LCD Display Inverter

On March 11, in Detroit, the United States, Tesla model Y slammed into a large white truck again, and the roof was cut off. The scene was very tragic.

Looking at China, in the past month, a reporter interviewed nearly 20 Tesla owners across the country who had experienced serious car accidents, and found that most of the accidents were similar: sudden acceleration, brake failure, serious injury to the car owner, no parking without hitting a strong obstacle, and more Run 8 kilometers away. Tesla’s response was unknowable: the driver kept pressing the accelerator pedal and did not brake. With such an advanced car, won’t the “black box” data end as soon as it is made public? Either there is a fight, or the data disappears. It’s okay to do it once or twice, but again and again.

These should be “3.15” things, we can’t control them. However, a recent news that “Tesla’s 5nm automatic driving (AV) chip will be developed in cooperation with Samsung” made people ask: Don’t talk about automatic driving, it is now “normal” driving with 14nm! Is 5nm a lifesaver? Isn’t it a bigger challenge?

Looking at the discussions of some of the top company executives in the semiconductor industry on chips and packaging, systems and data, reliability and safety, ADAS and autonomous driving, especially field verification, may help us understand something.

5nm is an unknown world

The domestic report on Tesla and Samsung seems to be wrong: “Developing a 5nm semiconductor for its in-vehicle infotainment (IVI) or media control unit (MCU) for self-driving cars.”

The industry is trying to develop 5nm automotive AI chips, and challenges with process variations, electromigration, electromagnetic interference, power delivery issues, and detection abound. We have never put an advanced node chip into such an extreme environment as a car in the past. Do we really understand the “road conditions” ahead and how to deal with it?

Dennis Ciplickas, vice president of advanced solutions at PDF Solutions, said: “5nm chips are a whole new world that we’ve never been to before. There are challenges to getting, absorbing and connecting all the data. The key to advanced technology is really to understand what’s missing. Data. For example, 5nm intermediate process (MOL) has electrical interactions in 3D and you can’t see physical detection at all. This is one of the main reasons we’ve been pursuing an inline ‘detection design’ to get sensitive measurements of leaks value, and in turn discover latent defects that have the potential to translate into actual defects. To respond properly, the existence of defects must first be known, which means new data must be created. Representing data as artifacts of the manufacturing process is not enough, it needs to be differentiated data.”

Advanced packaging is also new to automotive applications. While multi-chip modules have been around for decades, the packaging is not like what we’ve seen with sensor fusion or some 7/5nm chips. Packaging also has an impact on reliability, which is another layer of complexity that must be dealt with.

Andre van de Geijn, business development manager at yieldHUB, said: “It depends on which part of the car the chip is used in. If it’s in an entertainment system, you can use the same components as millions of mobile phones, and you can trust those components. If your mobile phone High failure rate, can get another one. Many companies say it’s just one component and I can take the module out and put in a new one to replace. It’s completely different from the packaging in the car management system. The ones that make the seat move buttons fore and aft Companies will develop entirely new technology that can replace those buttons when they fail. But if it’s a car management unit, that’s a whole different story.”

Gal Carmel, general manager of the automotive division at proteanTecs, said: “Advanced packaging adds another layer of complexity, as it lacks visibility and relies on high-density architectures, which limit the fallback of redundancy, which automotive systems require. Also, the artificial intelligence (AI) part of the chip is growing. Not only does it involve packaging and advanced nodes, but the chip architecture is AI-driven and uses in-field inference and training to continuously improve the hardware architecture. Using this feedback loop, it is possible to reduce Hardware redundancy and optimized complexity.”

Uzi Baruch, vice president and general manager of the automotive business unit at OptimalPlus, said: “Encapsulation does add complexity to the hierarchy and assembly of components that need to be related in three dimensions. This in itself introduces a semantic concept of the data. It has multiple vectors , one of which is also a hierarchical element. It does add complexity because a component cannot be seen as a single unit, the hierarchy of components has to be considered. If you don’t do that, you’ll get very limited things out of the analysis. However, if you do it right, you can pinpoint the problem.”

“System-in-package or 3D integration will introduce new failure modes due to interactions between components,” Ciplickas said. “Chip-to-chip communication is different than chip-to-board communication, and electrothermal/mechanical interactions are also different. A car Manufacturers say that under stress, SRAM fails in a predictable manner, but they actually measure it on the instrument. This leads to PCB design rules for how mount points are constructed in the ECU housing. The challenge is what happens when you put a 3D or 2.5D integrated package into a harsh environment. So it’s not just communication between chips. The thermal distribution of these packages is different than you might think, changing its expansion and stress, and then changing performance, Because we know that stress changes the behavior of the device. Understanding how individual chips behave in wafer sort testing, and then understanding what’s in the package — doing package-level evaluation, and putting it all together — is a huge challenge. The use of advanced 2.5D integration in cars is a new endeavor, especially in safety-critical systems.”

Software can have unintended consequences

Now, the whole industry is talking about software-defined cars, and the more advanced cars (like Tesla) use software to control certain functions, and OTA (over-the-air technology). If you update part of a complex software system, it has the potential to affect everything in that system. If you add a lot of software, the performance of the vehicle will degrade if the hardware does not support it, and it will even affect other vehicles on the road.

“Software is challenging because it doesn’t follow any physical rules,” Ciplickas said. “Hardware sounds hard, but it actually obeys some boundary conditions. With software, you can change everything, and that can create a lot of unexpected Consequences.” What are the consequences?

“Using deep data, we can virtualize hardware to better sense the impact of software operations,” Carmel said. “With these virtualizations, you can move to an adaptive software model that targets the vehicle’s ECU performance and field degradation.” Customization. Artificial intelligence (AI) applications increase the proportion of on-chip AI to meet the needs of software. This will help reduce redundancy and ensure optimized functionality. In addition, on-the-ground inference and training will continuously improve the interface between hardware and software Interaction.” The redundancy of various systems in the vehicle, especially sensors, radars, etc., is critical to safety, and how to optimize it is indeed a matter of going backwards.

The biggest challenge is security

When it comes to design, reliability and security go hand in hand. The biggest challenge is how to build safety mechanisms into these systems?

According to Ciplickas: “Security has many aspects, but some techniques and tests used to optimize reliability can provide tools to improve security. For example, debug monitors or drift and shift monitors can detect certain types of attacks, regardless of Is it detected at t = 0, or is it anomalous behavior or drift detected in the field. These monitors are already used for system operation and optimization, and there is an associated infrastructure between the two, although they are applied in very different ways .”

“It’s an opportunity to use the data, because the more valuable the data you provide, the better the signature of the chip,” Carmel said. “Ultimately, that data will help you understand if there is an anomaly. This may be more useful for shutting down the vehicle in question.” Urgency. Using deep data, you can create 24/7 fleet visibility and identify issues as soon as they occur.”

Now, with the sheer amount of data moving through automotive systems, is it really possible to detect a very slight anomaly?

Carmel continued: “Deep data based on general chip telemetry provides insight into the operation, performance, reliability margins and performance degradation of actual chips and systems. This real-world data does not depend on the transfer of contacts, but is dependent on the output of field operations.”

Ciplickas said: “With regard to signal and noise, the industry has developed techniques to find the signal. If you look at the sensor data on the device while processing the wafer or wire bonding, you can get all kinds of signals. In these The anomalies found in the signal are sometimes tiny blobs of light. Using machine learning techniques, those tiny blobs can be found in an otherwise ‘good’ sea of ​​noise. Rather than thinking of it as a device that makes wafers Seeing as a system operating in the field, it is possible to understand these tiny spots of light.”

Jay Rathert, Advanced Director of Strategic Partnerships at KLA, said: “The biggest challenge is not seeing the defects themselves, but understanding which defects are relevant and which may be potential defects. It is valuable to connect design and improve testing. The closer the two are, the better And have an iterative or collaborative, aligned dataset, the better you can understand what inputs and outputs need to happen on the chip. We need to bridge design and test and eliminate all kinds of problems in the product life cycle .”

Using upstream data to understand downstream signals is a very powerful technique that can help predict potential problems.

“People tend to tend to look at predictive models, but they ignore the fact that the feature set — actually contributes a lot to the ability to predict things,” Baruch said. “It can filter noise, discover what’s important and what’s not, and any What is the root cause of the problem. We often use a shift-left model (putting tests at the end of the traditional software development process to the premise). To avoid looking for a needle in a haystack, a good model can help you discover what is a priority problem , as long as you decide which angle to look at. You build these models, you predict something, and you need some people to fix the wrong properties in those models.”

It’s far from the time to build fully self-driving cars

ADAS and autonomous driving don’t seem to be the same thing. To get into fully autonomous driving, one must start thinking about system-of-systems (SoS) collaboration. What happens if cars use different generations of chips and software?

“The fundamental point of going from ADAS to AV is to understand what kind of failures they have experienced,” Carmel said. “Ultimately, the performance envelope needs to be defined. Every vehicle has its own performance envelope because it has different hardware, different software, different level. When you know exactly how to define this performance range and create a balance between reliability and safety, you can control the fleet. Using deep data, we can define the independent function and outline an autonomous driving hierarchy.”