SDVs And AI Forcing Big Changes In Automotive

The automotive industry is undergoing a fundamental transformation that includes everything from software-defined vehicles, the injection of AI into nearly every facet of the design and use case of a vehicle, and a complete overhaul of traditional relationships between different tiers and OEMs.

The switch to software-defined vehicles is a top priority for the automotive ecosystem. It enables faster time-to-market and faster updates, keeping vehicles current with new protocols, standards, and features. With a traditional hardware-defined approach, flexibility is more limited because it depends on fitting those features around pre-defined hardware. It’s also more time-consuming and expensive, which has put traditional carmakers at a competitive disadvantage.

“There are some obvious technologies and concepts, like continuous integration/continuous deployment (CI/CD), DevOps changes, and the like,” said David Fritz, vice president of hybrid-physical and virtual systems, automotive and mil/aero at Siemens EDA. “Virtual platforms are playing a big role in that. At the same time, they’re still trying to wrap their brains around multi-fidelity virtual platforms in terms of what it means. From our perspective, it’s opening the door to introduce the work that we’ve been doing over the years in terms of the virtual platforms of all fidelities, pulling all that together into a framework that allows verification of super complex systems at different levels of fidelity. How do you know that when you put all of these simulations together in a full virtual car that it’s actually meeting your requirements? How do you know that as you continue to refine your software architecture, your hardware architecture, that it’s actually making things better and that it works? How do you measure that? You don’t have to wait for the car to exist to do it. If you want to realize the software-defined vehicle promise, then you need a combination of multi-fidelity virtual platforms and verification of those platforms.”

For the automotive industry, this is a radical change. “Technologies like zonal architectures, chiplet-based computing, and ADAS/AD are gaining traction as OEMs accelerate their shift toward software-defined vehicles,” said Adiel Bahrouch, director of business development for silicon IP at Rambus. “The value chain is navigating challenges around in-vehicle cybersecurity, supply chain security, regulatory compliance for market access, open-standard APIs, and system integration — especially as the pressure to deliver personalized, secure, and scalable mobility intensifies, while time-to-market shrinks at an unprecedented pace.”

Alongside these shifts is a relentless push toward increasing levels of autonomy, as well as the digitization of the cockpit to improve the user experience. Vehicle architectures are also starting to evolve to a more centralized system management, and in electric vehicles, that includes efficient battery management systems to extend range and battery life.

While the rate of adoption of these changes can vary from one carmaker to the next, nearly all are heading in the same direction. “In the ADAS world, the architecture is still zone-based systems, and the move to central processing systems has been a slow adoption,” said Amit Kumar, director of product marketing and management for automotive, Tensilica Product Group at Cadence. “Some OEMs have advanced the centralized architectures in a few models, but most are still zone ECUs, or even more primitive ECUs like we had before the concept of zonal architectures.”

Robert Day, director of autonomous vehicles at Arm, began seeing changes in the automotive ecosystem over the last few years, starting with software-defined vehicles, distributing ECUs into different zones, changing the architecture in the vehicle, from going through distributed ECUs into different domains, with an eye toward centralized compute serving distinct zones.

“There was a lot of discussion about how the vehicle architecture is going to change, and how this will reduce cost,” Day said. “And because the wiring configuration won’t be as long and heavy, we can start using things like Ethernet for communication, and then the zones control a lot of the sensors and things within their zone. That changes the software architecture, and that’s where we were looking at the move to the software-defined vehicle, where the vehicle is defined by software, but it’s also updatable. Your vehicle can get better after it’s left the forecourt, because you can get updates and upgrades. And that was the big thing for the last three years or so.”

Arm’s part in enabling this was starting the SOAFEE initiative, which is centered on software. It details how to efficiently develop and test code using things like cloud-native tooling, as well as how to get that software to the vehicle, including technologies such as OTA orchestrators. These are technologies that traditionally have not been used in automotive, but they can help with the software-defined vehicle.

“Then we started to bring in virtual prototyping environments in the cloud,” Day said. “So you can develop and test with either relatively good performance or relatively high fidelity, depending on what you’re doing, and then have a high degree of confidence that your software is working before you even deploy it to the vehicle. That was part of what SOAFEE was trying to achieve. Bringing in virtual prototyping environments is fueling the whole shift left, where we’re able to develop and test our software long before hardware is available. Then, as you get to integration and deployment, it’s more predictable. And when the car has been sold, you continue that.”

Day sees the next generation of this as the AI-defined vehicle, building on the software-defined vehicle approach. “A lot of the workloads are going to be AI workloads, edge AI workloads running in the car, and that’s going to help determine whether it’s software-defined or AI-software-defined. That’s going to help the car of the future meet our requirements as consumers.”

Infineon also continues to see automotive customers migrate to SDV architectures to deliver a more personalized vehicle, over-the-air software updates and other features and functions in the vehicle. “AI also plays a significant role in the SDV era and will continue to make vehicles safer, more convenient and intelligent on the road—and off. Additionally, we see OEMs and our customers implement new safety, security and comfort, and some level of autonomous driving into today’s vehicles,” noted Bill Stewart, vice president of marketing, Americas at Infineon Technologies.

AI everywhere
The next challenge is to determine where else AI can play a role in the vehicle’s design and operation. “We’re applying AI to predict the performance of AI,” said Siemens’ Fritz. “Back in the day, we had YACC (yet another compiler compiler). You had compilers creating compilers. This is the same sort of thing. We’re using AI to prove the performance of AI. That’s a big thing that we’ve invested a lot into, and it’s really starting to bear fruit, to prove some of this stuff out long before you have any kind of silicon.”

AI also will play a significant role in the ADAS and autonomous driving space. There are new network models and new types of networks, and Cadence’s Kumar pointed to CNN-based networks like YOLO for object detection and SSD-ResNet for segmentation as proof points. “DETR-ResNet is coming up fast,” he said. “For vehicle behavior, RNNs are being used. And spiking neural networks are becoming popular for certain applications, especially within the vehicle. Cadence has been testing and running all kinds of neural network workloads across DSP cores, which contain MAC units. We also have an AI co-processor in our product portfolio that supports neural network engines to offload unsupported layers and operators with far less energy and power consumption.”

But unlike data centers, AI in a vehicle needs to be super-efficient. Level 5 vehicles can have more than 40 sensors and billions of lines of code, and if they burn up too much power, the range of a vehicle will drop significantly. Just designing a system that complex to be power-efficient is a huge challenge.

“To address this growing complexity while accelerating time to market, automotive manufacturers are seeking a scalable SoC architecture so they can quickly iterate on designs, scale systems up and down based on their processing needs, and deploy features that make modern cars safer,” said Mick Posner, senior product group director for chiplet and IP solutions in Cadence’s Compute Solutions Group. “Increasingly, they’re showing interest in chiplets to meet these needs. While multi-die is mainstream, chiplets aren’t yet, and the industry is at a pivotal point where we can assist with this transformation.”

Cadence announced a chiplet-based physical AI platform earlier this year, and is engaging with early adopters. The platform’s architecture includes a flexible base system chiplet, a configurable AI accelerator chiplet based on the Neo NPU, and support for a CPU chiplet. It also accommodates optional support for a fourth domain-specific chiplet. The platform’s chiplet framework facilitates chiplet communication, including SoC-level control of the security, safety, and control subsystems. Unified reference software is included for system bring-up and for development of production software use-cases. Last year, Cadence taped out its first Arm-based system chiplet, and currently has silicon in-house.

Fig. 1: Cadence’s Arm-based chiplet. Source: Cadence, Arm

Higher levels of autonomy require AI. “AI is figuring out what’s going on around the car so that it can make decisions as to what to do,” said Arm’s Day. “As you go up the levels of autonomy, there’s going to be more and more of that. There’s a lot of discussion right now within the autonomous driving community about having end-to-end AI powering it. So it’s not just the perception. It’s also the planning decisions, and then, what does the car need to do? We’re starting to see more use cases that could use AI, and people are putting it out there. The first step is ADAS. We need it. You can’t look around you and figure stuff out without having a lot of AI in the training, but also a lot of the inference in the vehicle. Now we’re looking at driver interaction with the vehicle.”

Changes in the cabin
AI is infiltrating the cabin, too, and it’s proving to be a differentiator for OEMs. “UX and UI are both part of cabin comfort and also enable safety functions with multiple sensors now within the cabin,” said Cadence’s Kumar. “One such emerging technology is in-cabin sensing, which enables the child presence detection feature recently mandated by Euro NCAP and is also a growth engine for automotive, both now and in the future. Driver monitoring systems (DMS) are getting more sophisticated, and penetration of such systems has been on an upward trajectory.”

Arm’s Day also is seeing AI use cases, like chat bots and LLMs coming into the cabin, so that the user can communicate with the vehicle, as they’re used to doing at home now, with a sort of Chat GPT interface. “There’s a lot going on around voice recognition in the cabin,” he said. “Then there are things like the DMS that will use AI to figure out what it is you’re doing as a driver, or an occupant monitoring system. All of that is going to start using AI. We can’t imagine all the workloads right now, but we’re going to see some new and interesting things coming up. For example, your owner’s manual for your vehicle. There’s no reason you should ever have to read that. You should just be able to ask the car what the tire pressure is on your tires. That’s a relatively straightforward, interesting use case. Then there’s the control of your infotainment system and your navigation system. All of that will have the same sort of things that we have in our homes. And in fact, there’s quite a lot of talk right now about having, essentially, the AI agents that you use in your phones and in your home follow you into the car, which is almost context-sensitive AI. It knows who you are. It knows where you’re going, for example. And then you can, as a driver, start asking it questions, like, ‘How long is it going to take me to get to Starbucks? Can you order me a latte?’

Robotaxis provide another avenue for AI. Relying on strong and sophisticated perception sensing and high-performance compute, robotaxis are becoming mainstream in some cities. “So far, this segment has been led by Waymo, but Tesla is kickstarting a robotaxi program in the United States, and Uber and Nuro are partnering to form a robotaxi fleet,” Cadence’s Kumar said. “These are clear indications of this new trend in automotive Level 5 vehicles.

Automotive security
As vehicle connectivity and complexity increase with autonomous driving, vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) deployments, so does the attack surface. “Preventing remotely exploitable vulnerabilities in vehicle update, V2V, and V2I is critical, along with preventing extraction of proprietary autonomous driving algorithms by competitors,” said Nicole Fern, principal security analyst at Keysight. “This has motivated the adoption of hardware security features such as secure enclaves and dedicated roots of trust.”

However, the inclusion of more security features does not automatically make an automotive solution secure. “Hardware attacks such as fault injection and side-channel analysis have successfully compromised systems utilizing hardware-backed security features, such as secure boot,” Fern said. “A good example of this is the Tesla autopilot hack from TU Berlin, which uses fault injection to extract code and data from the system, along with cryptographic keys used for authentication with Tesla’s back-end infrastructure.”

Fortunately, people are taking security in the automotive environment very seriously now. “They now understand that it underpins automotive safety, i.e., ISO 26262 compliance,” said Mike Borza, principal security technologist and scientist at Synopsys. “Without a secure environment, you don’t have ways to make guarantees about safety and the safe behavior of systems. That’s very similar to how it’s being used in the AI world. It’s also true that you have the conflation of these two things now. People are using image-based processing to determine where vehicles are, how they’re following other vehicles. In the case of reactive or dynamic cruise control systems. So now you have these things rolled up together, and people are going to go down that path in the automotive space. We’re also seeing significant care around maintaining the integrity of data streams within the automobile and moving data from one place to another in a way that they’re using technological means to enforce integrity, such as encryption and authentication — technical authentication in the sense of verifying that the information is intact when it’s received. And so the cameras that are using MIPI links, for example, are now incorporating MIPI security into those camera streams. It’s the same thing with the displays, because if you modify the data to make things that you should be able to see disappear when backing up your car, that’s a big problem if the vehicle is using a camera instead of a rear-view mirror, or instead of actually looking over your shoulder.”

Conclusion
All this technology is creating an upheaval in the automotive industry, and there is no sign it will slow down anytime soon.

“A lot of things we predicted, like cozying up with OEMs so there’s a tighter relationship with the OEMs, is happening,” said Siemens’ Fritz. “The concept of OEMs driving their ecosystem, rather than the way it was, it’s really coming true, even though in different regions it’s happening at different paces. One Japanese OEM, in particular, is pretty much locked in 100% on one Tier One, and that Tier One is locked in pretty much 100% with everything going on in a particular Tier Two. It’s almost as though the Tier Ones and Tier Twos are becoming captive to the OEMs, because that’s the only way they can control their future. And that’s interesting, because that forces everybody to get more focused and understand their contribution. That also opens the door for an awful lot of these smaller SoC and AI accelerator companies to squeeze their way in there. So we’re going to see consolidation in the next few years as those companies get rolled into these large OEMs, pulling in their whole ecosystem under their umbrella.”

This is both regressive and forward-looking. “It’s reminiscent of the Henry Ford days, which started when he took everything and put it right there in Detroit,” Fritz said. “He shipped in his own materials and produced everything and got completely vertically integrated, as opposed to distributed integrated, which is where we were recently. A Tier One would have a few OEM customers and try to leverage the same thing into each of their OEMs, and that’s how the OEMs lost control. What’s happening now is turning that upside down, back to the OEM saying, ‘I’m fully vertically integrated. I get to dictate all of it.’”