Chips Good Enough To Bet Your Life On

Semiconductor Engineering sat down to discuss automotive electronics reliability with Jay Rathert, senior director of strategic collaborations at KLA; Dennis Ciplickas, vice president of advanced solutions at PDF Solutions; Uzi Baruch, vice president and general manager of the automotive business unit at OptimalPlus; Gal Carmel, general manager of proteanTecs‘ Automotive Division; Andre van de Geijn, business development manager at yieldHUB; and Jeff Phillips, go to market lead for transportation at National Instruments. What follows are excerpts of that conversation.

SE: As cars become increasingly electrified, and as more data being collected in cars is digitized, more advanced electronics are needed. There is no history of using a 7nm or 5nm chip in these kinds of extreme environments. What do you expect will happen?

van de Geijn: I’ve been to a number of automotive sites, and what we’ve seen from a testing point of view is that more information from the foundry is coming over to the assembly and test sites. There we have to merge data to make sure that we do the right things and make sure the parts are working. But we also see those customers are starting to work in different ways. In the past you had a single microcontroller in your motor or engine control unit. These days there is more than one microcontroller. If one is failing, especially for those critical systems, the other one is going to take over. Even though you can do all kinds of testing and all kinds of reliability things, the FMEAs (failure modes and effects analyses) show the only way to cover that is to have redundancy in your microcontrollers for those critical items. They need to take over if one of the parts is failing.

Rathert: I was in a meeting with the head of the electronics R&D group at a large carmaker a few years ago. He said you haven’t built a car until you’ve argued over trying to drive a nickel’s cost out of a part. I come from an aviation background and redundancy is the way that we normally take care of reliability issues — three flight computers voting on who’s right. But we don’t have that luxury in an automobile, where we’re trying to save a nickel. We feel a lot of pressure through the whole supply chain to get each individual device to be reliable. Redundancy seems the obvious path, certainly for the higher-end parts where a die can be $15, $20 or more. But car companies are loath to put multiple units in there if they don’t have to. They’d rather look for other solutions.

Phillips: Autonomy is going to happen. It’s inevitable. But the path to get there is going to come down to two key questions. One is, ‘What do you mean by reliable?’ The other is , ‘What do you mean by safe?’ If the expectation is that safe and reliable mean zero accidents out of the gate, that’s just not going to happen. We’re going to see the testing process play out on the streets. We’re going to see failures happen. On the other hand, if we’re comparing that to human drivers, we’ve already seen technology have an impact on safety and a reduction of crashes. It’s just a matter of whether we’re going to shift as consumers away from comparing technology against ourselves and aim toward perfection. But it’s going to take a long time before every car everywhere is fully autonomous.

Ciplickas: Moving to the advanced technologies is incredibly exciting, if a little bit scary. Trusting a car to do what I can do, and all the compute it will take to really emulate that, is not easy. But at the same time, the possibilities it opens up are amazing. There are certainly some issues with advanced technologies like 7nm. We’ve seen what types of things happen in 7nm die —the different types of variations, the defects you can get in the middle of line, and all the interactions between process modules. Comprehending that in the way you test, the fault models you use, diagnosing and finding defects, and building functional safety that comprehend these factors, is a huge challenge. We are going to move toward autonomy, and we’re eventually going to get there. I personally think it’s going to be better than what humans can do right now, and that’s going to drive the changeover. But the innovation it’s going to take to make that happen, given the way the advanced technologies behave, is a big challenge and also a big opportunity.

Baruch: I’ve been walking many production lines in the last couple of years, and those production lines are suddenly evolving. It’s almost like having an iPhone inside of your car, because the complexity of producing each and every part is so tremendous. There are many opportunities for failure. Getting to the point where you can repeat the process in a controlled manner and find root-cause issues inside those [production] lines, or between plants or different suppliers, is creating a significant challenge for reliability. Whatever you’re producing needs to be repeatable, and you need to be able to trust it. That’s causing a significant shift in how these companies are operating.

Carmel: Increasing the performance envelope above L2+ requires some trust in the vehicle, once the system takes control. One benefit of these systems is that they are 100% focused, as opposed to drivers who have many distractions from things like data-push and cellular technology. Our job is, eventually, to increase the availability of those systems. So OEMs are looking for scalable ways to support the growing performance envelope of advanced SW technology, and in parallel maintain electronic reliability in advanced nodes. But what they need is data to know when and how to fail gracefully, and balance availability with safety in their fail-safe mechanism. It’s a trade-off, and perfection does not exist today, but we are working hard to change that equation by obtaining accurate visibility on these systems’ health.

SE: Latent defects typically were documented in stable environments in the past, but that gets more difficult in automotive because cars operate in very different, often harsh environments. Is the solutions redundancy, which the carmakers don’t seem to want to pay for, or is there some other approach needed?

Rathert: We believe the solution is a multifaceted approach. There is no one single approach that solves everything. Well-designed, well-built, low-defectivity devices that are well-fabricated using tight processes are the foundation to build on. But we don’t think that’s the only answer. Nor do we think test by itself is the only answer. The merger of these, along with real-time diagnostics and the capability to look across the supply chain and find weak points from design to the final system — all of these things coming together are the industry’s best hope to create a zero-defect solution. It’s not just one of us.

Phillips: Given the breadth and depth of technology and communication standards that are in an autonomous vehicle, there’s not going to be one or two companies that come together with a holistic end-to-end test solution for cars. It’s going to be different people in the ecosystem working together, collaborating on how to interoperate and integrate our solutions. There will be strengths from manufacturing and test analytics applied to radar, lidar, and I/O, using cloud and infrastructure processing. But there also are a lot of different technology vectors that have to come together.

SE: Is the data that’s being collected from all the sensors in a vehicle good enough to understand whether there will be a problem in the future? And if so, can you make that prediction early enough to avoid problems?

Baruch: There’s a lot of data coming from different angles and different suppliers. We recently finished what was essentially a hackathon with one of the premium automotive brands in Germany, combining data from a car with manufacturing data to see if we could predict a failure in a sensor, and to build a predictive model to find failures. There are a couple of important findings. First, data exists, but it has to be leveraged in a connected way. Today, some of the areas where the data is coming from are siloed. If it’s from the same manufacturer, then great. That’s not the reality, though. You see components coming from many suppliers and sub-suppliers. So the data is there, but it’s not clear whether they’re willing to share it to solve a particular problem that another company or the customer has. However, what we do see is a growing need by OEMs to get visibility down the chain. They are forcing their suppliers, all the way down to the semiconductor companies, to supply data along with what they’re producing. It’s a sort of digital signature for the component they are producing. That is the start for unlocking the potential of connecting all those data points.

Ciplickas: I agree. There’s data out there collected at every step of the manufacturing, test and assembly process. There’s tons of data produced from inside the factories at every millisecond or microsecond, all the way through to the field, where there are in-die sensors. The trick is putting it all together in a single analytics framework. There are huge challenges to actually share that data. Every individual owner of that data uses it for their own purposes, but there also are a ton of benefits to putting it together in terms of time-to-root-cause and improved quality screening. There are practical issues about moving it around the world in a secure or confidential way. And then there’s the privacy issue. Breaking that down is going to be tricky, but the benefits in yield, quality, safety and security are worth it. What we found working in the fab and in the assembly line, and with our characterization vehicle test chips, is the closer you get to the actual physics of the system, the more likely you can build a predictive model. For example, this can be a sensor on a tool that shows why a certain voltage is what changed or glitched or moved or ramped differently, which then led to this effect, which led to that defect. Standard metrology might not show you that, but the sensors that are measuring the physics can. So building up an understanding of why something behaved the way it did is critical, and a physics-based understanding is a key to getting a predictive model.

Rathert: Even in adjacent silos, it sometimes can be difficult to clean or align the data. So the dream and vision is there, but the practical reality is we’ve got a few steps to go to make it all work.

Carmel: The amount of data that’s flowing from the car to the cloud is rapidly growing, and this is good. This data is crucial to design better systems. GDPR (the EU’s General Data Protection Regulation) regulates exactly what data you can move and how you can move it. The OEMs ensure that the data being moved does not reveal any private information or add security risks. In other markets we don’t have these regulations, so it’s good that autonomous vehicles will protect this information. But the problem is finding unique data within all of this. If we cannot build a model that eventually helps us build better systems, then the data isn’t worth much. We need to create a learning curve from a particular set of failures and understand which data is relevant for which entity in order for them to build better models. And then it’s a very powerful tool for every entity along the way.

van de Geijn: To certain extent, I agree. But you also want to test the autonomous systems by themselves and make sure they are working. You can collect tons of data, but it needs to be relevant to the problem you want to want to fix. Collecting a lot of data in certain areas makes sense. Still, there are a lot of things that can happen to automotive products. You cannot totally predict that a single particle on a single transistor can mess up a function in a product just by using the data from sensors or other systems. You still need redundancy built into your product in case something unexpectedly fails. It is as important as collecting all that data. It’s another area that I see a lot of companies working on, and they are using those kinds of things to prevent products from failing.

Related
Making Chips To Last Their Expected Lifetimes
Lifecycles can vary greatly for different markets, and by application within those markets.
Growing Complexity Adds To Auto IC Safety Challenges
Building together a verification approach for safety and security in automotive chips is complex, multi-faceted, and extremely difficult.
Sensor Fusion Challenges In Cars
As more pieces of the autonomous vehicle puzzle come into view, the enormity of the challenge grows.