Four-stage HDI means four build-up cycles on each side, four lamination events, and four chances for a laser via to miss the pad below. Here is where the yield goes, and which process controls actually hold it.

What Four Stages Means in Practice
A 4-stage HDI build is usually written as 4+N+4: four build-up layers on top, four on the bottom, and a conventional core in the middle. Each build-up layer is added in its own lamination cycle, then laser drilled, desmeared, filled, plated and imaged. A 20-layer 4+12+4 board therefore goes through the press eight times in total counting the core laminations, and every one of those cycles adds registration error, resin flow and handling risk.
The alternative, a single-stage or two-stage HDI, cannot escape a fine-pitch ball field with the routing density that a large processor package demands. Four stages are chosen when the fan-out needs it, not because more stages are better.
Stacked Versus Staggered
Staggered vias land on the layer below but not directly on top of the previous via; stacked vias land pad-on-pad, so the connection runs vertically through several build-up layers. Stacked structures save routing space, which is why high-density escapes use them, but they demand two things that staggered builds do not: full copper fill in every via, and planarisation flat enough that the next laser via starts on metal rather than on a dimple.
| Structure | Routing benefit | Process demand |
|---|---|---|
| Staggered | Moderate | Standard fill, relaxed planarity |
| Stacked | High, saves a layer per escape tier | Void-free fill, ≤ 5 µm dimple, tight registration |
| Stacked with skip via | Highest | Fill plus controlled depth drilling |
Registration: The Error Accumulates
Each build-up cycle registers its laser vias to the copper features below. If a stage is off by 25 µm, the next stage is drilled relative to a surface that has already moved, and by the fourth cycle the accumulated offset can be 75 to 100 µm. On a 0.10 mm laser via with a 0.25 mm target pad, that is the difference between a reliable connection and a via that clips the pad edge.
Three controls keep this in range. First, X-ray target drilling after each lamination, so every cycle is drilled from a datum that is measured rather than assumed. Second, per-stage scaling compensation: laminate shrinks differently in warp and weft, and the artwork for each stage is scaled to the measured shrink of that stage, not to a generic value. Third, registration measurement on production panels, not only on the first article. A coupon that passes does not prove the panel did.
Via Fill and Planarisation
Stacked vias need solid copper, not a plated barrel with a hole in the middle. The fill process pushes conductive paste or copper into the via, cures it, then grinds the surface flat. Two failure modes matter:
Voids in the fill. A void under the next via becomes a stress concentration. It may pass electrical test and fail in thermal cycling, which is the worst kind of failure because it escapes to the customer. Cross-section at the start of every lot and X-ray on production panels is the only practical detection.
Dimples and over-grinding. A dimple deeper than about 5 µm leaves a gap under the next via; over-grinding thins the copper and changes the impedance of the layer. Both are process-window problems, and both are measured on the same coupon as the fill cross-section.
Laser Drilling the Small Via
A 0.10 mm via in 60 to 100 µm of dielectric is a shallow hole with a taper. The energy has to be enough to reach the target pad cleanly without damaging it, and the target pad has to be large enough to absorb alignment error. The practical rule is a capture pad at least 2.5 times the via diameter, with the pad tied to the plane below by spokes if it sits in a ground plane.
Desmear matters more than it looks. Resin smear left in the bottom of the via becomes a resistive interface after plating, which shows up as an intermittent rather than an open. Permanganate desmear followed by a plasma pass is standard for this class.
Lamination and Resin Flow
Each build-up cycle presses prepreg over a surface that already has copper features and filled vias. Resin has to fill the spaces between traces without starving the bond to the layer below, and it has to do that on a board that is already 2 to 3 mm thick. Thin cores and thick builds pull in opposite directions: the core wants low pressure, the build-up wants flow.
CTE mismatch between the build-up dielectric and the core also matters. On a 20-layer board that sees reflow four times during assembly, delamination at the interface shows up as a blister after the second or third pass, not the first.
Where the Yield Goes
| Defect | Typical cause | Detection |
|---|---|---|
| Via-to-pad misregistration | Accumulated stage error, shrink compensation | X-ray, cross-section |
| Void in stacked via | Fill process, cure profile | Cross-section, X-ray |
| Dimple under next via | Planarisation depth | Cross-section, surface profilometry |
| Resin void between traces | Press profile, prepreg flow | Cross-section, C-SAM |
| Pad damage from laser | Energy, target pad size | Cross-section, electrical test |
| Delamination after reflow | CTE mismatch, moisture | Thermal cycling, C-SAM |
Process Controls That Actually Hold
Per-stage scaling. Measure shrink per lamination, scale the next artwork accordingly. Generic compensation is not enough at four stages.
Cross-section at lot start. One panel per lot, every stack, cut through the via field. It is the only way to see fill voids and dimples before they multiply.
Thermal cycling on qualification lots. 500 cycles from −40°C to 125°C, then cross-section again. Failures that electrical test cannot see appear here.
Traceability per press cycle. Which press, which program, which prepreg lot. When a lot fails, the question is which cycle moved, and that cannot be answered without records.
What to Send With the Design
The stackup drawing with the build-up structure, the via types per layer pair, the impedance targets, and the escape requirement at the processor package. From those, the question of whether four stages are needed at all, and where the yield risk sits, can be answered before the first panel is pressed.
Related: 20-layer 4-stage HDI PCB for AI accelerator cards, 12-layer edge AI module PCB and 112G PAM4 insertion loss.