Automakers Take On More Responsibility

Carmakers shake up supply chain with their own test and system-level integration strategies.


Chip and EDA companies are scrambling to deal with stiff safety regulations and harsh environmental conditions for automotive chips, but automakers are making big changes of their own to ensure all those components work together as expected.

The result is a significant shift of responsibilities of companies in the automotive supply chain. Carmakers traditionally have left verification, validation and testing of chips and subsystems to their suppliers. In years past, it was enough for those companies—chipmakers and Tier 1 and Tier 2 suppliers—to certify that every component they deliver complied with required safety standards and could interoperate well enough to be plugged in with no further testing. But as the amount of electronic content increases and the complexity of these systems grows, not to mention the deployment of these devices in safety-critical systems, carmakers are taking on a much bigger role to ensure these components work in the context of other systems.

There are several key trends driving these changes. Among them:

  • Complexity is rising significantly, particularly as systems fail over to other systems within a vehicle in accordance with ISO 26262. Testing those systems independently is no longer sufficient, and automakers are in the best position to verify that all of the pieces work together.
  • The cost of errors and defects in assisted and autonomous could dwarf issues that have caused costly recalls in the past, particularly as technology takes growing responsibility for on-road maneuvering.
  • The automotive supply chain currently is divided into chipmakers and suppliers with limited experience in advanced semiconductor technology, and leading-edge semiconductor companies with limited experience in the automotive world. This is particularly evident at 10/7nm, which is where the AI systems are being developed to manage autonomous vehicles.

While carmakers can help guarantee that the integration of all of those pieces works, they are on a steep learning curve, as well. This is brand new technology, and there are almost certain to be some unexpected speed bumps along the way.

“You need to take a holistic view,” said Burkhard Huhnke, vice president of automotive strategy at Synopsys. “You can’t just judge the hardware. What’s different from the past is there is a serious issue in the supply chain. We have to come together as an industry. If there are safety islands on an SoC but those are not available to other systems, there is a huge problem.”

Dealing with problems in context
Moving data quickly and efficiently has been a core focus for semiconductor companies, and it is an essential building block of the von Neumann architecture. But moving enormous amounts of data being generated by automotive sensors is both inefficient and time-consuming, and that will become more obvious as automakers push to autonomy levels 3, 4 and 5.

That means more data has to be processed in place and mined for what’s critical, what is important but not time-sensitive, and what is considered useless. Still, all of that needs to be dealt with appropriately by systems, and systems of systems. Testing for that is well beyond what has ever been done in electronics.

For chipmakers and Tier 1/2 suppliers, this means verifying chips that manage LTE, WiFi or other wireless communications protocols will perform as advertised without interruptions from such factors as electromagnetic interference or other types of noise, said Neil Hand, director of marketing for the IC Verification Solutions division of Mentor, a Siemens Business. And for carmakers, it means putting all of the pieces in place to be able to figure out what can go wrong.

Most carmakers, however, are on a steep learning curve. After nearly a century of mechanical engineering, they now are being pulled deep into advanced electronics, from sensor fusion to AI systems being developed at leading-edge nodes with advanced packaging.

“You need to be able to do multiphysics simulation for the actuation and have sample sets, and be able to handle emulated sensor data to the multiphysics output so you can close the loop,” Hand said. “But that still doesn’t take into account the whole system of systems.”

Fig. 1: Automotive electronics cost as a percentage of total car cost worldwide, 1950 to 2030. Source: Statista

This marks an enormous shift in responsibility for carmakers, which in the past they subcontracted that kind of verification, validation and testing to their supply chain.

“Traditionally you’d have OEMs or systems integrators who could take off-the-shelf chips, put them together, and the world would be a happy place,” Hand said. “Now you’re seeing car companies start to do that integration testing themselves and taking on the liability to make sure it all works as a system. That’s a new set of skills and requirements for those teams.”

A real-world example
One of the more ambitious projects to surface in this arena involves Mazda Motor Corp., which trails larger companies in the development of electric vehicles as well as overall automotive market share, which is in the low single digits.

“With its scale, Mazda can’t match the kind of investment that Ford or Nissan or GM can bring, so they wanted to be a little creative in minimizing cost and raising the quality of the product,” said Michi Kubo, a test engineer for National Instruments who helped Mazda develop a custom simulation testing suite that included a robot arm to simulate human interaction by pressing buttons during tests.

Mazda’s Electronics Testing & Research Group, which developed the test suite, designed it to test both the logic (function) and the robustness of the systems under test. The company wanted to build a model-based simulation without the high, ongoing cost required to build and maintain a “digital twin” model. A digital twin model mimics every aspect of every component in a system, allowing interactions among them to be tested virtually rather than physically.

“Mazda brought a lot of automation to the test side from the production side and tried to emulate how the driver would interact with the system—which was difficult without a human in the loop,” Kubo said. “That is why they developed a robot arm.”

The test suite included a hardware-in-the-loop (HIL) test system, image processor, speech synthesizer, noise generator and GPS simulation. “A lot of the effort is focused more on system integration—making sure components work together correctly, not testing the individual component or just assuming all the components should work together when you assemble them based on the specifications given to the supplier,” said Kubo.

Mazda went one step beyond simply evaluating the function and likely long-term reliability of functions built into various components, however. It also tested the functions and robustness of the interface between human and machine, using the robot arm to manipulate physical controls and simulators to mimic voices of old drivers and young, male and female, speaking Japanese and English in turns.

“We needed to evaluate the robustness of the system to identify the extent to which instructions could be correctly recognized for these different variations,” according to a white paper.

The resulting modular system was able to save 90% of the time and effort required to test integration of systems, reducing testing costs by millions of yen per year, thanks largely to the flexibility and programmability of the NI platforms and the existence of Mazda’s internal Electronics Testing & Research Group.

Mazda wanted to apply a model-based design approach to try to eliminate the need for physical testing, and shift as much into virtual testing as possible.

“Getting real human input makes some of that complicated,” Kubo said. “Another issue is the fidelity of the model. We tried to get the model as close to 100% accurate as possible compared to the physical environment. It’s not possible to get to 100%, but as long as you can test the right functionalities well enough, it’s okay to have a model that’s 70% accurate or 80%. The situation many customers are facing is that they have to either outsource all of it or just do manual testing, which is insufficient. Some of what we did was not sophisticated, like the robot arm, but it was effective at pushing a button the way a human would push a button without interrupting the simulation or having a human have to get there and do it at just the right time. It worked well, which is what matters.”

Simulation and ADAS testing
This kind of experimentation is essential in assisted and autonomous vehicles. Road testing is still considered essential, but it’s also limited. While some companies claim to have millions of miles of real-world testing, some experts say it will take billions of miles before there is enough data to standardize on what works and what doesn’t.

In the meantime, the best solution is more simulation and modeling. The key is making sure that all of the electronics in a car will be robust enough to handle stresses from extreme heat and cold and vibration, as well as dealing with electromigration and other electrostatic effects with circuits that may be “on” for extended periods of time.

“The automotive test flow is directly related to complexity,” said Anil Bhalla, vice president of marketing at Astronics Test Systems. “The big question now is what we need to shift around. What we want to do above all is catch failures. But if you characterize a device from -40°C to 150° C, most of that effort is around qualification, not production. So now you need to deal with that in the flow or in some final test or wafer or module.”

This is particularly complicated in the autonomous vehicle world because chips are both analog and digital. There are power semiconductors, mixed signal and MEMS sensors, as well as digital logic and memory.

“So now the question becomes which vehicle will use which product, because in automotive they typically want two or three suppliers,” said Derek Floyd, director of business development at Advantest. “And with power and analog, there are low power and high power versions. And then you get to AI, which is new for automotive.”

This is good news for the test industry. Automotive is expected to be one of the big drivers projected for test equipment over the next few years. Frost & Sullivan estimates the global market for electronic test equipment, now at about 150 companies that generated $10.8 billion during 2017, will rise to $13.16 billion by 2022, growing at 4% CAGR on average.

Simulation is expected to benefit, as well. All of the major EDA companies offer a range of simulation tools and services for highly automated vehicles. In addition, Mentor’s parent company, Siemens, reinforced its product lifecycle management (PLM) simulation software in 2017 with the acquisition of TASS International, which focuses on ADAS testing. The range of EDA and verification/validation tools from Mentor build on its own PLM simulation testing products.

“Even at the lower, component level you need inputs from other systems, whether that means sensors with data on obstacles or something else,” said Mentor’s Hand. “You have to pretend with something that approximates the right data to get that part of the simulation right. On the output you have the actuation, which involves more than just the electronics—you could have data from sensors that goes to the chip and an actuator that needs to use the braking system so it can affect what’s going on with the car. You have to simulate that whole loop, but the complexity goes up an order of magnitude beyond the verification/validation level when you start to look at all the system in the car.”

The changes in the automotive supply chain are just the beginning. The challenges in this part of the market are so new that it’s hard to predict how far the pendulum will swing for automakers taking on a greater role. In the past, carmakers like Mazda would not have to go to such lengths to build their own physical/digital testing environment.

But it’s also not a simple process to automate. Using approaches such as a digital twin and simulation-based system integration are still too new in the automotive world to have well-established best practices or easy-to-use tools to build the models and test processes.

“It’s something we’re still trying to understand,” Hand said. “It’s a complex problem people have not been able to break down enough to do something that should be as simple as tracing failure modes throughout the system. There’s no hierarchical approach to propagate that information upwards. For now, it’s a tough problem, but for the most part, it follows the same model of other things we’ve been able to develop as the industry has advanced, so we’re pretty optimistic.”

—Ed Sperling contributed to this report.

Leave a Reply

(Note: This name will be displayed publicly)