High-performance GPU PCB board for AI server systems

1.Classification

In a typical AI server(such as the NVIDIA DGX architecture),the PCBs are classified by their specific roles:
◈ UBB(Unit Baseboard): The"main deck"that hosts the entire GPU platform.It connects multiple GPU modules and manages high-speed communication between them.
◈ OAM(Accelerator Module):The individual mezzanine card that carries the actual GPU chip and its local memory.
◈ GPU/CPU Substrates:Small,extremely high-density carriers(IC Substrates) that sit directly under the silicon die.
◈ Switch Carrier Boards:PCBs that house NVSwitch chips to facilitate ultra-fast"all-to-all"communication between GPUs.

2.Differences from Standard PCBs

AI server PCBs demand manufacturing tolerances and material performance is far exceed the capabilities of standard server or consumer-grade electronics:

3.Technical Requirements

In order to support AI product technology requirement, these boards must meet below items. 
◈ A.High-Speed Signal Integrity
AI servers use interconnects like PCIe 5.0/6.0 and NVLink,operating at speeds of 56G or 112G PAM4.
Requirement: Requires"Ultra-Low Loss"materials with a Dissipation Factor(DF) <0.002 to prevent signal degradation.
Precision:Impedance tolerances are tightened to +/- 5%(compared to the standard +/- 10%).
◈ B.Advanced Thermal Management
With a single AI server rack generating thermal loads comparable to an entire household, extreme thermal management is mandatory. PCBs must utilize high Glass Transition Temperature (Tg > 180℃) materials and a low Coefficient of Thermal Expansion (CTE) to maintain structural integrity and prevent warping or delamination during continuous, high-intensity operation.
The AI Standard: Because AI GPUs run at near-maximum TDP (Thermal Design Power) 24/7, the board is under constant thermal soak.
Design:Integration of copper-filled thermal vias and sometimes metal-core layers for liquid cooling interfaces.

◈ C.Extreme Power Integrity
Modern GPUs demand massive current delivery—often hundreds of Amperes—at sub-1V levels. To support this, AI PCBs utilize heavy copper planes and ultra-thick power layers to minimize IR Drop (voltage sag due to resistance), ensuring stable power delivery and preventing 'voltage starvation' during peak computational workloads.
◈ D.HDI & Precision Drilling
To fit thousands of pins from a GPU BGA into a small area,HDI(High-Density Interconnect) technology is mandatory.
Requirement: Multiple"steps"of laser drilling(e.g.,4-step or 8-step HDI) and Any-Layer via structures that allow signals to travel vertically between any two layers without taking up space on others.

To achieve the ultra-fast speeds and thermal stability required by AI GPU servers,designers must move away from standard FR4 materials.They instead use Advanced High-Speed Laminates and Substrates characterized by low signal loss and high thermal resistance.
Here is a detailed breakdown of the specific materials used in AI high-speed designs.
◈ 1.High-Speed Material Hierarchy (By Loss Level)
In PCB engineering,materials are classified by their Dissipation Factor(DF)—the lower the DF, the less signal is "lost"as heat.

◈ 2.Key Material Leaders&Series
A.Panasonic Megtron Series(The Industry Standard)
The Megtron family is currently the dominant choice for AI GPU Baseboards(UBB) and Switch boards.
● Megtron 6:The workhorse for PCIe 5.0. It provides a balance of cost and performance.
● Megtron 7/7N: Specifically designed for 112G PAM4 signaling.It features extremely low transmission loss and high thermal reliability for multi-layer counts(30+layers).
● Megtron 8:The newest generation,optimized for the next wave of 1.6T networking and Blackwell-generation(NVIDIA) GPU interconnects.
B.Rogers(High-Frequency Specialists)
While Megtron is often used for the"digital"layers,Rogers Corporation materials are used when"RF-like"precision is needed.
● RO4000 Series:These are ceramic-filled laminates.They are often used in Hybrid Stack-ups(e.g.,Rogers layers for high-speed signals and FR4 layers for power)to save cost while maintaining performance.
● CLTE-XT:Used in environments with extreme temperature fluctuations because it has incredibly low CTE(Expansion),ensuring the microvias don't crack under heat.
C.Isola
● Tachyon 100G:A direct competitor to Megtron 7,designed for high-speed digital backplanes and line cards.
● I-Tera MT40:Often used in high-speed designs that also require heavy thermal cycling.

4.Critical Material Properties for AI

When choose these materials for AI PCB,engineers will consider below three specific parameters for them:
◈ Dielectric Constant(Dk):
● Target: Lower and stable. A lower Dk allows signals to travel faster.
● AI Requirement: Usually 3.0 to 3.8.the signal travels roughly 15% to 20% faster.
◈ Dissipation Factor(DF):
● Target:Minimal.At 112Gbps,a high Df will cause the signal to"die"before it reaches the other end of the board.
● AI Requirement: DF <0.002.
◈ Coefficient of Thermal Expansion(CTE):
● Target: Matches Copper.If the board expands faster than the copper vias when it gets hot(GPU heat),the vias will snap.
● AI Requirement:Z-axis CTE<3.0%.
◈ Copper Foil:The"Skin Effect"
At AI speeds(high frequency),electrons don't flow through the copper;they flow on the surface(the Skin Effect).
Standard Copper:Too rough.The"peaks and valleys"on the surface slow down the signal.
HVLP(Very Low Profile)Copper:To combat the 'Skin Effect' at high frequencies, AI PCBs utilize HVLP (Hyper-Very-Low-Profile) copper.1 By polishing the copper surface to a mirror-like smoothness (Rz < 1.5μm), the signal path is streamlined, significantly reducing resistive loss and phase jitter compared to standard, rough-surfaced foils
Summary Table:The"Gold Standard"Stack-up for AI

In the context of high-performance AI hardware, a Hybrid Stack-up is a strategic design approach where different types of laminate materials are combined within a single PCB.
Since materials like Megtron 7 or Rogers 4350B is the 5 to 10 times cost more than standard FR4, using them for every layer of a 30-layer board would be prohibitively expensive.
◈ 1. How a Hybrid Stack-up Works
In a hybrid design, engineers "mix and match" materials based on the function of each layer:
High-Speed Layers: The layers carrying critical 112G PAM4 signals (NVLink or PCIe 6.0) use Ultra-Low Loss (ULL) materials.
Power and Ground Layers: The layers responsible for delivering 700W+ of power or providing structural stability use Standard High-Tg FR4.
Bonding Sheets (Prepreg): Specialized "No-Flow" or "Low-Flow" prepregs are used to glue these different materials together during the lamination process.
◈ 2. Why Use Hybrid Stack-ups in AI Servers?
Cost Optimization: You only pay for expensive "super-materials" where they are technically required for signal integrity.
Structural Integrity: Ultra-low loss materials are often more brittle or have different thermal expansion rates. Mixing them with standard FR4 can actually make the board more mechanically robust.
Thickness Management: AI PCBs are already very thick (often >3mm). Standard materials allow for easier control of the overall board thickness compared to using specialized laminates exclusively.
◈ 3. The Challenges of "Mixing" Materials
Creating a hybrid board is significantly more difficult than a uniform one due to:
Coefficient of Thermal Expansion (CTE) Mismatch: Different materials expand at different rates when they get hot. If not balanced correctly, the board will warp (bow and twist) or the copper traces will delaminate (peel off).
Symmetry Requirement: Designers must keep the stack-up symmetrical from the center (e.g., if Layer 3 is Megtron 7, Layer 28 should also be Megtron 7) to prevent "potato-chipping" during the soldering process.
Drilling Precision: The laser drill behaves differently when hitting ceramic-filled Rogers material versus resin-heavy FR4, requiring the manufacturer to constantly adjust laser parameters.
◈ 4. Comparison of Materials in a Hybrid Board

5.Application

While we’ve focused heavily on the AI Server as the primary driver, the ultra-high-spec "AI PCB" is rapidly becoming the standard for any industry moving toward autonomous decision-making or high-speed data processing.
Here are the primary application domains where these high-end boards are currently deployed
◈ 1. AI Infrastructure & Data Centers
This is the "Mount Everest" of PCB design. These boards aren't just components; they are the physical medium through which LLMs (like GPT-4 or Gemini) are born.
GPU Accelerator Cards: Boards for NVIDIA H100/H200 or AMD MI300X that utilize HBM (High Bandwidth Memory). The PCB must manage the massive 2.5D/3D packaging of these chips.
Switch/Network Boards: To handle the "East-West" traffic between GPU clusters, network switches use these materials to support 800G and 1.6T Ethernet speeds.
◈ 2. Autonomous Mobility (Automotive AI)
Self-driving cars are essentially "AI servers on wheels." They require the same signal integrity but with much higher reliability standards.
AD/ADAS Modules: Boards that process real-time data from LiDAR, Radar, and high-res cameras.
Safety-Critical Reliability: Unlike a server in a cooled room, these PCBs must meet IATF 16949 standards, surviving extreme vibrations and temperature swings from -40℃ to +125℃ without signal drift.
◈ 3. Edge AI & Smart Devices
Not all AI happens in the cloud. "Edge" devices need to perform inference locally to reduce latency and save bandwidth.
Industrial Robotics: PCBs that enable computer vision for high-speed assembly line sorting or autonomous mobile robots (AMRs) in warehouses.
Smart Medical Imaging: Real-time AI processing in portable ultrasound or MRI machines, where the PCB must handle high-speed data without generating EMI (Electromagnetic Interference) that could distort the medical images.
◈ 4. Advanced Packaging (Chiplets)
The line between a "PCB" and a "Chip" is blurring.
IC Substrates: AI PCBs are scaling down to become the "interposers" for chiplet-based architectures. Here, the manufacturing moves from standard etching to semiconductor-grade lithography, with line widths smaller than 10μm.

Send inquiry