Looking At Test Differently

Wilhelm Radermacher, executive advisor at Advantest, sat down with Semiconductor Engineering to discuss how the impact of rapid market changes, advanced packaging approaches and increasing complexity on test strategies and equipment. What follows are excerpts of that conversation.

SE: As we move into new markets where use models and stresses on devices are different, do we have to change the way we test? Does it have to be done in more depth? Automotive makers want chips to last 15 years with zero defects.

Radermacher: For many of these areas we know what the objective is, but we don’t know how to do it. There’s safety involved when you are talking about autonomous cars, and cars will need to drive themselves not only in good weather, but also in situations involving icy roads, fog, and heavy traffic in cities. You want to make sure it’s absolutely safe. That’s a hard problem to solve. The equipment has to have a lot of redundancy, and it has to be able to test itself in operation to make sure it’s working properly. This presents a whole new set of challenges for test, and the solution needs to include different partners that work in this industry. On one side, you can argue this is a very different kind of test, which has nothing to do with the kind of test that’s used in manufacturing. But at the same time, the electronics that create the problem need to be used to solve the problem. So we need to work very closely with all the different parties to get things like autonomous driving to work, and that includes doing test in parallel to make sure that the systems are healthy and that they work when a vehicle is in operation. That’s a big challenge.

SE: What changes in terms of how test equipment is used, and what kind of equipment gets used for what?

Radermacher: In automotive there is a lot of electronics that also has been used in the consumer area. The difference is that in automotive you have much more stringent temperature range requirements. At the same time, you also need to guarantee that your equipment is healthy, and that the test programs that run on top of the equipment-and all the parts used to make sure you do not have any escapes in chips-are healthy, as well. That’s all the standard stuff we do today, and it’s already challenging. But we also need to work together with our partners to discuss new test challenges, to define test methodologies, and to define all the test equipment that will be required to support the test methodologies.

SE: Does the test equipment that you have today still work well enough? Is it just a matter of using more of it, or does that have to be changed?

Radermacher: If you go back 10 to 15 years, the mobile space appeared to be a huge challenge. There are chips with a lot of different radios and multi-core IP blocks. When you put all of that together, how would you test that and would it change the test equipment? The result was that we changed the test equipment a lot, and that’s also what will be expected in the next 10 to 15 years in the automotive space. That doesn’t mean test equipment used today won’t be used anymore, but for the new challenges we need to either augment that equipment through innovation and by adding to it, or in some cases building new equipment. For example, system-level test is an interesting field. There is a lot of discussion about this. It becomes obvious that we need to do system-level type test earlier in the supply chain, ideally at the wafer-level. And it has to be done efficiently, in parallel. It’s a necessity for test equipment to be able to this. But today’s equipment will not support that well enough.

SE: Define system-level test. There are a lot of terms floating around for system-level these days.

Radermacher: Part of the problem is that this is a changing area. People would love to have one definition, although if you did have a single definition that would probably not serve it. What you’re trying to do is put any device, subsystem, or system-different levels of hierarchy-into a situation that is as close as possible to what you see in the final operation of that device or equipment. You integrate enough of the hardware and software, and potentially other components, that you excite all the possible use cases and problem areas to avoid escapes of this device. Let’s assume you have a mobile phone and do calls on the phone in the final version-that’s what I would call system test. System-level test is trying to mimic enough of the final environment, so you put a mobile phone chip onto a mobile phone board and run enough software on it so you avoid some weird escapes that won’t found by structural/functional test. That is what most people think about system-level test. Where confusion often comes in, especially in the last couple of years, is system-level test for let’s say an RF module, which is a very different subject than a system-level test for a CPU, a GPU, or a baseband device. You would see very different equipment that would be used. Some of the issues may be different, but they both would still be considered system-level test. Is it a multi-chip advanced package device? Is it an RF module with stacks of components on it? Or is it a soft part of a graphics computer? If you know what the answer is, you can get eliminate some of the confusion. What they all have in common is to bring hardware and software close enough to the environment so you excite the issues that would not be found by functional or structural test.

SE: What about advanced packaging? One of the issues with 2.5D or fan-outs is that some of the leads are inaccessible from the outside. The more complex the advanced packaging gets, the harder it’s going to be to connect and test directly. And does it all have to be tested together, or can some of it be done separately?

Radermacher: That’s exactly one of the issues. Some of these components that maybe something one company sells to another one, and you’re putting multiple devices into an advanced package. You can’t completely test them individually upfront because they wouldn’t even drive enough current over distance to be put in a tester. What that means is you do enough structural-type test on a device before you add it to another device to make sure you’re not wasting value of all of the components, particularly if you have a certain failure rate. If you have 99.99% confidence that a device will work with something with the kind of structural test you can usually do, that is sufficient to make sure you’re not wasting other devices. All of these pieces have a certain dollar value even before you put them together. So if you have 99.9% confidence in these parts, that usually is sufficient to ensure the final component when you put them all together is at a quality level that you can sell. Then you can test the integrated components, where access between them is not so important anymore. That’s a practical approach that is used these days, and it needs to be developed further. The system-level test also needs to be pulled up front to the wafer level. Otherwise, it will take too long, and in the end it will be too expensive because test at the end is always expensive. It’s not the final answer, but today it is a very valuable answer for making sure you are delivering quality to the end customer.

SE: When you look at a lot of the new markets that are coming up compared to PCs going back 20 years ago or mobile phone going back even 2 years ago, the big difference is that there’s a lot of smaller batches of chips being manufactured that are unique. With a mobile phone chip you may produce billions, or hundreds or millions of a single design. What does that do to testing?

Radermacher: That’s a good point. System-level test means you have to provide the specific environment for that chip. It used to be that you could test all devices on just one ATE-type test equipment by just writing a test program ant maybe maximum load ports specifically for this device. That is still something you need to strive for. If you have multiple different devices, you want to keep the amount of customization that’s needed to test one device to a minimum. But there’s not one answer on how to get there. For example, you can see that most devices, even small batch-type devices, share some common core-like an Arm core. We need to use this type of intelligence inside the chip, with an infrastructure built into the chip that supports keeping the test as generic as possible externally.

SE: How much of this comes from Advantest versus the manufacturers working out-for example, developers of an Arm core or a GPU-how is all of this supposed to behave based upon all the specs?

Radermacher: The other party you didn’t mention is the EDA companies. They’re helping the companies that put the different IP together. That’s why the discussion of ATE and EDA working together is now real. If you don’t have a means to automatically design test into the chip, there is no way to provide to a broad field of different, small types of devices the capabilities to test them on generic equipment later on.

SE: So you’re coming at this from both inside-out and outside-in?

Radermacher: Yes, you need to make sure that while you’re designing the circuitry for the product at the end of the module that you put in the hooks and architectures that allows you afterward to apply the necessary equipment from outside. Of course, that equipment needs to support that new way of testing. A simple vector blast of a chip test will not do a lot of good here if you have, for example, a protocol interface in the future that goes into a DFT-type architecture.

SE: The flipside of that is that it’s going to take a lot more to effectively test from two standpoints. One is time. The second one is flexibility in that you need to be able to configure these tests much more on the fly than what has been done in the past.

Radermacher: You’ve described a big challenge. You probably would agree that if you go back 10 years ago, people very much doubted we could keep scaling test to keep pace with , and that you would be able to test WiFi combo devices in less than a second and still improve quality. A lot there has been used to drive parallelism, and a lot has also been changed on the equipment itself to make it smart and adaptable to different technologies. That needs to be continuously moved forward. But now we need to make some more inroads into this area of working between EDA and ATE together, and also between DFT and all customers. This may be for small batch-type devices, only working if they use common cores. But it certainly is part of what is also architecturally correct for what we call IoT. Every IoT device is a unique design-not designs from a platform with some core IP that takes care of some of these things. At the beginning people knew IoT only as cheap devices that could do everything, but over time there has to be architectural innovation in terms of platforms. That is still yet to happen. If you look at the automobile companies, behind the scenes are platforms with reusable parts, which they use to put out different cars. The same needs to be happening in the area of devices manufactured in small batches.

SE: In developing Advantest testers, which are typically huge multi-year efforts, does it now become more modularized with smaller upgrades and changes? Or is it a full system that has to be redone every few years?

Radermacher: When you look at the test system today, like standard platform-type equipment, it has changed a lot. There are modular cards that are used to build out the system, and for what you describe here we will need additional modularity in the future. If you look at a more recent generation of equipment, you can see a back-end that’s kind of generic, and for all the different RF application areas there are front-end modules that are integrated in the test pads, which can be changed. These kinds of approaches are needed for hardware and also software to innovate faster than the multiple-year effort that it takes today. At the same time, the reason you exchange a digital card in a tester only every five to six years is because to get enough value out of this, you can’t change the whole tester every year. It would not provide an economical solution. So to handle the new challenges you need a mixture of modular architecture with a stable back-end, that is hopefully generic enough for 5 to 10 years, and then you can change out hardware and software modules.

SE: This is probably the most interesting time then in ATE history, right?

Radermacher: Yes. At Semicon West, the main topic of discussion in recent years was whether Moore’s Law is coming to an end. Last year’s show was much more enthusiastic about the overall electronics market opportunities. But with these opportunities come a ton of challenges-like the challenges we discussed in the automotive area and the extremely short time to market. You look at the length of time the big vendors take to bring out the next generation of mobile phones compared to even two years ago. That time has shrunk. Along with that, the device has now 5 billion transistors and is integrated in an advanced package. It’s very challenging from an engineering and market standpoint, but it’s also where money is being made. That provides a good opportunity for innovative people working together, and it’s an absolute necessity that groups across the globe work together. Even 10 to 15 years ago, you’d say you need to work together, but if you had a smart engineering team you still could do decent work without anyone else’s help. Today, with work between foundries on one side, the OSATs that use equipment later on, and the upstream customer that designed the product with EDA, there’s a lot of opportunity to solve these issues and add value.

SE: So you have a whole bunch of new partners here that you’re getting a lot closer to than you’ve been in the past?

Radermacher: Yes, and some of these relationships are new for us and still need to be developed. For example, there hasn’t been a lot of engagement in the past with automotive. If you go to one of the big automobile companies that are integrating electronics, they aren’t the experts in that field-they rely on the Nvidia’s of the world to do this. So they need to develop a relationship with the end customer as well as the Bosch and Nvidia’s of the world. That’s exciting for young engineers, and also for the ones that have been here for a while.

SE: Change is happening everywhere on a massive scale.

Radermacher: Just a year ago you heard a lot of skepticism. Now you see a much more positive view of the opportunities lying ahead of us. In the digital space, for example, it used to be that infrastructure-wise you had servers and computers of different sizes, but all of a sudden the architecture of the products are much closer to the end customer. You don’t just have big server farms. You also have edge computing and people using computer hardware just to mine bitcoins. There is an explosion of opportunities.

SE: As you go into 7nm, 5nm and beyond, there are a lot more things you need to make sure are working-like dynamic power, heat and security. Are you being drawn into these new challenges with testing?

Radermacher: When you talk to people dealing with finFET-type devices, often they’ll say it’s the same, but that’s not completely true. The amount of scan data is exploding. The voltages are getting smaller. Getting enough coverage is difficult. There’s a lot more challenges there. We need to do more work on the parametric test for processes.

SE: In the past when companies looked at testing, they allotted certain percentage of dollars and time. Is this changing? Does the economic formula change for the customer as this gets more complicated?

Radermacher: If you look at the history on the percentage of cost versus manufacturing cost, it has actually come down, despite that we have much more challenging test these days. It has scaled quite well. But right now, we’re at a stage where we see, at least recently, that there is more test needed than in the past, and the quality expectations are also growing. In years past people started to see huge returns in the mobile phone world. It was unacceptable for wireless LAN connections. The big companies that drive this market simply have set a new level of quality. When that happens, all of a sudden the economic formula is not the most important part of this. At the same time, test can only be a certain percentage of the overall manufacturing cost for some products at the low end. Otherwise, the economics wouldn’t work. Still, when it comes to quality, if you can’t provide it any other way, then people will spend more for test. For a tester company in the long run, what is more important is that we keep striving for adding more value. If you have a tester company integrated at the front end to help with the early phase of a new device to get the yield up faster, then there’s a strong economic value and the customer will pay for that.

SE: That is one of the big changes. You need to get a product to market very quickly, and it has to work well enough and be unique enough in order to win market share. Then they can modify it afterward. But it also requires more iterations than in the past.

Radermacher: On the high volume products-using the cell phone manufacturers as an example-they need to provide a relatively large volume of chips to their end customers very early after the first wafer. That’s a big change.

SE: We’re also seeing that more people are designing test strategies upfront than they ever were in the past.

Radermacher: Yes. One of the opportunities that we have seen over the past year is that design people who never before used testers are the ones who develop test vectors. They need to make sure that by the time they release the wafer to production that the test is developed to a level so that when the chip comes back that they are able to get enough data back to them to debug their own device. That’s just the beginning.