224Gbps AI Server PCB: From Signal Collapse to Stable Channel

1. The Reality We See in Projects

At 224Gbps PAM4, PCB is no longer just a carrier.

In most of the AI server projects we’ve been involved in over the past two years, the real bottleneck is not the chip, but the board itself.

Typical failure points are very consistent:

• Via stubs introducing reflection
• Material loss exceeding channel budget
• Stack-up inconsistency affecting impedance
• Manufacturing variation between batches

At this speed, small issues are no longer “minor.” They directly determine whether the system works or not.

2. The Challenge

This project came from an AI HPC customer developing a next-generation GPU-based system.

Their initial prototype had already gone through design iteration, but failed during signal validation.

What we saw on the board:

• Eye diagram closing at 224Gbps
• Insertion loss beyond acceptable range
• Strong reflections caused by via stubs
• Impedance drift across layers in hybrid stack-up
• Risk of reliability issues under thermal cycling

In short, the system could not pass SI validation.

3. What We Actually Changed

This was not a simple fabrication job.

We worked together with the customer from a channel perspective and focused on what actually affects signal behavior.

1. Channel-Level Adjustment 
The total channel length was around 14 inches. By re-evaluating the structure and loss distribution, we reduced insertion loss to a level that could support 224Gbps operation. Return loss improved noticeably, and the eye diagram reopened with sufficient margin.

2. Material System 
We moved to a Megtron 7 based structure with low-profile copper. This was mainly to reduce both dielectric loss and conductor loss, especially over longer routing paths. The difference was not theoretical — it showed clearly in the channel performance.

3. Backdrilling 
This was one of the key turning points. We controlled residual stub length below 8 mil, with tight depth control during backdrilling. All high-speed signal vias were included. After this step, reflection-related issues were significantly reduced.

4. Stack-Up Optimization 
The board used a 24-layer hybrid stack-up. We rebalanced the structure to ensure: dedicated low-loss signal layers, continuous reference planes, symmetry in lamination. This helped stabilize impedance across the entire routing path.

5. SI Collaboration 
We did not treat this as “build to Gerber.” We worked with the customer’s SI team to review channel behavior before fabrication and validate after layout. That made a big difference — problems were addressed before they became physical defects.

6. Manufacturing Consistency 
For this type of board, consistency matters as much as design. We focused on: keeping impedance within ±5%, controlling lamination quality (void-free), using materials with good CAF resistance. This ensured the results were repeatable, not just a one-time success.

4. Before and After

Before optimization:

• Eye diagram was closed
• Reflection was strong
• Channel loss exceeded limit
• Signal behavior was unstable

After optimization:

• Eye diagram clean at 224Gbps
• Reflection significantly reduced
• Loss brought within channel budget
• Stable transmission across full channel

This was not tuning — it was recovery of the signal path.

5. Where This Applies

This type of solution is not limited to one project. We are seeing the same requirements in:

• AI server interconnect boards
• GPU cluster backplanes
• High-speed switching platforms
• 112G to 224G upgrade designs

If your system is moving into this range, these issues will show up sooner or later.

6. What This Project Proves

At 224Gbps, PCB is no longer a passive part of the system. It directly defines whether the system can pass validation.

And solving it is not about one single factor. It comes from the combination of:

• Material selection
• Stack-up design
• Signal path control
• Manufacturing execution

7. Final Thought

Most manufacturers can build a PCB. But when the signal starts to fail, very few can actually fix it.

ULTRONIU Electronics 
From PCB to System — we focus on what makes the signal work.