如何降低 Co-Packaged Optics（CPO）的测试成本

Key Takeaways:

• For FA, engineers need to identify boundaries and be able to decouple individual components to identify the root cause of problems.
• Underperforming lasers are a substantial issue, making the “golden reference” a hot topic in test.
• Training in electro-optics is critical to understanding tradeoffs in design and test.

Experts At The Table: Testing cost for optical components is high, approaching the manufacturing cost. Loopback capability is an important part of failure diagnosis. Semiconductor Engineering sat down 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. What follows are excerpts of that conversation. To view part one, click here. Part two is here.

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

SE: How do you determine where failure originates in electro-optical systems?

Yamaguchi: We always look at the reference plane. For the testing, these are the physical, electrical, and optical boundaries at each of the measurement planes — for example, determining where this boundary sits. Is it at the die pad, package balls, the connectors, or the fiber interface? We also need to think about de-embedding everywhere between the boundaries — for instance, probe, socket, trace, connectors, patch cord — whatever we are introducing, and making sure the same boundaries are used at testing insertions 1, 2, and 3. We need to align those boundary conditions, as well. So then we emulate the missing half. For example, to test the optical engine alone, you replace the neighbors with a known characterized substitute. If we have a golden reference laser source, that can be used.

Griffin: We are seeing customers add monitor photodiodes into their light paths, so that if they see an unexpected result from the transmit or receive path, they can trace it back. Some customers doing loopback testing want tap capability so they can route to characterization instrumentation or simply monitor power. That’s useful if they aren’t planning to do a full characterization test and just want to do loopback or mission-mode testing. The downside is that if it fails and you don’t meet your bit error rate or you have a functional issue with the program, you may not know why. So being able to tap into that loopback capability could be important.

Leventhal: I wanted to follow up about having a golden reference for a laser source. Really, the job at the end of the day is not to test the laser source. We’re looking for how to test the part. That laser source is basically the functional equivalent of the device power supply on the electrical side, and we don’t call that a golden reference. We call it a device power supply, and it’s a given that thing is supposed to work as it should, just like everything else in the test system. Customers want to default to getting the laser source out of the equation and to use the same one that they’re using in mission mode, but that laser source is not an ATE instrument, so things are going to need to evolve. As Dave talked about, FA is following your nose, and you need to be able to discern the scent of things that should be in the equation of causing trouble and things that should be just off the table because it’s a given they’re known-good.

SE: In what ways is co-packaged optics (CPO) failure diagnosis harder than near-packaged optics (NPO) FA?

Griffin: When people say NPO and CPO, there isn’t always a standard definition of those terms. The two share similar challenges, in that module test requires alignment of multiple optical connectors, and we still need to deliver power to the device. Some customers will say, ‘With NPO, I can detach the optical engines.’ That might be a key distinction. So you could argue that the test coverage before that attachment becomes less valuable, because if I detect a failure caused by a bad optical engine, I can just remove it. But we’ve also heard from customers that reworking and reattaching further downstream is a real pain point, which means they still want known-good optical engines, regardless of detachability.

Leventhal: With CPO, you have a lot more stuff integrated onto the package, and you need to have ways of decoupling them. And so many of the themes that we’ve talked about over the course of this discussion come into play here. DFT would help, as Dave discussed — understanding what information you need to get out of the testing on the individual pieces so you can correlate it to what’s going to fail in the package.

Yamaguchi: A retimed NPO has its own DSP, so the optical engine can be compliance-tested on its own, like a pluggable, and a failure is easy to locate. A linear-drive NPO removes the DSP, and the host ASIC SerDes drives the optics directly through linear drivers and TIAs (transimpedance amplifiers). The electrical and optical sides then form one signal path, so a failed link can’t easily be traced to either side. The test complexity is the same as for CPO.

Armstrong: Another concern is high power dissipation. For CPO, your electrical signals can’t drive very far, and hopefully with an NPO they can. And because you have higher reach in the electrical, you have to manage that. The thermal of all these optical parts is a big deal.

SE: How does the economics of tests change when you have an inexpensive photonics device integrated with a very expensive ASIC or GPU?

Leventhal: You can’t have an inexpensive optical engine taking down a $40,000 or $50,000 GPU or ASIC, so it’s an interesting change in terms of your test budget. We tend to look at test costs as a certain percentage of the manufacturing cost of that chip, but this blows that completely out of the water. Now you essentially have like an infinite test budget for these chips that are going to surround your big ASIC. Of course, you always look for ways to reduce test cost, but you do have the ability to over-test at the beginning to make sure that you’re not throwing away the really expensive chip. And then, over time, we will get the yield learning information as we scale.

Yamaguchi: The device value is increasing, the scrap cost is increasing, the test complexity is increasing, the package complexity is increasing. So, given this situation, additional wafers or optimal engine-level testing become more justifiable, and that’s how the economics are shifting toward pre-assembly screening rather than packaging test, which we have traditionally served in the industry.

Griffin: I completely agree. I would add that the end customers for these AI accelerator and network switch devices are highly motivated to deploy as many of these as possible into data centers so they can generate more tokens. These are some of the most complex semiconductor devices ever made. If you look at the cost of an optical engine in isolation, you get a different answer than if you look at the extremely complex package that it’s going to be integrated into, and the core functionality it provides. It reduces power and enables disaggregated compute architectures, which makes that device extremely critical. So we’re going to see continued focus on testing these systems efficiently.

Leventhal: For the industry, it’s no longer a challenge to prove that CPO works. Meta presented strong reliability data earlier this year showing it’s actually quite reliable. It works. Now the question is, ‘How do you take what we have today to high-volume manufacturing?’ That’s an ecosystem problem. It’s not just a problem in a single piece of the puzzle. So we’ll deal with that in the shorter term by relying on additional testing, screen-in, burn-in, more insertions, more engineering controls, while the ecosystem gets dialed in and matures. NPO plays a really important role in this. It’s a lower hurdle to jump over to get up and running. Learnings that come out of that will bootstrap efforts to bring CPO to HVM. But it won’t be a slam dunk. And you asked what we haven’t talked about at this point. The answer is standards. The industry is moving so fast that standards bodies can’t keep up. Where standards come into play is not to invent the technology, but to be able to manufacture at scale, where we have multiple sources for parts.

Yamaguchi: The industry is looking into standards and MSAs to enable the technology and collaborate with ecosystem partners. Standards don’t change the measurement science itself, but they align the platforms, and that is where the industry can help most. Some of the test criteria also include questions like how much optical power needs to be transmitted and what kind of bias you need to apply. That helps us prepare the manufacturing requirements and the deliverables.

Griffin: With almost every customer we talk to, there are two different groups within that company — the photonics group, which has been testing devices but typically not at the scale of semiconductor high-volume manufacturing, and the semiconductor high-volume test group. The test group is really concerned about uptime, software support, service support, ensuring they can deploy many test cells in high volume, and there are a lot of decisions you would make about a test cell for flexibility, modularity, and serviceability that may force a trade-off on optical performance. So from an ecosystem perspective, these groups have to talk to each other within the company to make educated decisions regarding acceptable trade-offs. No one knows their device better than the customer, the designers, and the manufacturing teams, so they need to understand that if I choose a certain connector type for more flexibility or higher channel count, but it results in slightly lower optical power at the DUT, that may impact the design. These conversations require multiple groups at the same company with vastly different backgrounds talking to each other, and I think we’re just now starting to see that happen because the production ramp is starting. Those conversations will do more to move the industry forward than anything else.

Armstrong: According to the surveys we’ve done in the IPSR, the cost to test an optical component is approaching the cost of manufacturing, and as the complexity rises, it may exceed the cost of manufacturing. That means the industry will have to realize that they’re going to have to pay for quality and known-good die, and as Ira alluded to, there should be money on the table because the reality is the cost of testing the GPU on a percentage basis is peanuts compared to the cost of testing the optical components. So the economies have to shift a little bit, and there could be different companies involved. Also, it takes a community. Each of us needs to work with partners to grow a structure that makes sense for all of us to help the industry move forward in a unified, consistent fashion that is scalable and high-volume production-capable.

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.