2026-11-02
Per-unit microlens inspection cadence varies by an order of magnitude across systems in production u
2026-10-27
Wafer-level silicon lens production has lowered front-end cost, but device-side delivery now hinges
A silicon photonics module team recently ramped an 800G project. Wafer-level silicon lens yield was healthy. Yet delivery cadence on the module side never aligned. After dicing, lenses queued for inspection, then queued again for sorting. At the module factory, they were re-inspected on arrival. Turnaround stretched at every handoff.
The team lead put it plainly: “Wafer processing has brought cost to a reasonable level. But whether the device side can deliver reliably—that bottleneck is not at the wafer fab. It is in back-end inspection.”

For optical module and silicon photonics module makers, delivering silicon lenses looks like a volume question. In practice, three hard constraints must hold simultaneously.
Yield constraint. Lot-to-lot pass rates must remain consistently high. Any batch-to-batch swing disrupts module build planning and customer commitments.
Consistency constraint. Surface form PV/RMS and roughness Ra must fall within a narrow window—both across an array and between arrays. Outside that window, module coupling efficiency fluctuates. Industry data shows that in 400G/800G modules, a curvature radius deviation exceeding just 0.1% of design value can reduce coupling efficiency by more than 5%. A single out-of-spec lens that enters a module forces rework across the entire optical path.
Delivery constraint. Data center traffic continues to rise sharply, and the 400G-to-800G transition is moving fast. In 2026, 800G optical modules are running lead times of 40+ weeks, with some allocation queues extending toward 2028. When lens arrival slips by a week, module delivery slips by a week.
These three constraints compound. What the device side needs is not occasional inspection data. It needs every batch—and every lens—backed by data, delivered on a predictable cadence.
Southeast Asia is where these constraints are now colliding. Johor, Malaysia has become a volume manufacturing node for 800G, 1.6T, and silicon photonics transceivers. PRO3C, a subsidiary of Taiwan-listed FICG, entered mass production at its Senai facility in the first quarter of 2026 and is moving toward full-scale output in the second half, with eight SMT lines already operating. The site is explicitly structured to handle manufacturing, mass production, and delivery, while a planned Singapore centre handles innovation and supply chain engagement. As this Johor–Singapore corridor scales, the inspection handoff between wafer-level lens supply and module-level assembly becomes a regional delivery variable—not a local factory detail.
Wafer processing has lowered production cost and raised front-end efficiency. But when the device side absorbs that output, three new challenges emerge.
1. Post-dicing unit counts surge; inspection capacity falls short.
A single wafer yields tens of thousands of lenses after dicing. The wafer fab provides a lot-level pass/fail conclusion. The device side needs a per-lens disposition—every unit, individually. Inspection workload jumps by an order of magnitude compared with the small-die, small-batch era.
2. Array consistency demands per-lens verification.
Wafer-level processes inherently produce high array density. Surface form variation between adjacent lenses must be confirmed lens by lens. Sampling is no longer accepted by module factories. A sample that passes does not prove lens number 1,001 will pass—especially when the consistency window has compressed from 5 μm PV in 2015 to 0.5 μm in the 800G era.
3. Delivery windows have compressed.
The available delivery time for 800G silicon photonic modules keeps shrinking. Inspection time at the lens stage consumes module-side capacity directly. If inspection is a queuing step, it becomes a delivery tax on every downstream commitment.
For the device side to achieve stable volume delivery, inspection must stop being a queue and start functioning as a stabilizer: sufficient cadence, complete data, and responsive service.
Breaking the device-side delivery chain into discrete steps shows where inspection determines outcomes.
| Delivery Step | Key Action | Requirement on Inspection Equipment |
|---|---|---|
| Post-dicing sorting | Per-lens surface form PV/RMS and Ra | Second-level cadence; continuous operation |
| Array consistency confirmation | Within-array and array-to-array comparison | Full-field 3D point cloud; minimal stitching error |
| Factory full-inspection report | Per-lens data archiving and traceability | Data management; automatic report generation |
| Exception batch traceback | Batch or serial-number回溯 | 48-hour on-site response; 72-hour corrective plan |
| Spare parts and maintenance | No line stoppage while waiting for parts | Local spare parts delivery in 3–5 days |
In this chain, inspection equipment cadence determines throughput for the first three steps. Service response determines risk coverage for the last two.
The required inspection paradigm is now well defined. Moving from sampling to full inspection is not simply “measuring more.” It imposes three hard requirements that did not exist before: per-unit cadence compressed to the second range; continuous 24-hour operation without drift or mechanical wear; and per-lens data traceability with automatic archiving—not manual transcription. Under sampling, inspection was an after-the-fact gate. Under full inspection, it becomes an in-line mandatory step: measure immediately after completion, sort immediately after measurement, isolate non-conforming units on the spot.

For module makers operating in the Johor–Singapore corridor, the inspection decision is not a laboratory preference. It is a delivery-capability decision.
The MCZX Qiuhou R series (Mingcha Zhixin (Shenzhen) Technology Co., Ltd.) is designed around this exact production constraint. It uses reflective laser interferometric non-scanning imaging: a single exposure acquires the full-field 3D point cloud without Z-axis mechanical reciprocation.
Relevant production parameters:
Per-lens cadence: approximately 1 second; 10,000 lenses in approximately 3 hours
Vertical resolution: 0.5 nm
Stitching error: <0.5 μm, supporting array-level consistency verification
Built-in stitching algorithms and data archiving for per-lens traceability
48-hour on-site response; corrective plan within 72 hours; local spare parts delivery in 3–5 days
Note: Accuracy and cadence data are based on specific samples and test conditions. Actual results vary by sample and application environment.
For a device maker, the practical calculation is straightforward. A metrology tool that sustains one-second cadence and continuous operation turns wafer-level capacity into module-level delivery. A scanning-based tool that requires tens of seconds per lens turns the same capacity into a sorting backlog.
The urgency calculation has also shifted. For a manufacturer already committed to wafer-level volume, replacing an inspection tool with one that has sufficient cadence—and can run continuously across a production day—may now be more urgent than adding another dicing line.
Singapore and Malaysia are scaling silicon photonics manufacturing on the assumption that high-volume delivery is achievable. That assumption depends, in part, on whether back-end inspection can keep pace. Equipment that combines second-level cadence, per-lens traceability, and regional service response is no longer a quality accessory. It is part of the delivery infrastructure.
MCZX focuses on industrial vision inspection. The Qiuhou R series uses reflective laser interferometric non-scanning imaging, acquiring full-field 3D point clouds in a single exposure with 0.5 nm vertical resolution, approximately 1-second per-lens cadence, and 48-hour on-site response. Free verification testing is available by appointment.