Co-Packaged Optics Testing: Why Failure Isolation and Early Screening Matter

Testing optical components for co-packaged optics (CPO) can approach the cost of making them, according to Dave Armstrong of ISE Labs. That changes the calculation for manufacturers assembling optical engines with high-value ASICs and GPUs: a test step that looks expensive in isolation may help avoid scrapping a far costlier package. The harder question is how to find faults reliably without letting test time and equipment costs grow alongside system complexity.

That problem brings several engineering concerns together. Test teams need consistent measurement boundaries, ways to isolate individual components, and access to optical signals during diagnosis. They also need to decide when a laser source should be treated as part of the system under test and when it should serve as a known-good input. These tradeoffs came up in a discussion with Matt Griffin of Teradyne, Ira Leventhal of Advantest America, Armstrong, and Kazuo Yamaguchi of Keysight Technologies.

Fig. 1: (L-R): Teradyne’s Griffin; Keysight’s Yamaguchi; ISE Labs’ Armstrong; Advantest’s Leventhal

Failure analysis starts at the measurement boundary

When a link fails, the test result is only useful if engineers know which parts of the signal path it covers. Yamaguchi describes the reference plane as the physical, electrical and optical boundary for each measurement. Depending on the test, that boundary might be at a die pad, package ball, connector or fiber interface. Measurements taken at different stages need to use aligned boundaries, or engineers may confuse changes introduced by the test setup with faults in the device.

Everything between those planes can affect the result. Probes, sockets, traces, connectors and patch cords may need to be de-embedded, with the same boundaries maintained across successive test insertions. To test an optical engine by itself, engineers can replace its neighboring components with characterized substitutes. A known-good laser source can supply the missing optical input, allowing a test to focus on the engine.

That approach depends on drawing a clear line between the component being evaluated and the equipment supporting it. Leventhal argues that the laser source is functionally closer to a device power supply than to the component under test: the test system assumes it works as specified. In practice, customers may prefer the same laser source used during normal operation, but that source may not be an automated test equipment instrument. As test methods mature, teams will need to decide which elements are trusted inputs and which remain possible causes of failure.

The distinction matters because a fault investigation often proceeds by eliminating parts of the path. If boundaries are unclear, or if each test insertion uses a different setup, comparing results becomes less reliable. A known-good substitute can make the diagnosis more focused, provided its characteristics are understood and the test conditions match the intended measurement.

Loopback helps, but it needs diagnostic access

Some manufacturers are adding monitor photodiodes to optical paths so they can trace unexpected transmit or receive results. Other designs provide taps in loopback paths, letting engineers connect characterization instruments or monitor optical power. That can support a simpler loopback or mission-mode check when a full characterization test would be unnecessary.

There is a tradeoff. A system may fail its bit error rate target or show a functional problem, yet a basic loopback result may not reveal whether the cause sits in the transmitter, receiver, or another part of the path. A tap gives engineers a way to inspect signals inside the loop and gather evidence about where performance degrades. The extra access is especially useful when the test is intended to screen devices quickly and cannot measure every characteristic in detail.

For failure analysis, the broader principle is to make components observable and separable. Design-for-test features can provide measurements from individual pieces that engineers can compare with the behavior of the complete package. Without those signals, a failed end-to-end link may identify a problem without narrowing down its source.

NPO and CPO pose different isolation problems

The terms near-packaged optics (NPO) and CPO do not always have consistent definitions, but the approaches share practical test challenges. Both require careful alignment of multiple optical connectors and a way to power the device. One possible distinction is whether an optical engine can be detached from the package. Detachability may make a faulty engine easier to replace, but it does not remove the value of screening it beforehand: rework and reattachment later in manufacturing can be difficult.

CPO places more components together in the package, increasing the need to isolate their behavior during testing. Leventhal points to design-for-test methods that capture information about individual components and help correlate those measurements with failures in the assembled package. That evidence can help a team distinguish a component defect from an interaction introduced by integration.

NPO architectures also vary in how much of the signal path can be tested separately. In retimed NPO, the optical engine has its own digital signal processor (DSP), so engineers can test it for compliance much like a pluggable module. A failure is easier to locate because the optical engine can be evaluated on its own.

Linear-drive NPO removes that DSP. The host ASIC’s SerDes drives the optics through linear drivers and transimpedance amplifiers (TIAs), joining the electrical and optical sections into one signal path. A failed link is then harder to attribute to either section, making diagnosis more like CPO. The architecture’s integration can therefore shift work from isolated component tests to tests that must account for the complete path.

Thermal behavior adds another complication. Armstrong notes that CPO limits how far electrical signals can travel, while NPO can allow greater electrical reach. That reach still needs to be managed, and the heat generated by optical components matters in either design. Test results need to account for operating conditions that could affect performance across the package.

Package value makes more testing easier to justify

Leventhal puts the economic change plainly: an optical engine that costs relatively little can still jeopardize a GPU or ASIC valued at $40,000 or $50,000 if it fails after integration. Traditional rules of thumb that set test spending as a share of a chip’s manufacturing cost may not fit this situation. Manufacturers may accept more testing early in production to avoid losing expensive devices, then use yield learning to refine the test flow as manufacturing scales.

Yamaguchi sees the same pressure in the rising value of devices, scrap, test complexity and package complexity. That can make additional wafer-level or optical-engine-level screening more attractive than relying on tests after packaging. The decision depends on whether early measurements catch failures that would otherwise consume more expensive materials and manufacturing effort.

Griffin adds that these optical links are part of complex AI accelerators and network switches used in data centers. Looking only at the optical engine’s price misses the role it plays in the assembled system, including lower power use and support for disaggregated compute architectures. Manufacturers are therefore likely to keep focusing on how to test these systems efficiently.

Scaling production will require coordinated test choices

Leventhal says the industry has moved beyond the question of whether CPO can work. He cites reliability data presented by Meta earlier in the year and describes the next challenge as reaching high-volume manufacturing. In the near term, manufacturers may use more test steps, screening, burn-in and engineering controls while production methods mature. NPO can provide an intermediate path: its production experience may inform later CPO work, though it does not remove the remaining scale-up challenges.

Shared standards and multi-source components are another part of that work. Yamaguchi says measurement science itself may not change, but common platform definitions can help companies align equipment and manufacturing expectations. Agreed test criteria can also clarify requirements such as transmitted optical power and applied bias. The industry is still developing those approaches as the technology moves quickly.

Production decisions also bring together teams with different priorities. Photonics engineers may know device behavior but have less experience with semiconductor-scale test operations. High-volume test teams focus on uptime, software and service support, and the practical deployment of many test cells. Choices that improve flexibility, serviceability or channel count can affect optical performance at the device under test. Designers and manufacturing teams need to understand those consequences together.

Armstrong says surveys by the IPSR indicate that testing an optical component already approaches its manufacturing cost and could exceed it as complexity rises. The implication is that manufacturers may need to spend more to establish known-good components before assembly, while sharing the work across companies and technical teams. For CPO, the path to lower total costs will depend on whether better isolation and earlier screening prevent expensive package failures without creating a test flow too costly to scale.

Read parts one and two of the discussion:
CPO Testing Faces A Moving Target Problem
Test vendors are adapting to changing connector architectures and extreme alignment tolerances as co-packaged optics heads toward mainstream.

Scaling CPO Testing Beyond A Single Die
Fiber alignment, connector durability, and simultaneous optical and electrical testing pose challenges for high-volume manufacturing.

The post Bringing Down Test Costs For Co-Packaged Optics appeared first on Semiconductor Engineering.