sales@ipcb.com
As AI servers continue to develop towards faster speeds, higher performance, and larger capacities, the parameters and performance of PCBs are also improving accordingly.
In the previous article（PCB: The AI Evolution of the “Mother of the Electronics Industry" (Ⅱ)）, we mainly introduced the types of PCBs and the application areas of AI PCBs. In this article, we will specifically introduce the improvement of PCB performance
under AI computing power.
(1) PCIe Models: Basic PCB Requirements under a General Interconnect Architecture
GPU cards are installed through the PCIe slots on the server. GPU cards are interconnected via the PCIe bus, an internal bus and a computer expansion bus standard.
It is a high-speed serial, high-bandwidth expansion bus, typically used on motherboards to connect peripherals such as graphics cards, solid-state drives, various acquisition cards,
and wireless network cards. PCIe is not limited to motherboards; it is also used for interconnections between many chips.
(2) NVLink/SXM Models: NVIDIA's solution for high-performance GPU interconnection. It uses a proprietary protocol laid on the circuit board,
similar to how GPUs are mounted—directly on the circuit board. GPUs are interconnected via NVLink links.
(3) Value Quantity
The value of components such as PCB, memory, power supply, and SSD all increases. When upgrading from a regular server to an AI training server,
the value of components such as memory, SSD, PCB, and power supply increases several times over. This is driven by the extreme demands of AI training on hardware performance:
Memory: AI training needs to process massive datasets and complex neural network models simultaneously, resulting in a geometric increase in the requirements for memory capacity,
bandwidth, and speed. Regular servers typically use tens of GB of DDR4 memory, while AI training servers require hundreds of GB or even higher capacity
DDR5 (or more advanced specifications) memory to support high-concurrency data exchange during multi-GPU parallel computing. Its value increases several times over with
the upgrade in capacity and specifications.
SSD: AI training involves frequent dataset loading and intermediate calculation result caching, requiring storage speed and capacity far exceeding those of regular servers.
While typical server SSDs are mostly 100GB-level SATA interfaces, AI training servers utilize terabyte-level NVMe high-speed SSDs (or even storage-class memory, SCM)
to meet the read/write demands of several gigabytes per second. This performance upgrade of storage media directly drives a significant increase in its value.
PCB (Printed Circuit Board): AI training servers integrate multiple high-power GPU cards and high-speed interconnect modules, placing stringent requirements on
the number of PCB layers, materials, and signal integrity. Typical server PCBs are mostly low-layer, conventional board materials, while AI servers require high-multilayer PCBs
(even rigid-flex PCBs) and high-speed copper-clad laminates to ensure the stability of high-speed signal transmission and high-current power supply.
The increased complexity of the manufacturing process and the higher material costs multiply its value.
Power Supply: AI training servers consume extremely high power (up to several kilowatts per unit), requiring high-power, high-efficiency power supply modules.
While typical server power supplies operate at a few hundred watts, AI server power supplies jump to over 2000W and must meet the 80PLUS Titanium efficiency standard to
minimize energy loss. This upgrade in power and performance significantly increases their value.
This increased value of components is essentially a result of the "high computing power, big data, and high power consumption" characteristics of AI training forcing hardware
performance upgrades. It is also one of the core reasons why the hardware cost of AI training servers is significantly higher than that of ordinary servers.
It is these customized technological upgrades for AI scenarios that have significantly increased the material costs and process complexity of AI server PCBs, ultimately driving their value to multiply several times over. This is not only the result of PCB's own technological iteration but also a reflection of its value as the core carrier of AI computing power.
In today’s era of full AI penetration, computing power competition has become a key determinant of enterprises’ core competitiveness. As the "neural hub" of AI servers, PCB performance upgrades directly determine the efficiency of AI computing power release. We specialize in the R&D of AI-specific PCBs, leveraging high-layer design, high-speed substrate application, and precise process control to provide stable, efficient interconnection solutions for AI servers. Partner with us to seize the computing power advantage in the intelligent era.