Is ADAS The Edge?

Debate is brewing over whether ADAS applications fall on the edge, or if they are better viewed squarely within the context of the automotive camp.

There is more to this discussion than just semantics. The edge represents a huge greenfield opportunity for electronics of all sorts, and companies from the mobile market and from the cloud are both rushing to stake their claim. At this point there is no single instruction-set architecture that owns this space, and it’s not clear exactly how this market will be carved up. There are end devices that can do pre-processing, and various levels of servers and local clouds before data is processed in the large cloud operations of companies such as Google, Amazon and Microsoft.

Whether it is referred to as adaptive driver assistance systems or advanced driver assistance systems, ADAS is a ubiquitous term in the semiconductor industry lexicon today. And with so much technology development focused on IoT and automotive applications, this is a big growth opportunity. But how exactly compute architectures evolve to handle a growing volume of data generated by sensors in vehicles, including streaming data from multiple cameras, radar and LiDAR systems, is not well defined at the moment. And that, in turn, raises questions about where the edge starts and where it ends.

“The car is definitely becoming an edge device,” said Mike Fitton, senior director of strategic planning at Achronix. “You are collecting data from different things. LiDAR and radar produce a huge amount of data, so you need to do edge analytics. You may have as many as five LiDAR streams and eight radar streams. You also have machine-learning algorithms for positioning, and all of this needs to have very low latency. So you want to do more of the processing at the edge, whether that involves ResNet or YOLOv2, YOLOv3, plus 5G for V2x and V2V.”

So where exactly does the edge start and where does it end? Opinions vary greatly, and automotive appears to be one of the crossover points.

“The edge is between the IoT device and the cloud,” said Lip-Bu Tan, president and CEO of Cadence. “It’s a mini-cloud, but it’s not so massive and it will be energy-efficient. There will be an automotive cloud and different vertical clouds.”

Chris Shore, director embedded solutions, Automotive and IoT Line of Business at Arm, has a somewhat different view. “The scale of the compute required to implement something like ADAS takes it out of the realm of edge compute. That said, unless you do the compute in the car, it doesn’t work for a ton of reasons in latency, security or safety. In that respect, it’s almost the ultimate example of edge computing. It does not work unless you can do the compute at the edge. In that respect, it’s a tricky question. Still, you’re putting into the car a serious amount of compute—more than you and I have on our desks, compressing all of that, putting it in a car. That’s serious edge stuff. And to me it’s so specialized, I would almost treat it as a separate field.”

Others look at ADAS in the context of a more traditional system that semiconductor designers are familiar with.

Jeff Miller, a product marketing manager at Mentor, a Siemens Business, sees ADAS as a system, and at the edge of that system are sensors and actuators. “If you’ve got an imager with a built-in computer vision/machine learning model built into the camera, then I would consider that edge because processing is happening right there at the camera module. If it’s all coming back to a big central processing system that’s taking streams off multiple cameras and multiple sensors, I would say that’s the cloud, or at least the gateway. Obviously, these things also connect up to the cloud to get traffic data and other sorts of things, but I think of that as the gateway tier if you’re thinking about that as an IoT model.”

But then again, the ultrasonic range-finding sensors, LiDAR, and radar are all complex devices themselves, and often do quite a bit of processing, so those would be classified as intelligent edge devices, Miller said.

Given the different regulatory requirements for automotive applications in general, it is easy to understand how most people think about these types of applications differently. “Call it a mission profile, but the required operating temperature ranges, and the safety certifications that come along with it, can be different even if you’re deploying something that seems similar. You take that same thing and deploy it to an automotive use case, it comes with a different binder of paperwork than if you’re deploying it into an aviation use case, a security use case or something else. This is an important consideration for companies to keep those kinds of multiple applications in mind, particularly as developers try to preserve flexibility to try to make sure that their design is useful in a range of application areas, or target markets for these types of things,” he said.

Generally speaking, whether an automotive application should be classified as an edge application depends on what is being done with it, as there is a trend in the automotive industry that says vehicles are basically edge devices and they can return data to a central server about things that they see on the road, said Marc Greenberg, product marketing group director at Cadence.

“You might have a system that can mark out potholes, objects on the road, things like that,” Greenberg said. “Perhaps the first vehicle crashes into it, but the second vehicle goes past it and may have been able to avoid it. In that respect, that is a system where there’s a huge amount of data being generated in the vehicle being processed. Then there’s an abstract of that data that’s going up to some central server, such as, ‘I was expecting to see the road layout look like this, and now there’s a new object in it,’ and upload that to the central server so the next vehicle that comes by can anticipate that extra object already being there.”

Geoff Tate, CEO of Flex Logix, meanwhile views the edge as anywhere outside of the data center. “A server in a Walmart is in the edge. A car is in the edge. Everything’s in the edge other than inside of the giant data centers. So a car is definitely an edge device by that definition, and that’s going to be one of the biggest drivers of compute. There’s a lot of silicon that’s going into cars now.”

Here too, there are still lots of tradeoffs. “If you’re inside the car, you know where the cameras have to go, but there is a debate as to whether the processing should be right at the camera or whether it should all be in a central location,” Tate said. “Comparatively, if you have surveillance cameras in a bank, should you have every camera have its own inference, or should you have the server where the cameras are already all wired because they’re recording all the information for security purposes? Should you put the inference in the server? In all edge applications, there are still lots of decisions and tradeoffs people will make.”

Conclusion
Automotive challenges many traditional rules of the game for semiconductor developers, and it is changing at a time when compute models are undergoing rapid changes, as well. But regardless of definitions, processing at the edge, near the edge or at end devices before the edge all point to the need to prioritize and parse data processing. How that gets labeled will likely change over time, but when definitions are confusing, that usually is accompanied by uncertainty in the technology.

“The important aspect of this is that there are people making sensors that go into these systems, and those sensors have built in intelligence, and they are designed and built for a very specific purpose,” said Mentor’s Miller. “That process of bringing the right design for that market is something that we see over and over again where engineering teams are targeting a very specifically, i.e., this isn’t a general purpose device, this is for space applications, this is for automotive applications, etc., and while they try to preserve component reuse as much as they can, the products that they’re selling are really targeted at those specific applications. For that reason, it is imperative that the right trade offs are made for each application. Nobody wants radiation hardening in their automotive parts, they want functional safety.”