Using Better Data To Shorten Test Time

More sensors plus machine learning are making it possible to focus testing where it has the most impact.

popularity

The combination of machine learning plus more sensors embedded into IC manufacturing equipment is creating new possibilities for more targeted testing and faster throughput for fabs and OSATs.

The goal is to improve quality and reduce the cost of manufacturing complex chips, where time spent in manufacturing is ballooning at the most advanced nodes. As the number of transistors on a die increases, and more dies are added onto a board or into a package, it takes more time to test all of those devices. That, in turn, adds to the overall cost, or it reduces confidence that the device under test will work reliably over its lifetime if the test time remains constant.

One solution is to differentiate between what really does have to be tested thoroughly, versus those places where there is enough confidence in devices that have been proven in silicon that testing is not so critical. This amounts to pruning of a test suite, but making that determination requires gathering enough useful data to be able to make that call. In the past, this was difficult to do, but as more sensors are added and more data is generated from those sensors, there is an increased ability to look at that data with much more granularity and to identify patterns from that data.

“If you can build a model that’s 99.99% accurate at predicting when a chip is going to pass burn-in, then you could skip burn-in on that chip and save the money,” said Dennis Ciplickas, vice president of characterization solutions at PDF Solutions. “You could set a target skip rate of 20%, 30% or 50%—whatever percentage you decide on—and save that burn-in cost. The more data you have, the better the prediction you can make. The challenge is how do you connect the right data to do that. So if all of your data is running at an OSAT, and your wafer sort and assembly and your final test is there, then you can integrate all of that and make predictions from it. But if you do wafer sort in one place and assembly and final test in another, you now have to merge data from multiple sites and you need a system to do that.”

Burn-in test is used to detect early failures in various components in a chip. In the past, this kind of data was shared through a data exchange, and while helpful, that isn’t sufficient to eliminate testing.

“The idea behind a data exchange network originally was to build a database so that you could have visibility into your supply chain,” Ciplickas said. “But if you have all of that data, machine learning now lets you take many different features and put them together in order to do something new, like prediction. So now you can control burn-in cost. Having data from the sensors was the first step. Having the data be able to flow through different sites together in a coherent, connected way was the next part. And then being able to use all for that for prediction is the next step.”

In effect, this becomes a much more granular way to characterize various components throughout the design-through-manufacturing flow, and to precisely eliminate what testing isn’t necessary.

“What this enables you to do is wring out first-order, second-order and maybe even third-order areas where there is reduced efficiency,” said Doug Elder, vice president and general manager of OptimalPlus. “Now you can take silos where you did wafer sort and final test and connect all the data sources to figure out where the failures come from. So you can reduce your test set and areas like burn-in, but that’s just one area where we’re seeing benefits. You also can run machine learning algorithms against these to fix test and improve your recipes.”

Image by <a href="https://pixabay.com/users/TheDigitalArtist-202249/?utm_source=link-attribution&utm_medium=referral&utm_campaign=image&utm_content=3964871">Pete Linforth</a> from <a href="https://pixabay.com/?utm_source=link-attribution&utm_medium=referral&utm_campaign=image&utm_content=3964871">Pixabay</a>Image by Pete Linforth from Pixabay

In the past, manufacturing issues may have taken months or years to show up, frequently after use in the field. Some of those problems could be patched through software, such as antenna issues in smartphones, which worked well enough until the technology was replaced. But as more chips are used in industrial and automotive applications, where they are expected to function according to spec for a couple decades, that approach is no longer adequate.

“If you can identify these issues back at the inspection and burn-in level, then you can decrease your burn-in test time by 10% to 20%,” Elder said. “That’s a big number for manufacturers. You can do adaptive test time reduction, where in real-time you statistically look at what tests are not failing. That way you can decrease the test time, and you can add that back into the test cycle with a closed-loop system, which can include everything from wafer sort to final test. The other area where we’re seeing a lot of extra cost is in re-test. How many times can you re-test something before you damage it? We saw one case where a device was re-tested 17 times until it finally passed.”

That’s bad for the device under test, but it’s also expensive and time-consuming. In a foundry, every fraction of a second has a cost attached to it.

Better data, better results
A big change in this equation is better data. This is a relative term, because “good” data can mean different things in different markets, and often even within the same market. What makes data good is its usefulness for a particular operation or process or device. But making that determination isn’t always obvious, and it frequently requires a fair amount of domain expertise.

“We find that some of the smaller companies we talk with don’t have expertise in every area,” John O’Donnell, CEO of yieldHUB. “This is happening in automotive, but it’s not just confined to that market. Some companies could be very strong in design, but they’re not as strong in test.”

That creates a problem with complex chip designs, and it will become even more critical for chiplets once that approach begins gathering steam. But it also highlights one of the challenges across the design-through-manufacturing chain, which is making sure everyone involved understands what has been done in other parts of the chain.

“Someone might ask, ‘Do we really need to spend time on this particular block or this part of the die? What is important is that everyone on the team has to see this,” O’Donnell said. “You need collaboration as well as analysis, because you might be an expert for a particular part of the chip but not on another part. That needs to be maintained in the knowledge base. It also allows you to decrease the number of tests required to save costs because you have enough confidence in a particular area, so you can add your knowledge to the analysis of the system. What you don’t want to do is eliminate something and months later no one can figure out why that was eliminated.”

Another way to ensure the quality of data is with a continuous feed of that data. So rather than all external data, some of that data can be measured while a device is in operation.

“With in-circuit data, you have the ability to improve performance and make the right decisions on a continuous basis,” said Shai Cohen, CEO of proteanTecs. “You can measure defects parametrically both in the process and in memory. That provides high coverage, but you need to add in multi-dimensional Agents, which can be measured and processed with machine learning. So now you can have Agents for a specific design, and you can reconstruct that data for better visibility into what’s happening.”

Garbage in, garbage out
None of this matters if the sensors aren’t accurate enough. Consider the Boeing 737 MAX, for example, where faulty sensors were the likely culprit behind two crashes and a number of scares. One of the problems, though, is that data needs to be shared more widely than is currently done for errors to be identified, and system vendors tend to guard that data as a competitive edge.

“In test, the trend is to follow chip design where you standardize on test, but OEMs can’t trust that it works,” said Doug Farrell, principal solutions manager for transportation at National Instruments. “They’re still reluctant to share data. That will have to change because you can’t have everything in a single company. This is especially true when it comes to autonomous driving, where the Tier 1’s and the OEMs are competing.”

One solution is to continually test the sensors to make sure the data coming out of them is of good quality.

“For the people who operate fleets of vehicles, at the end of the shift they can diagnose the sensors and do calibrations on them,” Farrell said. “This is essential because we’re seeing companies moving from pure simulation directly to putting those sensors in vehicles. There is supposed to be a middle step, but they’re skipping it. A lot of companies don’t have the resources, but that’s irresponsible from our perspective.”

The data itself also has to be stored in a way that prevents future problems.

“What we’re trying to avoid is silent data corruption,” said PDF’s Ciplickas. “When that happens, and you don’t know it, you’ve got egg on your face.”

For safety-critical applications, the result can be much worse than that. There is liability now associated with the functionality of those devices, and data is the best way to determine what went wrong and why.

“The fundamental difference between the automotive industry and the semiconductor industry is that test is only one vector in automotive,” said Uzi Baruch, general manager of OptimalPlus’ Electronics Division. “That’s more than what you typically see in the semiconductor business. And it’s not just limited to electronics. It’s the full assembly line where you have multiple touch points.”

Design for inspection
An important piece of this puzzle involves inspection, which is becoming more difficult at advanced nodes and in certain types of advanced packaging. As a result, there is renewed emphasis on putting more sensors in more places—both in the equipment as well as in the packages—and on making those sensors smaller, faster and much lower power.

“There already is sampling going on today at the fab and OSAT level,” said Subodh Kulkarni, CEO of CyberOptics. “Right now we are doing functional check of packages. The question is how much fallout there will be before we get to 100% inspection. This is necessary because the packages are so expensive. The need is clearly there to inspect a package for a reasonable cost.”

This becomes more difficult as new materials are introduced and new structures are added onto packages, such as microbumps and increasingly skinny pillars. In addition, different materials require different inspection techniques, because some diffuse light differently than others. Both of those can require recalibration of inspection equipment, such as the optical cameras, said Kulkarni.

“It’s becoming more difficult with advanced packaging inspection and module inspection, particularly for things like HBM,” he said. “In the past, 2D inspection was good enough. Now you need 3D optical inspection for memory modules.”

That takes more time and it generates significantly more data that needs to be analyzed.

“If you look at the welding operations in automotive, at the end of the welding there is a visual inspection,” said OptimalPlus’ Baruch. “You could analyze the data and go from there. But now you go back and look at in-circuit test and expand it. The goal is to shift left on the line so that you can classify images and determine what is a good weld and what is not. Ultimately, what you want to get to is the point where the machine can go back and fix the weld because you can predict those characteristics that end up in a bad image.”

Conclusion
Using sensor data and machine learning together is just beginning to catch on in the manufacturing world, but the opportunity for improving quality and reducing test time is enormous. That also can help reduce redundancy in design, which is costly in terms of parts, power and weight.

“The goal is absolutely to put an end to the growth in redundancy,” said Raanan Gewirtzman, chief business officer at proteanTecs. “That’s particularly important with ISO 26262, which requires some level of redundancy. But that can be handled by better measurement by agents, and we can keep adding more of them.”

That requires a whole different way of looking at and using data, but the promise of better coverage for less money is one that is getting a lot of attention everywhere these days. The big question is how far can this approach be extended.

Related Stories
Gaps Emerge In Automotive Test
Reliability requires different parts to work in sync, and much more time-consuming testing and simulation.
Automotive, AI Drive Big Changes In Test
DFT strategies are becoming intertwined with design strategies at the beginning of the design process.
Big Shifts In Big Data
Why the growth of cloud and edge computing and the processing of more data will have a profound effect on semiconductor design and manufacturing.
Using Better Data To Shorten Test Time
More sensors plus machine learning are making it possible to focus testing where it has the most impact.
Data Confusion At The Edge
Disparities in processors and data types will have an unpredictable impact on AI systems.
Data Analytics Knowledge Center
More top stories, special reports, blogs and white papers on Data Analytics



Leave a Reply


(Note: This name will be displayed publicly)