· Sophia Reyes · Engineering  · 7 min read

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.

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.

Quick Answer

Edge AI inference modules demand PCBs that simultaneously handle 50-150W in compact form factors with high-speed memory interfaces (LPDDR5 at 6400 MT/s) and accelerator interconnects. Key requirements: minimum 8-layer HDI stackup, 2oz+ power planes, thermal via arrays under processor (min 0.3mm pitch), controlled impedance for memory at 100 ohm differential, and strategic power plane partitioning to isolate switching noise from sensitive analog circuits.

Quick Reference: Edge AI PCB Requirements Summary

ParameterTypical Edge AIHigh-Performance Edge
Layer count8-1212-16
Power dissipation15-50W50-150W
Copper weight (power)2 oz3-5 oz
Min trace/space4/4 mil3/3 mil
Via typeMicrovia + throughStacked microvia HDI
Memory interfaceLPDDR4X/LPDDR5LPDDR5/LPDDR5X
Accelerator I/OPCIe Gen4/Gen5PCIe Gen5 + custom
Thermal solutionVia array + heatsinkEmbedded Cu + vapor chamber
Board size70x70mm to 100x100mm100x100mm to 170x170mm

Why Edge AI Creates Unique PCB Challenges

Cloud AI spreads 300-700W across large server boards (300x400mm+) with unlimited cooling budgets. Edge AI packs 50-150W into modules smaller than a credit card. This density creates three simultaneous engineering constraints:

  1. Thermal density: 10-30 W/cm2 concentrated under a single package with limited airflow
  2. Power delivery: 50-150A at 0.7-1.0V core voltage through short, low-impedance paths
  3. Signal integrity: LPDDR5 at 6400 MT/s and PCIe Gen5 at 32 GT/s in millimeter-scale routing lengths

All three constraints compete for the same stackup real estate. More power planes mean fewer signal layers. More thermal vias mean less routing area. Getting this balance right is the central challenge of edge AI PCB design.

AI HARDWARE PCB EXPERTISE

Designing an Edge AI Module?

AtlasPCB has fabricated boards for AI inference deployments requiring HDI, heavy copper, and controlled impedance. Up to 30 layers with 3/3 mil trace/space.

Get Engineering Review ›

Thermal Management: Via Arrays and Beyond

Thermal Via Array Design

For AI processors dissipating 50-150W through an exposed thermal pad:

Design ParameterMinimumRecommendedAggressive
Via drill diameter0.3 mm0.3 mm0.25 mm
Via pitch1.0 mm0.6 mm0.5 mm
Via fillTentedNon-conductive fillConductive fill
Thermal resistance (array)15 C/W8 C/W4 C/W

Design rules:

  • Fill all thermal vias (IPC-4761 Type V or VII) to prevent solder wicking during assembly
  • Place vias on same grid as solder paste stencil openings to avoid paste bridging
  • Connect via bottoms to a continuous copper pour on the bottom layer (heat spreader interface)
  • Minimum copper connection from via to ground/power pour: full flood (no thermal relief on thermal vias)

When Via Arrays Are Not Enough

Above 100W in compact form factors (under 100x100mm), via arrays alone cannot achieve sufficient thermal resistance. Additional solutions:

  • Embedded copper coin: 2-4mm thick copper slug press-fit into a milled cavity. Thermal resistance: 1-2 C/W. Adds $30-80/board.
  • Metal-core layer: Aluminum or copper core layer in the stackup center. Good for uniform heat spreading across the board area.
  • Direct die attach: Skips the package thermal interface — die bonded directly to board-level heat spreader. Requires bare die assembly capability.

Power Delivery Network Design

The Edge AI PDN Challenge

A typical edge AI ASIC draws:

  • Core: 50-100A at 0.75V (VDD)
  • I/O: 10-20A at 1.2V (VDDIO)
  • Memory: 5-10A at 1.1V (VDDQ)

At these currents, every milliohm of PDN impedance translates to millivolts of droop. With 5% tolerance on 0.75V core supply, total PDN impedance budget is just 0.375 mOhm — including VRM, decoupling, and PCB planes combined.

PDN Stackup Strategy

For a 12-layer edge AI module:

LayerFunctionCu WeightNotes
L1Signal + component1 ozBGA fanout, passive placement
L2Ground (ref)1 ozContinuous pour, signal reference
L3Signal (memory)0.5 ozLPDDR5 byte lanes
L4Power (VDD core)2 ozSolid pour, max area
L5Ground2 ozPower return, thermal spreading
L6Power (VDDIO)2 ozI/O and memory power
L7Ground2 ozIsolation between domains
L8Power (VDDQ)2 ozMemory power, separate domain
L9Ground1 ozSignal reference
L10Signal0.5 ozLow-speed I/O, misc routing
L11Ground1 ozBottom reference
L12Signal + component1 ozBottom-side passives, connectors

Key principles:

  • Every signal layer immediately adjacent to a ground plane (no signal-signal adjacent)
  • Power planes sized for current density: min 2oz for 50A+ total
  • Separate ground plane between analog and digital power domains
  • Decoupling capacitor placement: first tier (0201) within 1mm of BGA power balls, second tier (0402) within 5mm

HEAVY COPPER + HDI

Up to 5oz Copper with 3/3 mil HDI

AtlasPCB combines heavy copper power planes with fine-line HDI signal routing in the same stackup — exactly what AI modules demand.

View Full Capabilities ›

Signal Integrity: Memory and High-Speed Interfaces

LPDDR5/5X Routing Requirements

ParameterLPDDR5 (6400 MT/s)LPDDR5X (8533 MT/s)
Impedance (single-ended)40-50 ohm40-50 ohm
Impedance (differential)80-100 ohm80-100 ohm
Max trace length25 mm20 mm
Length matching (intra-byte)+/- 0.5 mm+/- 0.25 mm
Length matching (inter-byte)+/- 2 mm+/- 1 mm
Routing layerStripline (preferred)Stripline (mandatory)
Crosstalk budget< -30 dB< -35 dB
Dielectric (Dk)Standard FR-4 OKLow-Dk preferred

PCIe Gen5 Routing

For the host interface (typically PCIe Gen5 x8 or x16 to the edge AI module):

  • 85 ohm differential impedance (+/- 10%)
  • Maximum trace length: 200mm (connector-to-connector)
  • Pair-to-pair skew: < 5 mm
  • Via stub: < 8 mil (backdrilling required for through-hole vias)
  • Loss budget: < 8 dB at 16 GHz Nyquist

HDI Strategy for BGA Fanout

Modern edge AI processors use 0.5-0.65mm pitch BGAs with 600-2000+ balls. Fanout requires:

  • Layer 1-2 microvias for signal escape from inner BGA rows
  • Stacked microvias (2-3 levels) for power/ground connections to inner planes
  • Via-in-pad on all BGA pads (IPC-4761 Type VII)
  • 3/3 mil trace/space minimum for signal routing between via pads

Typical HDI structure: 2+N+2 or 3+N+3 buildup with 0.1mm laser-drilled microvias.

HDI FOR AI

5+N+5 HDI with Stacked Microvias

Our HDI process supports the fine-pitch BGA fanout that edge AI processors require — laser drill to 0.075mm, via-in-pad, and stacked configurations.

Explore HDI Options ›

Material Selection for Edge AI Boards

Application TierMaterialDk (10 GHz)Df (10 GHz)TgUse Case
StandardHigh-Tg FR-4 (Tg170+)4.20.018170CLPDDR4, PCIe Gen4
Mid-rangeMegtron 4 / Tachyon 60G3.80.008200CLPDDR5, PCIe Gen5
High-endMegtron 6 / Tachyon 100G3.60.004210CLPDDR5X, 112G SerDes

Practical guidance:

  • For designs under 10 Gbps per lane: standard high-Tg FR-4 is sufficient and cheapest
  • For 10-25 Gbps (PCIe Gen5, LPDDR5): mid-range materials provide adequate margin
  • For 25+ Gbps (112G SerDes, LPDDR5X): high-end materials are mandatory

Most edge AI modules in 2026 operate at the mid-range tier — LPDDR5 memory and PCIe Gen5 host interface are well served by Megtron 4 class materials without needing ultra-premium laminates.


Design Checklist: Edge AI Module PCB

  1. Stackup: minimum 8 layers; 12+ for complex designs. Every signal layer referenced to ground.
  2. Power planes: 2oz minimum for core voltage. Separate domains for VDD, VDDIO, VDDQ.
  3. Thermal vias: Filled array under processor thermal pad, 0.3mm drill, 0.6mm pitch minimum.
  4. HDI: Via-in-pad for BGA, 2+N+2 minimum buildup for 0.5mm pitch packages.
  5. Memory routing: Stripline on layers adjacent to ground. Length-matched within +/- 0.5mm (byte).
  6. PCIe routing: 85 ohm differential, backdrilled vias if through-hole used.
  7. Decoupling: 0201 capacitors within 1mm of power balls, bulk capacitors within 5mm.
  8. Board thickness: 1.6-2.0mm typical. Thicker boards need backdrill consideration.
  9. Surface finish: ENIG for BGA reliability; consider Immersion Silver for loss-sensitive pads.
  10. Testing: Impedance coupon, thermal cycling validation per IPC-6012 Class 3.

ATLASPCB

Ready to Build Your Edge AI Board?

From stackup consultation to final delivery — we support the full complexity of AI hardware PCBs. Upload your design for engineering review and quote.

Get Instant Quote ›

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 PCB layers do edge AI modules need?
Minimum 8 layers for basic designs with one accelerator. Complex multi-chip modules need 12-16 layers. Typical stackup: 2 signal + 2 power + 2 ground + 2 signal, with HDI microvias for BGA fanout.
What thermal via design works for AI chips?
Use filled thermal via arrays with 0.3mm drill on 0.6mm pitch directly under the chip thermal pad. For 100W+ dissipation, combine via arrays with metal-core inserts or embedded copper coins for thermal resistance below 5 C/W.
What copper weight is needed for AI inference PCBs?
2oz copper minimum on power planes for 50W designs, 3oz for 100W+. Inner ground planes at 1oz minimum. Outer layers typically 1oz with local thickening at high-current paths.
How do you route LPDDR5 for edge AI?
LPDDR5 at 6400 MT/s requires 100 ohm differential impedance, max 15mm trace length matching within byte groups, and stripline routing on adjacent layers to ground planes. Keep memory traces on layers immediately adjacent to the processor BGA layer to minimize via stubs.
  • edge AI
  • PCB design
  • thermal management
  • power delivery
  • HDI
Share:

Related Posts

View All Posts »
Edge AI PCB Design in 2026

Edge AI PCB Design in 2026

The inference flip is here — two-thirds of AI compute now runs at the edge. Learn how to design PCBs for edge AI processors including NPU power delivery networks, thermal via arrays, LPDDR5X routing, and HDI stackup strategies that ship reliable products.

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 for Edge AI Inference Modules

PCB Design for Edge AI Inference Modules

Complete design guide for edge AI inference module PCBs — covering HDI stackup requirements for 0.4-0.65mm BGA escape, thermal via arrays for 5-25W dissipation in compact form factors, DDR4/5 and high-speed SerDes routing, and material selection that balances performance with cost.

PCB Design for AI Edge Inference

PCB Design for AI Edge Inference

Engineering guide for PCB designers building AI edge inference hardware. Covers thermal management for NPUs, power delivery network design for transient AI workloads, and high-speed memory interface routing for edge ML accelerators.