Dissecting the AI Computing Power Beast: From GPU to 64‑Layer PCB, Uncovering the Extreme Challenges of HDI and Back‑Drilling Processes

With the explosive emergence of large‑language models such as ChatGPT, DeepSeek and Gemini, global computing power demand has witnessed an exponential surge. While everyone is talking about the shortage of GPU chips, hardware engineers are confronted with another severe underlying challenge: how to support such staggering computing‑power density?

The answer lies within those unassuming green circuit boards. The extreme requirements of AI servers for signal transmission speed, power integrity and thermal dissipation are forcing a revolution in the PCB industry centered on layer count and interconnect density. High‑layer‑count PCBs and HDI technology have become the indispensable 'invisible cornerstone' for AI‑hardware R&D.

I. Dissecting the AI Server: 'Blood Vessels' and 'Skeleton' of the Computing‑Power Beast

Unlike ordinary servers, AI servers feature an extremely complex hardware architecture. According to an in‑depth analysis by Sinolink Securities, the PCBs inside a high‑end AI server fall into three core categories: GPU board assembly, CPU motherboard assembly, and auxiliary components

Take NVIDIA DGX A100 as an example. It can be broadly divided into five hardware sections:

1.Fan Module: The fan module of the DGX A100 consists of 8 fans, a configuration largely consistent with that of conventional 8U‑form‑factor servers.

2.Hard Drives: Hard drives and the front control board (for signal transmission between control circuits and external devices) are located beneath the fan‑module section. The DGX A100 is equipped with eight 3.84TB hard drives, delivering a total internal storage capacity of 30TB.

3.GPU Board Assembly: The rear section houses the assembly area for key components of the entire AI server. The most critical segment is the GPU board assembly, which also marks the core difference between AI servers and conventional servers. From the architecture of the DGX A100, the GPU board assembly mainly comprises three parts: GPU components, module board and NVSwitch. All three parts involve different types of PCB products.

4.CPU Motherboard Assembly: This is the core component for all servers, both conventional and AI servers. It consists of parts including the CPU motherboard, system memory, network adapters, and PCIe Switch. The CPU motherboard, system memory and network adapters account for the major PCB consumption in this section.

5.Power Supply Module: Six power supplies are fitted at the lower‑rear section of the DGX A100. Heavy‑copper PCBs are adopted inside the power supplies.

In terms of value content, the PCB value per unit of a conventional server is approximately RMB 2,425, while that of a DGX A100 reaches as high as RMB 15,321, representing a 532% increase. Among this value growth, 80% of the incremental value comes from the GPU board assembly.

As a super base supporting eight GPUs, the GPU module board requires an extremely high layer count for routing. In the DGX A100, the GPU module board covers an area of approximately 0.3 m². It typically adopts a 26‑layer through‑hole board and must use Ultra‑Low‑Loss grade materials.

GPU accelerator cards serve as the core units carrying GPU chips. For the DGX H100, the OAM even applies 5‑order HDI technology to meet the demand for ultra‑high‑density interconnections between chips.

As for CPU motherboards, even general‑purpose CPU motherboards commonly feature designs with 10‑12 layers or more under the PCIe 5.0 / 6.0 bus standard, accompanied by a substantial unit‑price increase.

II. Technical Pain Points in High‑Layer‑Count PCB Design and Manufacturing

For hardware R&D engineers, designing high‑layer‑count PCBs for AI applications means overcoming multiple major obstacles. Even if the design is completed successfully, numerous challenges remain throughout the manufacturing process. Two typical examples are illustrated below:

1.Extremely Stringent Signal Integrity

In AI servers, GPU interconnections and PCIe buses operate at extremely high speeds. To mitigate signal reflection, the back‑drilling process is mandatory during the manufacturing of high‑layer‑count PCBs. Back‑drilling refers to partial‑depth drilling performed from one side of a multi‑layer board to remove unwanted inner‑layer electrical interconnections. It eliminates the adverse effects of stubs on signal integrity and guarantees high‑frequency signal integrity as well as impedance continuity. Take Sunkey as an example: it has launched a back‑drilling process for high‑layer‑count PCBs, supporting board constructions from 4 to 68 layers. In addition, impedance must be tightly controlled, with typical tolerance requirements of ±7% or even tighter.

2.Manufacturing Difficulties from Ultra‑High Aspect Ratio

To support high current and heat dissipation, PCBs for AI servers generally have board thickness ranging from 3.0 mm to 5.0 mm. This results in an extremely high drilling aspect ratio. At a mechanical‑drilling aspect ratio of 20:1 (hole depth / hole diameter), for a 5.0 mm‑thick board, the minimum hole diameter is merely 0.2 mm. This poses a severe challenge for via electroplating.

III. R&D Breakthrough: How Sunkey Breaks Through Technical Barriers with One‑Stop 'High‑Layer‑Count + HDI' Solutions

For hardware R&D engineers, designing AI‑oriented high‑layer‑count PCBs often means confronting a triple dilemma: “unmanufacturable, slow turnaround and high cost”. Leveraging its digital intelligent manufacturing capabilities, Sunkey has standardized services for high‑layer‑count PCBs and HDI boards, delivering a new solution for engineers.

1.Breaking Layer‑Count and Lead‑Time Bottlenecks: From Month‑Long Waiting to Weekly Delivery

AI‑server motherboards often feature 20 or more layers with board thickness typically ranging from 3.0 mm to 5.0 mm. When traditional manufacturers receive such high‑layer‑count orders, production scheduling is usually required, resulting in prototype lead times of 6‑8 weeks, which greatly slows down R&D iteration speed.

Sunkey’s current manufacturing capability covers 4‑ to 68‑layer PCBs. It can not only meet requirements for standard AI board cards, but also handle extremely complex circuit‑integration designs for aerospace and other fields. Furthermore, thanks to automated production lines and digital workflows, Sunkey has cut the prototype lead time for high‑layer‑count boards down to 15‑25 days — twice as fast as the industry average, helping R&D teams seize market opportunities.

2.Overcoming High‑Density Interconnection Challenges: Popularization of HDI and Back‑Drilling Technologies

AI chips (such as GPUs and Switch chips) feature extremely high pin density, which mandates the adoption of HDI technology. Nevertheless, high‑order HDI involves complex processes. The precision of laser drilling is difficult to control. Besides, the back‑drilling process required for high‑speed signals tends to leave stub residues, impairing signal integrity.

Sunkey supports high‑layer‑count boards with arbitrary‑order HDI blind and buried via designs. It is equipped with high‑precision laser drilling (minimum hole diameter: 0.075 mm) and via‑fill electroplating processes, which deliver perfect compatibility with high‑density BGA packages. Meanwhile, its high‑precision back‑drilling service effectively reduces signal reflection. Furthermore, LDI (Laser Direct Imaging) technology is adopted to achieve layer‑to‑layer alignment accuracy up to 15 μm for ultra‑high‑layer‑count boards.

3.Reshaping Costs: Making High‑Layer‑Count Prototyping Affordable

High‑layer‑count boards incur extremely high prototyping costs due to low yield rates and expensive raw materials. In addition, board warpage or open/short‑circuit defects can lead to scrapping of costly AI chips.

Sunkey exclusively adopts high‑frequency and high‑speed materials from well‑known laminate brands, including SYTECH, NOUYA, ITEQ and Panasonic, to guarantee high heat resistance and low coefficient of thermal expansion, suitable for high‑temperature operating conditions of AI servers. Benefiting from high operational efficiency, its prototyping cost for high‑layer‑count PCB is approximately 50% lower than competitors, greatly lowering the trial‑and‑error threshold for R&D.

Competition in AI computing power is ultimately a competition in hardware engineering. When you are designing next‑generation AI servers, accelerator cards or high‑performance computing units, there is no need to worry about PCB layer‑count constraints or long lead times. With industrial‑grade quality and highly competitive pricing, Sunkey’s high‑layer‑count and HDI services serve as a solid backing for your hardware innovation journey.