DEV Community

Cover image for PCB Requirements for AI Accelerator Hardware: What Your Fab Needs to Deliver
AtlasPCBEngineering
AtlasPCBEngineering

Posted on • Originally published at atlaspcb.com

PCB Requirements for AI Accelerator Hardware: What Your Fab Needs to Deliver

AI accelerator PCBs represent the most demanding boards in commercial electronics manufacturing. Every challenging specification converges simultaneously: 20-30 layers, HDI microvias, 5oz copper power delivery, ultra-low-loss laminates, and extreme thermal management.

Here's what these boards actually require—and why they cost $3,000-10,000+ for bare PCB alone.

Why 20-30 Layers is the Starting Point

A modern AI accelerator (think Nvidia B200-class) connects to:

  • 8-12 HBM3e memory stacks (1024+ I/O each 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: I2C, SPI, thermal sensors

Routing this requires:

Layer Type Count Purpose
Signal (high-speed) 8-12 HBM, PCIe, NVLink
Signal (low-speed) 2-4 Control, GPIO
Ground reference 6-8 Adjacent to every signal layer
Power planes 4-8 Multiple rails, high current
Total 20-30 Typical range

Every signal layer needs an adjacent ground reference for impedance control. That alone means 2 layers per routing tier. With 6-8 routing tiers plus 4-8 power planes, 20+ layers becomes the cost-optimized solution—not overengineering.

Power Delivery: 500A at Sub-1V

Standard 1oz copper on a 50mm-wide plane carries ~30A. To deliver 500A:

Copper Weight Current per 50mm Width Planes Needed
1oz (35um) ~30A 17 (impossible)
2oz (70um) ~60A 9
3oz (105um) ~90A 6
5oz (175um) ~150A 4

Practical designs use 3-5oz copper on 4-6 power/ground pairs. The BGA connection uses massive via arrays: 300-500 power pins, each via carrying 1-2A.

PDN impedance target: < 0.5 milliohm DC from VRM to BGA. At these numbers, every microhm counts.

Signal Integrity: 112G PAM4 Material Requirements

112G PAM4 (56 Gbaud, Nyquist at ~28 GHz) is brutally demanding on PCB materials:

Parameter Standard FR-4 Megtron 6 Megtron 7
Df @ 12.5 GHz 0.020 0.002 0.001
Insertion loss (6", 28 GHz) >20 dB 5-6 dB 3-4 dB
Surface roughness 6-8um 2-3um (VLP) 1-2um (HVLP)
Cost vs FR-4 1x 2-3x 3-4x

Above 10 GHz, copper surface roughness dominates loss. HVLP (Hyper Very Low Profile) foil is mandatory—standard RTF adds 3-4 dB/inch at 28 GHz.

FR-4 is completely unusable: a 6-inch trace at 28 GHz loses 20+ dB. The eye diagram is gone.

Backdrilling: The Mandatory Step Above 16 GHz

A through-via in a 24-layer (3.5mm) board creates a stub from signal exit to via end. At 28 GHz, even 0.5mm of stub creates destructive resonance.

Backdrilling removes the unused portion:

  • Target stub: < 200um
  • Drilling accuracy needed: +/-75um (3mil)
  • Required on: ALL high-speed signal vias not terminating at outer layers

Without backdrilling, via stubs create notch filters that kill channel performance above 16 GHz. This is non-negotiable for 112G signaling.

Thermal Via Arrays for 400W+ TDP

With 400W concentrated in a 50x50mm area:

Via Configuration Effective Thermal Conductivity
No thermal vias (FR-4 only) 0.3 W/mK
Via array (0.3mm, 1mm pitch) 3-8 W/mK
Dense array (0.2mm, 0.5mm) 15-25 W/mK
Copper-filled dense array 30-50 W/mK

AI boards use 1000-2000 copper-filled thermal vias under the die shadow, connected to 3-5oz internal planes for lateral heat spreading (copper: 400 W/mK along-plane).

The PCB conducts 10-30W to board edges; the remaining 370W+ exits through heatsink/vapor chamber above. But those 10-30W through-board prevent hotspot formation that would otherwise cause solder joint fatigue.

Manufacturing Yield Reality

Board Type Typical First-Pass Yield
8-layer standard 95-98%
16-layer HDI 88-93%
24-layer AI (Megtron 6) 75-85%
30-layer ELIC + Megtron 7 65-78%

Aspect ratio is the primary yield killer. 24 layers at 3.5mm with 0.2mm vias = 17.5:1—exceeding most fabricators' standard 12:1 capability.

Solutions: 0.25mm minimum drill (14:1), backdrilling to reduce effective depth, or HDI structure to avoid full-depth vias entirely.

The Cost Equation

A typical AI accelerator carrier board (24-layer, 200x200mm, Megtron 6, 5oz Cu, HDI 2+N+2):

  • 10 pieces: $3,000-5,000/unit
  • 100 pieces: $1,200-2,500/unit
  • 1000 pieces: $600-1,200/unit

These numbers reflect the convergence of every expensive PCB feature: premium materials, high layer count, heavy copper, HDI processing, tight tolerances, and low yield.

Summary Spec Table

Parameter AI Accelerator Requirement
Layer count 20-30
HDI structure 2+N+2 to 4+N+4
Trace/space 3/3mil (75um)
Copper (inner) 2-3oz
Copper (outer) 2-5oz
Material Megtron 6/7
Aspect ratio 14-16:1
Backdrilling +/-3mil stub accuracy
Thermal vias 1000+ per BGA footprint
Impedance +/-5% with TDR coupon

We've been seeing exponential growth in AI hardware board inquiries over the past 18 months. The complexity is pushing fabrication capability boundaries—particularly the combination of high layer count + Megtron materials + heavy copper that drives yield challenges. Our HDI production line handles up to 30 layers with 5+N+5 buildup and 5oz copper.

If you're designing boards for AI accelerators, GPU modules, or HPC infrastructure: upload your design files for a detailed DFM review and cost analysis.

Top comments (0)