Co-Packaged Optics Testing: Scaling Beyond Single-Die Production

Scaling co-packaged optics (CPO) testing means solving a physical access problem as much as a measurement problem. Most production setups still test one site at a time, and aligning fibers with optical engines is a key obstacle to testing several devices in parallel. Connector durability, radio-frequency probing and lengthy measurements add further constraints.

Those issues came up in a discussion with Matt Griffin, senior product manager for optoelectronics test at Teradyne; Ira Leventhal, vice president of research and venture at Advantest America; Dave Armstrong, fellow at ISE Labs, a subsidiary of ASE Group; and Kazuo Yamaguchi, silicon photonics researcher and engineer at Keysight Technologies. Their account points to a transition that will depend on both test equipment and the design choices made around device connections.

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

Parallel testing starts with fiber alignment

Whether a device can be tested in parallel depends on its power needs, connector placement and test stage. Armstrong said a large GPU with connectors on several sides can be difficult to test alongside another die: the tester has to supply and remove substantial power while making the required connections. Networked optics (NPO), by comparison, typically uses a single connector, which may make multi-site testing more attainable as volumes grow.

Armstrong sees later test insertions as promising places to introduce multi-site approaches. Those stages tend to involve longer, more complex testing, where parallelism could have a meaningful effect on throughput. Photonic integrated circuit (PIC) wafer testing, however, is generally single-site today. Probe-card capabilities may increase the number of sites that can be tested, though the path is still developing.

For optical engines, Griffin identified fiber alignment as the main bottleneck. Testing multiple devices at once requires a probe card or related arrangement that can align optical connections reliably across sites. Several companies are working on such systems, but production solutions remain mostly single-site. If multi-site testing becomes practical, it could help manufacturers raise output without adding the same number of separate test operations.

Connectors have to survive repeated handling

Optical probing should, in principle, avoid physical contact with the device, allowing repeated tests without wearing the optical interface. Early setup can bring risks: Griffin said teams sometimes crash fiber-array units (FAUs) into devices while bringing up a new system. Contamination is a more persistent concern, so FAUs need cleaning.

At module level, the device-side connector presents a different challenge. Optical engines may have receptacles attached, and those connectors are designed to be connected once and left in place. Each new connector design can require changes to both the connector process and the automation that handles it.

Leventhal compared the issue to repeatedly inserting a ball-grid-array device into a socket: too many cycles can damage the device’s contact points. A test-side connector must withstand many uses, potentially tens of thousands or even 100,000 insertions, while the receptacle on the device may be less durable. A hard-wearing test connector can therefore accelerate wear on the part it mates with if the two sides aren’t designed together. Connector designs will need to balance optical performance with mechanical life on both sides of the interface.

Electrical and optical contacts complicate the test setup

Making electrical and optical connections at the same time adds alignment and access challenges. At wafer and die stages, one difficult case is characterizing electro-optic devices at radio frequencies. Modulators and photodetectors are assessed through S-parameters, which requires optical access alongside high-frequency electrical contact. Probe cards have limited room for RF lines and limited bandwidth, while some device characterization extends beyond 200 GHz.

The package adds further variables. Leventhal pointed to warpage, the need to align along multiple axes, and the large number of BGA pins that must make reliable contact. Connections may need to approach from above, from the side, or from both directions. The device architecture also matters: Griffin noted that optical and electrical interfaces can sit on the top, bottom or both sides of a device. At module level, connections may span two to four package sides and several optical engines. A test system that can accommodate that range is harder to scale than one built around a single layout.

Combining operations could save time, but it takes work to coordinate them. Armstrong said the photonics test roadmap he chairs highlights AC testing and bit-error-rate testing as potential time sinks. With suitable design-for-test (DFT) features, a system could run optical loopback and electrical loopback together. Today, teams may loop back one side while running a bit-error-rate test on the other, creating a setup that can roughly double effective test time.

Griffin expects manufacturers entering volume production to use some form of loopback, whether electrical, optical or both. The automation work remains, and it’s not yet clear which approach will become standard. The practical benefit depends on whether devices can support concurrent tests and whether equipment can make all the necessary connections reliably.

Test time and reliability both matter

Insertion loss attracts attention because it is measured frequently, but Armstrong argued that it shouldn’t automatically lead the production-test priority list. Eye quality, receiver sensitivity and bit-error-rate performance deserve at least as much focus. Griffin added wavelength sweeps for optical engines, which can take substantial time depending on the requirements.

Yamaguchi emphasized that a device must be assessed for more than initial performance. A “known good die” should also be likely to remain reliable after assembly and through service. Some faults emerge only during integration: fiber-to-FAU coupling loss can change after attachment, resonances can drift relative to laser tunability, and thermal or packaging effects can alter behavior. Transmitter and receiver performance is measured using metrics such as TDECQ (Transmitter and Dispersion Eye Closure Quaternary) and receiver sensitivity.

That makes early diagnostic data valuable. Leventhal said test systems need to establish whether data works in both directions—electrical to optical and optical to electrical—and collect enough information to identify the source when it doesn’t. The next step toward higher-volume CPO testing will hinge on progress in multi-site fiber alignment, connector durability and concurrent test methods. Those choices will determine how much parallelism manufacturers can add, and how much test time they can remove, without losing confidence in the assembled device.

Read part one 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.

The post Scaling CPO Testing Beyond A Single Die appeared first on Semiconductor Engineering.