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

As optical engines move closer to expensive compute chips, testing them before assembly becomes a way to protect the value of the whole package. The challenge is that optical and electrical behavior increasingly share one signal path, while the cost of testing an optical component is already approaching its manufacturing cost. For co-packaged optics (CPO), the industry must improve fault isolation and production methods at the same time.

That changes the economics of test. A test plan that seems expensive when judged against a small photonics device may be reasonable when it prevents that device from damaging or wasting a much more costly accelerator or switch package. But the route to high-volume manufacturing will take more than adding test steps: teams need repeatable measurement boundaries, useful diagnostic access, and agreement across photonics and semiconductor manufacturing groups.

Define the measurement boundary before chasing a fault

Finding the source of a failure starts with deciding where each measurement begins and ends. Kazuo Yamaguchi of Keysight Technologies described these reference planes as physical, electrical, and optical boundaries. Depending on the test, a boundary might sit at a die pad, package ball, connector, or fiber interface. Engineers then need to account for everything between those planes, including probes, sockets, traces, connectors, and patch cords.

This matters because measurements taken at different stages are only useful for comparison when their boundaries line up. If test insertions use different reference planes, a change in the result may come from the device, the test setup, or the equipment between them. De-embedding those intervening elements helps teams separate their contribution. When part of the system is absent, a characterized substitute can stand in for it. A known laser source, for example, can help test an optical engine on its own.

The diagnostic goal is to isolate components without losing sight of how they behave together. Designers and failure-analysis teams need enough access to tell which element is responsible, then enough system-level testing to establish that the assembled path works.

Loopback helps locate problems, but it needs diagnostic access

Some customers are adding monitor photodiodes to optical paths so they can trace unexpected transmit or receive results. Others want taps in loopback paths, allowing them to route signals to characterization equipment or monitor optical power. That access can make a difference when a device gets a functional check or mission-mode test instead of a full characterization.

A loopback test can show that a path is failing, but the result alone may not explain why. A failed bit error rate or other functional check could point to a problem at several places in the path. Without a way to observe or tap the signal, engineers may have little evidence about the cause. Diagnostic access can narrow the search, although it adds design and test considerations that teams must weigh against the desired level of coverage.

The status of the laser source also needs careful treatment. Ira Leventhal of Advantest America argued that the test objective is the device under test, with the laser serving a role closer to a device power supply than to a special reference instrument. Customers may want to use the same laser source as in normal operation, but that source may not be an automated test instrument. The distinction matters: a reference component can help isolate the device, while a production setup must also account for the behavior of the source actually used in operation.

In failure analysis, teams need to identify which parts of the setup belong in the list of possible causes and which are treated as known-good. That judgment depends on consistent characterization and a clear account of what the test includes.

Integration makes some NPO links as hard to diagnose as CPO

Comparisons between near-packaged optics (NPO) and CPO can be slippery because the terms do not have one consistent definition. Both can require aligning multiple optical connectors and supplying power to the device. One proposed distinction is whether the optical engine can be detached. If it can, a failed engine may be removed, but that doesn’t eliminate the value of testing it before attachment: reworking and reconnecting a package downstream can be difficult.

CPO integrates more components into the package, making it important to decouple their behavior during test. Design-for-test features can expose information about individual pieces and help teams connect component measurements with failures in the assembled package. Without that visibility, a fault in one element can be difficult to distinguish from a problem elsewhere.

The architecture of an NPO link also affects how easy it is to diagnose. A retimed NPO has its own digital signal processor (DSP), so the optical engine can be tested on its own in a manner similar to a pluggable module. A linear-drive NPO removes that DSP: the host ASIC’s SerDes drives the optics through linear drivers and transimpedance amplifiers. Electrical and optical behavior then form one path, so a link failure is harder to attribute to either side. In that respect, its test complexity resembles CPO.

Thermal behavior adds another concern. Dave Armstrong of ISE Labs pointed to the power dissipation of CPO and the limited distance electrical signals can travel. NPO may allow greater electrical reach, but that reach brings its own management needs. With multiple optical components and a complex package, temperature must remain part of the test and design discussion.

Test spending shifts toward protecting the package

Leventhal framed the economic tradeoff with a potential $40,000 or $50,000 GPU or ASIC: an inexpensive optical engine failure should not take down that package. In the usual approach, test spending is considered as a share of a chip’s manufacturing cost. For a package whose value depends on many integrated components, that calculation changes. More testing early in the production ramp may be justified if it helps avoid scrapping the costly compute device.

Yamaguchi sees the same shift in the rising value of the device, the cost of scrap, and the complexity of both testing and packaging. That combination makes wafer-level or optical-engine-level screening more attractive before assembly. It moves some effort upstream from package test, where a fault can be more expensive to diagnose and recover from.

There is still a need to bring test costs down. Armstrong said surveys conducted through the IPSR put optical-component test costs close to manufacturing costs, with a possibility that testing could cost more as complexity rises. His point is that known-good optical components have value, especially when they are paired with much more expensive compute devices. The budget cannot be assessed by looking at the optical engine alone.

Scaling depends on shared production choices

The panel’s assessment is that demonstrating CPO can work is no longer the central question. Leventhal cited reliability data presented by Meta earlier in the year and said the harder task is scaling the technology into high-volume manufacturing. For the near term, manufacturers may rely on added screening, burn-in, more test insertions, and tighter engineering controls while production practices mature. NPO can provide a nearer-term route to deployment, with experience from those efforts informing later CPO production.

Standards and multi-source agreements can help align test platforms, components, and manufacturing expectations. They don’t alter the underlying measurement science, but they can clarify shared requirements, including optical power transmission and applied bias. That coordination matters when manufacturers need parts from multiple suppliers and production lines that can operate consistently.

There is also an organizational problem. Matt Griffin of Teradyne described customers bringing together photonics specialists and semiconductor test teams. Photonics groups may know how to characterize devices, while high-volume test groups focus on uptime, software and service support, and deploying many test cells. Choices that improve flexibility, channel count, or serviceability can affect optical performance at the device under test. Those tradeoffs need to be understood by the people designing the device and the people building its production test flow.

The practical direction is clear: establish consistent reference planes, build in enough access to diagnose failures, and screen components before expensive assembly when the economics support it. Whether those methods become affordable at scale will depend on test equipment, package design, and teams making compatible decisions across the production chain.

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

Read the earlier discussions:

• CPO Testing Faces A Moving Target Problem — Test vendors are adapting to changing connector architectures and tight alignment tolerances.
• Scaling CPO Testing Beyond A Single Die — Fiber alignment, connector durability, and combined optical and electrical testing complicate high-volume production.