Thermal Management on AI Server PCBs: Copper Coin, Thermal Vias and Heatsink Integration

Introduction

The artificial intelligence revolution is fundamentally a hardware challenge. As data centers scale up to train large language models and handle massive inference workloads, the thermal design power (TDP) of individual AI accelerators is skyrocketing. When examining what is an AI server, it becomes immediately apparent that moving data is only half the battle; moving heat is equally critical. For hardware engineers, thermal management on AI server PCBs has evolved from an afterthought into a primary driver of board architecture.

The generational leap from A100 to H100 saw TDPs rise dramatically, and the latest architectures push these boundaries even further. Managing 700W to over 1000W of heat dissipation on a single dense board requires more than standard FR4 and basic airflow. It demands advanced thermal management PCB techniques, specifically the strategic use of thermal vias, embedded copper coins, and robust heatsink integration. This comprehensive guide explores how layout engineers and manufacturers collaborate to keep next-generation AI silicon from melting down.

1. Table of Contents

• Introduction
• The Cooling Challenge in Modern AI Servers
• Thermal Vias: The Foundation of PCB Heat Dissipation
• Embedded Copper Coins: Localized Extreme Cooling
• Heatsink Integration and Board-Level Mechanics
• Comparing PCB Thermal Management Techniques
• Structural Diagram: Heat Flow in AI Server PCBs
• Material Synergy and Manufacturing
• Frequently Asked Questions (FAQ)

The Cooling Challenge in Modern AI Servers

In traditional server designs, processors dissipated 150W to 250W. The heat was manageable through standard thermal interface materials (TIMs) and air-cooled aluminum fin heatsinks. Today, high-performance AI chips, such as those built on the NVIDIA Blackwell architecture, operate at significantly higher power densities. This extreme power consumption creates massive thermal bottlenecks.

Heat transfer in a PCB occurs through conduction. The thermal conductivity of standard FR4 dielectric material is notoriously poor—typically around 0.25 to 0.3 W/(m·K). Copper, on the other hand, boasts a thermal conductivity of roughly 385 W/(m·K). The goal of thermal management PCB design is to create efficient "thermal highways" using copper to bypass the insulating FR4 and move heat from the junction (T_j) of the semiconductor to the ambient environment (T_a) as rapidly as possible.

If heat is not properly dissipated, several catastrophic failures can occur:

• Thermal Throttling: AI GPUs will lower their clock speeds to prevent damage, drastically reducing training and inference throughput.
• Solder Joint Fatigue: Continuous thermal cycling (expansion and contraction) causes micro-cracks in the solder joints under the GPU, eventually leading to open circuits. This is a massive concern in complex BGA assembly processes.
• PCB Delamination: Localized hot spots (often exceeding 130°C) can cause the dielectric resins to outgas and separate, permanently destroying 30+ layer HDI PCBs.

Thermal Vias: The Foundation of PCB Heat Dissipation

For most AI accelerator PCB design projects, thermal vias are the first line of defense. A thermal via is simply a plated through-hole (PTH) or microvia placed directly underneath a heat-generating component. Its sole purpose is to conduct heat away from the surface component and transfer it to internal copper planes or to the opposite side of the board where a heatsink is attached.

To maximize the efficacy of thermal vias, layout engineers must calculate and optimize the following parameters:

1. Via Diameter and Pitch
The standard diameter for a thermal via is usually between 0.2mm and 0.3mm. Making the holes too large can lead to solder wicking during the reflow process, which starves the component pad of solder and creates weak joints. The pitch (distance between the centers of two vias) is typically set at 0.8mm to 1.2mm to maximize density without compromising the structural integrity of the bare board.

2. Plating Thickness
The thermal resistance of a via (R_th) is inversely proportional to the cross-sectional area of the copper plating. While standard IPC class 2/3 vias might have 20μm to 25μm of copper plating on the walls, thermal vias in AI server boards often require heavy copper plating (up to 40μm or more) to increase the conductive mass. The formula for thermal resistance is R_th = L / (k × A), where L is the length of the via, k is the thermal conductivity of copper, and A is the cross-sectional area.

3. Via Filling and Capping (VIPPO)
To prevent solder wicking while maximizing heat transfer, thermal vias under high-density BGAs must utilize Via-in-Pad Plated Over (VIPPO) technology. The vias are filled with a specialized epoxy. While non-conductive epoxy has a low thermal conductivity (around 0.2 to 0.6 W/m·K), silver-filled conductive epoxies can be used to boost thermal conductivity to 3.5 - 8 W/m·K. After filling, the via is plated over with copper to create a flat, solderable pad.

Embedded Copper Coins: Localized Extreme Cooling

When the heat flux (W/cm²) generated by an AI ASIC or high-power voltage regulator module (VRM) exceeds the capability of thermal via arrays, engineers turn to copper coin PCB technology. A copper coin is a solid piece of copper that is physically embedded into the PCB directly under the hot component.

Because it is solid copper, a coin provides an uninterrupted thermal path with a thermal conductivity of 385 W/m·K—massively outperforming even the densest array of thermal vias. There are several geometries of copper coins used in server motherboard PCB manufacturing:

• I-Coin: A simple cylindrical or rectangular block embedded through the entire thickness of the PCB. It conducts heat straight from the top layer to the bottom layer.
• T-Coin: Shaped like the letter T, this coin provides a large surface area on one side (usually the component side) to gather heat, funneling it down through a narrower channel to save routing space on internal layers.
• U-Coin: Typically used for cavity-down designs, allowing components to sit inside the board while being surrounded by solid copper for multi-directional heat dissipation.

Manufacturing a PCB with embedded copper coins is highly complex. The PCB fabricator must route a precise cavity in the pre-pressed board. The copper coin is then press-fitted or laminated into the cavity. Tolerances are critical: if the coin protrudes even 20μm above the board surface, it will cause tombstoning or open BGA connections during assembly. If it sits too low, the thermal paste will not make adequate contact, creating an insulating air pocket.

Heatsink Integration and Board-Level Mechanics

Moving heat through the PCB is useless if it cannot be expelled into the air or liquid cooling loop. Heatsink integration on AI server boards is a meticulous mechanical engineering task. The sheer weight and mounting pressure of modern AI heatsinks present unique challenges to the PCB structure.

Thermal Interface Materials (TIM)
Air gaps are the enemy of thermal management. To bridge the microscopic imperfections between the silicon die, the copper coin/thermal vias, and the heatsink base, engineers use Thermal Interface Materials. TIM1 is applied between the silicon die and the heat spreader, while TIM2 is applied between the PCB/Heat spreader and the main heatsink. For AI servers, high-performance phase-change materials or liquid metal pads (conductivity > 10 W/m·K) are becoming standard.

Mechanical Stress and Warpage
Heavy copper heatsinks, vapor chambers, or liquid cold plates must be bolted down tightly to ensure minimum thermal resistance. This high mounting pressure (often exceeding 30-50 PSI) can bow the PCB. To combat this, AI server PCBs utilize heavy-duty metal stiffeners and backplates on the reverse side. Without these stiffeners, the board would warp under thermal stress and mechanical load, snapping the delicate BGA solder balls of the adjacent components.

Comparing PCB Thermal Management Techniques

Layout engineers must choose the right technique based on cost, required routing space, and thermal output. The table below outlines the primary differences.

Feature	Thermal Vias (VIPPO)	Embedded Copper Coin	Metal Core PCB (MCPCB)
Thermal Conductivity (Z-Axis)	Medium (~10-30 W/m·K equivalent)	Extremely High (~385 W/m·K)	High (~100-200 W/m·K)
Manufacturing Cost	Moderate	Very High	High
Routing Space Impact	High (Blocks internal routing channels)	Very High (Requires large keep-out zones)	Total (Cannot route internal layers)
Best Use Case in AI Servers	Standard ASICs, FPGAs, PCIe controllers	Core AI GPUs, High-power VRMs	LED lighting, not suited for complex logic
Assembly Complexity	Standard (Requires planarization)	Very High (Strict Z-axis tolerances)	Standard

Structural Diagram: Heat Flow in AI Server PCBs

To visualize how these components interact to remove heat from a high-power chip, reference the structural stackup diagram below.

=========================================================
                 [ LIQUID COLD PLATE / HEATSINK ]          <-- Final Heat Exhaust
=========================================================
                       | | | | |
               [ Thermal Interface Material (TIM2) ]       <-- Fills Microscopic Gaps
                       | | | | |
=========================================================
                  [ AI GPU / ACCELERATOR DIE ]             <-- Heat Source (1000W+)
=========================================================
                       | | | | |
               [ Thermal Interface Material (TIM1) ]
                       | | | | |
=========================================================
   [ PCB Top Layer ]       |||      [ PCB Top Layer ]
   [ Internal Routing]     |||      [ Internal Routing]
   [ Power Planes ]    [ SOLID ]    [ Power Planes ]       <-- FR4 acts as insulator
   [ Ground Planes ]   [ COPPER]    [ Ground Planes ]      <-- Copper Coin creates highway
   [ Internal Routing] [ COIN  ]    [ Internal Routing]
   [ PCB Bottom Layer]     |||      [ PCB Bottom Layer]
=========================================================
                       | | | | |
               [ Thermal Interface Material (TIM3) ]
                       | | | | |
=========================================================
              [ RIGID METAL BACKPLATE / STIFFENER ]        <-- Prevents Board Warpage
=========================================================

Material Synergy and Manufacturing

Thermal management does not exist in a vacuum; it must coexist with signal integrity. AI servers utilize ultra-fast interconnects. As discussed in our guide on what NVLink is and how it shapes routing, signals are traveling at 112G to 224G PAM4 speeds. This requires ultra-low-loss laminates.

The challenge is that many high-speed PCB materials (like Megtron 7 or Rogers) have different coefficients of thermal expansion (CTE) compared to copper. As the copper coin and the PCB heat up, they expand at different rates. If the materials are not carefully matched, the CTE mismatch will shear the copper vias or delaminate the board. Experienced fabricators managing GPU PCB manufacturing must meticulously calculate resin flow during pressing to ensure the copper coin bonds perfectly with the surrounding high-speed prepreg without causing resin starvation.

Frequently Asked Questions (FAQ)

Q: Can I use thermal relief pads on vias under an AI GPU?
A: No. Thermal reliefs (spokes) restrict the flow of heat and current. Under high-power AI GPUs, you must use solid, direct connections to the copper planes to maximize thermal mass and minimize electrical inductance.

Q: What is the maximum thickness a copper coin can be in a PCB?
A: Copper coins can be manufactured to match the full thickness of the PCB, which in AI servers can range from 3.2mm to over 5.0mm. However, thicker coins require highly specialized press-fit and lamination profiles.

Q: Are conductive epoxy-filled vias as good as copper coins?
A: No. Silver-filled conductive epoxies typically max out around 8-10 W/m·K. Solid copper is roughly 385 W/m·K. While conductive epoxy vias are significantly better than empty vias, they cannot match the sheer thermal transfer volume of a solid copper coin for extreme TDP applications.

Q: Does thermal management affect routing?
A: Yes, severely. Both thermal via arrays and copper coins create massive "keep-out" zones on all internal layers, forcing layout engineers to route critical high-speed differential pairs around these thermal structures, often requiring additional PCB layers to complete the routing.

Conclusion & Call to Action

As AI workloads continue to push the physical limits of silicon, thermal management on AI server PCBs remains one of the most critical aspects of hardware engineering. Implementing VIPPO thermal vias, navigating the manufacturing tolerances of embedded copper coins, and ensuring robust heatsink mechanics are essential for reliable AI infrastructure. A brilliant chip design is rendered useless if the PCB cannot exhaust its heat.

Designing complex, thermally demanding AI hardware requires a manufacturing partner capable of extreme precision, heavy copper plating, and advanced HDI capabilities.

Need to manufacture AI server PCBs? Get a quote from NextPCB →