· Thomas Webb · Engineering  · 7 min read

PCB Requirements for AI Accelerator Hardware

Designing PCBs for AI accelerators (GPUs, TPUs, custom ASICs) demands extreme specifications: 20-30 layers, 500A+ current delivery, 3/3mil routing, and advanced thermal management. Here's what your fabricator needs to deliver.

Designing PCBs for AI accelerators (GPUs, TPUs, custom ASICs) demands extreme specifications: 20-30 layers, 500A+ current delivery, 3/3mil routing, and advanced thermal management. Here's what your fabricator needs to deliver.

Quick Answer

AI accelerator PCBs require 20-30 layer builds with 2+N+2 or higher HDI structure, 3-5oz copper on power planes for 500A+ delivery, low-loss laminates (Megtron 6/7) for 112G PAM4 signaling, thermal via arrays (1000+ vias per BGA footprint) for 400W+ TDP dissipation, and strict impedance control with backdrilling for 16:1 aspect ratio vias. These boards represent the most demanding specifications in commercial PCB manufacturing—combining high layer count, fine-pitch HDI, heavy copper, and premium materials in a single design.

The Most Demanding PCBs in Commercial Electronics

AI accelerator hardware represents the convergence of every challenging PCB specification simultaneously: high layer count, HDI microvia structures, heavy copper power delivery, ultra-low-loss materials, extreme thermal management, and tight impedance control. A single AI training card can cost $3,000-10,000 just for the bare PCB—before any components.

This guide covers what it takes to fabricate these boards and how design choices impact manufacturability and cost.


Layer Count: Why 20-30 Layers is the Minimum

The Driving Forces

A modern AI accelerator (Nvidia B200, AMD MI300X, Google TPU v5) connects to:

  • HBM3e memory: 8-12 stacks, each with 1024+ I/O at 9.6 Gbps
  • PCIe Gen6 or NVLink: 112G PAM4 per lane, 16-72 lanes
  • Power delivery: 0.7-0.85V core at 500-1000A
  • Management/control: I2C, SPI, GPIO, thermal sensors

Routing all of this requires:

Layer TypeCountPurpose
Signal (high-speed)8-12HBM, PCIe, NVLink traces
Signal (low-speed)2-4Control, I2C, GPIO
Ground reference6-8Adjacent to every signal layer
Power planes4-8Multiple voltage rails at high current
Total20-30Typical range

Why Not Fewer Layers?

Every signal layer needs an adjacent ground reference plane for impedance control. That alone means signal + reference = 2 layers per routing tier. With 6-8 routing tiers needed for BGA escape and interconnect, plus 4-8 power planes for current density, you quickly reach 20+ layers.

Reducing layer count means wider boards (more routing space needed) or higher-density HDI (more expensive buildup). At AI-scale pin counts, 20-30 layers is actually the cost-optimized solution.

HIGH LAYER COUNT

Up to 30-Layer PCBs for AI and HPC Applications

AtlasPCB fabricates 20-30 layer boards with HDI buildup, Megtron 6 laminates, and 5oz copper power planes for AI hardware.

Get Quote for AI Board ›

Power Delivery: 500A+ at Sub-1V

The Challenge

Modern AI accelerators draw 500-1000A at 0.7-0.85V. At these currents, even milliohms of PDN impedance create unacceptable voltage droop. The PCB power delivery network must achieve:

  • DC resistance: < 0.5 milliohm from VRM to BGA
  • AC impedance: < 1 milliohm up to 100 MHz
  • Current capacity: 500A sustained, 800A+ transient

Copper Weight Requirements

Standard 1oz (35um) copper on a 50mm-wide power plane carries approximately 30A with acceptable temperature rise. To deliver 500A, you need:

Copper WeightCurrent per 50mm WidthPlanes Needed (500A)
1oz (35um)~30A17 planes (impossible)
2oz (70um)~60A9 planes
3oz (105um)~90A6 planes
5oz (175um)~150A4 planes

Practical designs use 3-5oz copper on 4-6 power/ground plane pairs, supplemented by wide copper pours on signal layers.

Via Current Distribution

The BGA-to-plane connection uses massive via arrays. A typical AI accelerator BGA has:

  • 300-500 power pins
  • 300-500 ground pins
  • Each via carries 1-2A average

Via resistance dominates at these currents. For a 0.3mm via in a 3.5mm board:

  • 0.5oz plating: ~6 milliohm per via
  • 1.0oz plating: ~3 milliohm per via

1000 vias in parallel: 3-6 microhm contribution — acceptable.


Signal Integrity: 112G PAM4 on a PCB

Material Requirements

112G PAM4 signaling (the Nyquist frequency is ~28 GHz for 56 Gbaud) demands:

ParameterRequirementStandard FR-4Megtron 6Megtron 7
Df @ 12.5 GHz< 0.0040.0200.0020.001
Dk variation< +/-3%+/-5%+/-2%+/-1.5%
Surface roughness< 2um Rz6-8um2-3um (VLP)1-2um (HVLP)
Insertion loss (6”)< 8 dB @ 28 GHz>20 dB5-6 dB3-4 dB

The copper surface roughness contribution to loss becomes dominant above 10 GHz. HVLP (Hyper Very Low Profile) copper is mandatory — standard RTF (Reversed Treat Foil) adds 3-4 dB/inch at 28 GHz.

Backdrilling: Eliminating Via Stubs

A through-via in a 24-layer board creates a stub from the signal exit point to the via end. At 28 GHz, even a 0.5mm stub creates destructive resonance. Backdrilling removes the unused stub portion:

  • Target stub length: < 0.2mm (200um)
  • Backdrilling accuracy: +/-3mil (AtlasPCB specification)
  • Required on: all high-speed signal vias not on outer layers

Without backdrilling, via stubs create notch filters that kill eye diagrams above 16 GHz.

SIGNAL INTEGRITY

Low-Loss Materials + Precision Backdrilling

Megtron 6/7 laminates with HVLP copper foil and backdrilling to +/-3mil stub accuracy for 56+ Gbaud signaling.

View Signal Integrity Specs ›

Thermal Management: Conducting Heat Through the PCB

Thermal Via Arrays

With 400W+ TDP concentrated in a 50x50mm die area, the PCB beneath becomes a thermal bottleneck. Thermal via arrays provide vertical heat conduction:

Via ConfigurationThermal Conductivity (W/mK)Typical Application
No thermal vias0.3 (FR-4 only)Not acceptable
Via array (0.3mm, 1mm pitch)3-8Moderate power
Dense via array (0.2mm, 0.5mm pitch)15-25High power
Filled via array (copper-filled)30-50AI accelerator

For AI boards, the area directly under the chip (die shadow) contains 1000-2000 copper-filled thermal vias in a tight grid, connected to internal copper planes that spread heat laterally to board edges and mounting hardware.

Heavy Copper as Heat Spreader

Internal 3-5oz copper planes serve dual duty: power delivery AND thermal spreading. A 5oz (175um) copper plane has thermal conductivity of 400 W/mK along its plane — 1300x better than FR-4. This makes the internal planes the primary lateral heat transport mechanism.


Manufacturing Challenges and Yield

Aspect Ratio

24-layer board, 3.5mm thickness, 0.2mm via drill: aspect ratio = 17.5:1

This exceeds standard capability. Solutions:

  • Use 0.25mm minimum drill (14:1 aspect ratio) — acceptable
  • Backdrilling reduces effective depth
  • HDI structure avoids full-depth through-vias for many connections

Registration

At 24+ layers, layer-to-layer registration stacking becomes critical. Each layer adds +/-25um registration error. After 24 layers, total stack can reach +/-75-100um, eating into annular ring budgets.

Solution: fabricators use sequential alignment (re-registering every 4-6 layers) rather than single-stack alignment.

Yield Impact

Board TypeTypical First-Pass Yield
8-layer standard95-98%
16-layer HDI88-93%
24-layer AI (standard materials)80-88%
24-layer AI (Megtron 6, backdrill)75-85%
30-layer ELIC + Megtron 765-78%

Lower yield directly impacts cost — and why AI PCBs carry premium pricing.

AI HARDWARE FABRICATION

High-Complexity PCBs for AI Infrastructure

Up to 30 layers, 5oz copper, 16:1 aspect ratio, Megtron 6. We manufacture the boards that power AI data centers.

Explore HDI Services ›

Specification Summary for AI Accelerator PCBs

ParameterTypical RequirementAtlasPCB Capability
Layer count20-30Up to 30
HDI structure2+N+2 to 4+N+4Up to 5+N+5
Trace/space3/3mil (75um)3/3mil minimum
Copper weight (inner)2-3ozUp to 3oz
Copper weight (outer)2-5ozUp to 5oz
Min laser drill0.075mm0.075mm
Aspect ratio14-16:1Up to 16:1
MaterialMegtron 6/7Megtron 6 available
Backdrilling+/-3mil stub+/-3mil accuracy
Board size300x300mm+Up to 1200x600mm
Surface finishENIGENIG, immersion silver
Impedance tolerance+/-5%+/-5% with TDR coupon

ATLASPCB

Building AI Hardware? Let's Talk Specifications.

Upload your stackup or design files. Our engineering team has experience with high-layer-count AI boards and can provide detailed DFM feedback within 48 hours.

Upload Design Files ›

Related Reading:

About AtlasPCB — We specialize in complex PCB manufacturing for HDI, RF, and high-reliability applications. Explore our HDI PCB manufacturing capabilities, or get an aluminum and metal-core PCB services . Every order includes free engineering review. Get your quote.

Reviewed by AtlasPCB Engineering Team — IPC-certified manufacturing specialists with 15+ years of production experience in HDI, RF, and high-reliability PCB fabrication. Content based on factory floor data and real customer design reviews.

Frequently Asked Questions

How many layers does a typical AI accelerator PCB need?
Modern AI accelerator boards (GPU/TPU carrier boards) typically require 20-30 layers. The layer count is driven by: (1) Power delivery requiring 4-8 dedicated power/ground plane pairs to carry 500A+ at low voltage, (2) High-speed signal routing at 112G PAM4 needing 6-12 signal layers with adjacent reference planes, (3) HDI buildup structure (2+N+2 or 3+N+3) to escape 2500+ pin BGAs at 0.65-0.8mm pitch, and (4) Thermal plane layers dedicated to heat spreading. Nvidia's H100/B200 carrier boards are reportedly 24-28 layers.
What PCB materials are needed for 112G PAM4 signaling in AI hardware?
112G PAM4 (used in PCIe Gen6, 800G Ethernet) requires ultra-low-loss laminates with Df below 0.003 at frequencies above 28 GHz. Megtron 6 (Panasonic, Df=0.002) is the minimum viable material; Megtron 7 (Df=0.001) is preferred for longer traces. Standard FR-4 (Df=0.02) creates unacceptable insertion loss at these frequencies—a 6-inch trace on FR-4 at 56 GHz loses 15+ dB, vs 4-5 dB on Megtron 6. The material cost premium is 80-200% over FR-4 but is non-negotiable for signal integrity at these speeds.
How do you handle 400W+ thermal dissipation on a PCB?
PCB-level thermal management for high-TDP AI chips uses: (1) Dense thermal via arrays directly under the die shadow—typically 1000-2000 vias in a grid pattern, plated with copper fill for maximum thermal conductivity. (2) Heavy copper internal planes (3-5oz) acting as heat spreaders. (3) Coin insertion or embedded heat pipes in critical areas. (4) Direct bonded copper (DBC) or insulated metal substrate (IMS) for the immediate die-attach area when applicable. The PCB alone can conduct 10-30W from the die to the board edges; the majority of heat must be removed by heatsink/vapor chamber attached from above.
What aspect ratio challenges exist in AI accelerator PCBs?
A 24-layer board at 3.5mm total thickness with 0.2mm vias creates an aspect ratio of 17.5:1—exceeding most fabricators' standard capability of 12:1 or even advanced capability of 16:1. Solutions include: (1) Increasing board thickness to use larger via diameters, (2) Back-drilling to reduce effective via depth, (3) Using sequential lamination with stacked microvias instead of full-depth through-vias, (4) Specifying 0.25mm minimum drill diameter for through-vias. AtlasPCB supports up to 16:1 aspect ratio with controlled drill accuracy.
  • AI hardware
  • PCB design
  • GPU
  • accelerator
  • HDI
  • power delivery
  • thermal management
  • high layer count
Share:

Related Posts

View All Posts »
Edge AI PCB Design

Edge AI PCB Design

Explore the unique PCB design challenges of edge AI hardware — from thermal via arrays managing 25W NPU chips in constrained form factors to DDR5/LPDDR5X routing and multi-rail power delivery networks for 45-80 TOPS AI accelerators.

PCB Design Requirements for Edge AI Inference

PCB Design Requirements for Edge AI Inference

Engineering guide covering PCB design requirements for edge AI inference modules — from NVIDIA Jetson and Google TPU edge boards to custom ASIC deployments. Covers power delivery network design, thermal via arrays, high-speed memory interfaces, and compact HDI stackups for AI at the edge.

Thermal Management PCB Design for AI Accelerator

Thermal Management PCB Design for AI Accelerator

Design guide for PCBs carrying AI GPUs and ASICs with 300-700W TDP. Covers thermal via arrays, power delivery network design, material selection for high-temperature operation, and heatsink interface optimization.

AI Accelerator PCB Design

AI Accelerator PCB Design

Engineering guide for PCB design supporting AI accelerators (GPUs, TPUs, NPUs). Covers thermal via arrays, high-current power delivery networks, and HDI routing for HBM interfaces. Practical design rules for 800W+ AI compute boards.