Bottom line first: inspection is not a contest for the widest coverage. AOI sees the board surface and the joints after assembly, X-ray sees the layers and barrels that the surface hides, and electrical test sees the opens and shorts that neither of them can. What actually drives inspection cost is the false-call rate and the way readings are judged — not the number of machines on the floor. Grade defects by inspectability and put numbers on every criterion; that saves more money than adding another full inspection.
1. What each imaging method can actually see
Write the machine limits down and the wrong tool stops being used — and nobody pays for defects the tool cannot see.
- AOI (optical): images the surface and resolves trace width and spacing deviations down to the 50 µm class. It is effective on nicks, foreign material, scratches, mask offset and legend over pads; it is blind to barrel copper thickness and layer-to-layer registration, and it cannot see copper defects covered by solder mask.
- X-ray (transmission): 2D shows registration and hole position, 3D-CT shows BGA joint voids and via fill. Commercial resolution sits in the 5–10 µm class: it flags density anomalies but cannot replace a cross-section for barrel copper.
- Electrical test: opens, shorts and net continuity only come from flying probe or a test fixture; impedance needs its own coupon and TDR run. The three are not interchangeable, and dropping any one leaves a blind spot.
2. False-call rate is a cost line, not a quality metric
A false call does no harm to the board, but it consumes review hours and engineering attention, and it dulls people to real defects.
- Typical level: consumer multilayer lines often run 3000–8000 ppm, higher on high-layer counts and fine lines. Only below 1500 ppm can inspection claim it is not delaying the schedule.
- Labour cost: roughly 3–5 minutes of review per 1000 decision points; one line's daily review load can absorb a full-time position.
- Main sources: one sensitivity setting for everything (trace, nick and foreign material sharing a threshold), light angle mismatched to copper reflection, low contrast on white solder mask, and panel datum drift.
- Countermeasures: set thresholds per defect type; calibrate each lot against a golden board; recalibrate on a part-number or mask-colour change instead of reusing the previous recipe.
3. Fix the X-ray judgement criteria in writing
Most X-ray disputes are not about whether something is visible, but about whether it is out of tolerance. The criteria must name the datum, the number and the way it is recorded.
- Layer registration: ±75 µm is the common criterion against the target; for a 44-layer backplane tighten it to board thickness and layer count, and record the offset vector per layer rather than one combined figure.
- Hole position: hold the hole centre within ±50 µm of true position; a missing barrel can only be confirmed by cross-section — a density anomaly on X-ray is an indicator, not a verdict.
- BGA voids: the usual reading is no more than 25% void area in a single joint (IPC-7095), and never concentrated at the pad interface. Beyond that it is a rework, not a release on the grounds that it looks acceptable.
- CT sampling: five boards from the first lot of a new stackup, one per lot once stable, and back to the first-lot rate after a complaint or a material change.
4. The economic boundary between sampling and full inspection
The inspection plan follows board value × lot size × consequence of failure. There is no single answer for every part number.
- High-value, high-layer: above 20 layers and in small lots, run full AOI plus full electrical test plus CT sampling — still cheaper than one replacement lot.
- High volume, mid layer count: full AOI, AQL sampling on electrical test, cross-section on the first article. Move the money saved from full electrical test into first-article and in-process control.
- Tighten by area: BGA, gold fingers and press-fit holes get their own tightened criteria; everything else stays standard, so the whole panel does not pay for a few zones.
- Traceability: record machine ID, shift and verdict per board, so a complaint can be traced to a lot instead of triggering a full recall.
5. Five moves for design and sourcing
- State the acceptance class on the drawing: IPC-A-600 Class 2 or Class 3, with the IPC-6012 Class 3 items marked separately.
- Supply a golden board or images of known-good product as the shared datum for AOI calibration and dispute comparison.
- Agree in advance which defects are concessionable and which must be scrapped, so judgement does not rest on individual experience.
- List the inspection items at RFQ stage (AOI, X-ray, electrical test, cross-section) and price them separately instead of guessing what is bundled into the unit price.
- Ask for monthly false-call and escape rates as a supplier scorecard item, rather than only raising the question after a problem appears.
AIPCB.cc is operated by Shenzhen Lingxin Technology Co., Ltd. The platform supports up to 44 layers, a maximum processed size of 610×1100 mm, aspect ratios up to 20:1 (25:1 for backplanes), impedance control tolerance of ±8% (±10% standard), trace accuracy of ±8 μm, heavy copper up to 13 OZ, and expedited prototype evaluation. For engineering and quotation, contact info@ipcba.com.cn or +86 138 2310 9230.
Send your Gerber files and stackup requirements; we review feasibility and lead time against your build-up complexity.