. According to industry chain research, the value of a single GPU is about US$100, or RMB 650 per GPU. Therefore, the value of a single GPU carrier board is RMB 5200.
2) NVSwitch, a basic module for communication between GPUs based on the NVLink standard. The carrier of NVSwitch is a product similar to a carrier board. The processing requirements are relatively simple. The key is to bear the performance of high-speed transmission of large amounts of data. According to industry chain research, the value of a single unit is about US$30, or RMB 195 per unit. Based on the calculation of 6 units in an A100, the value of a single unit is RMB 1170.
3) OAM, or OCP Accelerator Module, is a card used to host GPU chips . There's a one-to-one correspondence between OAMs and GPUs; for example, with the DGX A100 containing 8 GPUs, one AI server requires 8 OAMs. In terms of area, referencing the PCIe version's dimensions of 267.7mm*111.15mm (internal PCB dimensions are basically the same as the casing), the OAM's area can be calculated to be approximately 0.03 square meters. Regarding the PCB layout, since OAM involves high-speed, multi-line signal transmission for the GPU, according to industry research, the SXM version of the DGX A100 OAM requires 20 layers of Ultra Low Loss CCL material and a 4th-order HDI process, with a corresponding product price of 12,000 yuan/square meter. The PCIe version of the DGX A100 OAM has relatively lower specifications, requiring only 14 layers of Ultra Low Loss CCL material. The product, manufactured using a blending process with high Tg FR4 grade CCL material and a first-stage HDI process, has a unit price of 7000 yuan per square meter. Overall, based on the DGX A100 model configuration, the OAM (Original Equipment Maintenance) value per unit for a high-end AI server would reach 2880 yuan.
4) UBB, Unit Baseboard, is a PCB board used to mount the entire GPU platform. One AI server corresponds to one UBB. Based on the bottom surface specifications of the DGX A100 and industry chain research, we estimate that the UBB area is about 0.30 square meters, which requires a 26-layer through-hole PCB board. The CCL material uses Ultra Low Loss, with a corresponding unit price of about 10,000 yuan/square meter, and a corresponding unit value of 3,000 yuan.
In summary, the NVIDIA DGX A100 GPU board consists of four main parts: the GPU carrier board, NVSwitch, GPU accelerator card, and GPU module board. The total PCB area per unit is 0.624 square meters, corresponding to a PCB value of 12,250 yuan. Among them, the carrier board-level products account for 6,370 yuan (52%), and the PCB-level products account for 5,880 yuan (48%).
3. CPU motherboard assembly: Valued at 2845 yuan per unit, with the carrier board accounting for 46% and the motherboard accounting for 40%.
The CPU motherboard assembly consists of the CPU carrier board, CPU motherboard, and accessory boards, among which the functional accessory boards include system memory cards, network cards, expansion cards, and storage operating system driver boards.
1) CPU carrier board: According to industry chain research, the specifications of CPU carrier boards and GPU carrier boards are similar. If the value of a single CPU carrier board is US$100 and DGX is equipped with 2 CPUs, the value of a single machine is about RMB1300.
2) CPU motherboard: Primarily used to house the CPU chip, PCIe switch chip, TPM module, and various functional expansion cards. The specifications of this type of PCB board are mainly determined by the CPU platform design and bus standard. The DGX A100 solution primarily uses a 64-core AMD ROM CPU chip, and the bus standard remains PCIe 4.0. Therefore, the CPU motherboard still uses a 10-12 layer, Low Loss grade CCL material, and through-hole design. According to industry chain research, the unit price is approximately 3000 RMB/square meter. Based on the DGX A100's size specifications, the estimated CPU motherboard area is 0.38 square meters, thus the unit value of the CPU motherboard can be calculated as 1140 RMB.
3) Functional paneling: There are many types of panels. According to industry chain research, the panels generally used are 8-10 layer boards, Mid-Loss grade CCL, with a unit price of approximately 1500 yuan/square meter. The area and quantity, referring to DGX A100, are as follows:
CPU memory cards: The DGX A100 is designed with 32 CPU memory cards, totaling 2TB of RAM . Generally speaking, server CPU memory cards have a relatively unified standard size in the industry, and the estimated area of a single memory card is about 0.004 square meters/piece.
The DGX A100 network interface card (NIC) uses the Mellanox ConnectX series (with X-7 and X-6 options available). The standard configuration includes 10 NICs (8 single- port 200Gb/s Ethernet ports and 2 dual-port 200Gb/s Ethernet ports ). Based on NVIDIA's website, the Mellanox ConnectX-7 measures 68.90mm x 167.65mm, resulting in a per-NIC board area of approximately 0.012 square meters.
Riser Card: Servers use rippers to expand PCIe interfaces due to board placement design. The DGX A100 has a horizontal Standard networking card, so a ripper card is required. According to industry research, the area of this ripper card is approximately 0.01 square meters per card.
The DGX A100 will be equipped with two 1.92TB M.2 NVMe system drives , but the two drives are mounted on opposite sides of a single PCB board. Therefore, there is only one system drive board, with an area of approximately 0.01 square meters per board.
The above four parts combined result in a functional panel area of 0.27 square meters per unit, corresponding to a unit value of approximately 405 yuan.
In summary, the total PCB area used in the NVIDIA DGX A100 CPU motherboard assembly is 0.662 square meters, with a single unit value of approximately RMB 2,845. Among them, carrier-level products account for 46%, PCB-level motherboard products account for 40%, and PCB-level accessory products account for 14%.
4. Other accessories: Total value of a single unit: 226 yuan
Besides the GPU board assembly and CPU module assembly, other components include power supplies, hard drives, and front control panels. According to industry chain research, these products mainly use 6-10 layer, FR4/Mid Loss grade CCL specifications, with a unit price of approximately 1000-1500 yuan/square meter. Based on the DGX A100 specification, the usage and area calculations are as follows:
1) Power supply: In terms of usage, the DGX A100 is equipped with 6 power supplies. Referring to the specifications of Delta Electronics' 2200W server power supply DPS-2200-AB-2 (73.5*265.0mm), we estimate that the PCB board area of a single power supply is 0.019 square meters.
2) Hard drives: In terms of usage, the DGX A100 comes with 8 hard drives. Based on the industry standard of 3.5" drives, we estimate that the PCB area of a single hard drive is 0.008 square meters.
3) The front control panel, mainly used to control external devices, is a PCB board placed in the middle of 8 hard drives. According to industry chain research, we estimate that the area of this board is about 0.010 square meters.
Based on the combined GPU board assembly, CPU motherboard assembly, and accessories, we estimate the total PCB area used in the DGX A100 system to be 1.474 square meters, with a single unit value of 15,321 yuan. Among these, the GPU board assembly accounts for 12,000 yuan (80%), the CPU motherboard assembly accounts for 2,845 yuan (19%), and other accessories account for 226 yuan (1%). In terms of board level classification, the carrier board level accounts for 7,670 yuan (50.1%), and the PCB board level accounts for 7,651 yuan (49.9%).
