L5 vehicles need at least 10 more years of development.
Are we there yet? Governments, consumers, and engineers alike want to know how close the automotive world is to producing a fully autonomous Level 5 vehicle.
While some experts say such vehicles could hit the road in the next few years, they’re a shrinking minority. Most forecasts say a truly self-driving car is at least a decade away — and maybe much longer, because it requires disruptive technology that has yet to be invented.
At stake are billions of dollars-worth of automotive and tech industry profits, as well as an untold number of human lives. More than a million people around the world die each year in road deaths according to the World Health Organization, though the extent to which such deaths are fully attributable to driver error and could be prevented by high-tech cars is up for debate. Still, the European Union is reportedly planning to approve the sale of Level 4 vehicles — those that are fully autonomous within a specific geographic area or set of circumstances — as part of its attempt to reach zero road deaths by 2050.
There are currently no such cars on the market. Level 4 technology is employed in Waymo ridesharing, a taxi service by the General Motors subsidiary Cruise, and a handful of other projects, but not in consumer vehicles. Level 3, which requires the driver to take over when the system encounters a problem, currently can be found in select Mercedes models in Germany and a small number of Honda Legends in Japan. Advanced driver assistance systems (ADAS) in Teslas and other high-tech cars available to consumers typically are classified as Level 2.
Level 5 vehicles, which are self-driving in all road situations, do not yet exist. Moreover, it’s not clear when that will change. Forecasts are all over the map. Earlier this year, Oculus VR consulting CTO John Carmack and Stack Overflow co-founder Jeff Atwood publicly wagered a $10,000 charitable donation that Level 5 would be commercially available to passengers in major U.S. cities by 2030, with Carmack betting for and Atwood betting against. Tesla CEO Elon Musk has made several statements forecasting a Tesla Level 5 achievement, but has yet to prove it.
Fig. 1: Different phases of driving automation. Source: Synopsys
One of the big challenges is that some very complex systems need to work together as a single integrated system. That is made even more difficult by the fact that this collection of advanced technologies are new and still evolving.
“Today, the market is driven by EV, ADAS toward autonomous driving, connectivity, and shared mobility,” said Marc Serughetti, senior director of embedded software and solutions at Synopsys. “The challenge for the OEM is, given that this is what consumers want, how do they start delivering a product that enables all those capabilities? In the past, the way it was done in automotive, was the OEM said, ‘Here’s a new function. Let me throw in a new piece of hardware that addresses this function.’ But with these new trends, you can’t act like this. It doesn’t work anymore. So, they can’t say anymore, ‘I need this functionality. Let me add a new electronic control unit that’s going to have this function.’ It cannot work because those functions do not work independently. They are connected and interdependent.”
So how can all of this be achieved, particularly when so many pieces are in various states of development, with a constant stream of innovations and consumer demands? “A lot of companies, like Toyota and others, talk about software-first,” Serughetti said. “It’s the mentality of thinking about software first, and how that impacts the product and delivery. For the hardware, there are multiple levels of implication. First, you cannot have the same architecture you had before. You cannot have an architecture where all those ECUs are distributed all over and each have their function. You need a change in the electrical/electronics architecture of the car. This is why people are talking about central compute, zonal gateway, and domain controllers. There’s an entire evolution on that side that’s happening, and that’s to support that software.”
Putting all of these new pieces together takes time, however. Paul Karazuba, vice president of marketing at Expedera, believes much needs to be done to get to full autonomy, starting with much faster and more energy-efficient processing of an enormous amount of streaming data. “Simply from a hardware perspective, the platforms that would be required to deploy L5 in a car don’t exist, and nor do today’s solutions realistically scale to meet L5,” he said. “L5 is going to require a certain amount of AI processing. It’s probably going to be, depending on who you talk to, between 1,000 and 3,000 tera operations per second (TOPS) of processing power. Today’s most advanced solutions for AI processing that one would find in a car is about 250 TOPS.
Fig. 2: The role of artificial intelligence in self-driving vehicles. Source: Expedera
At that level, says Karazuba, the processor consumes about 75 watts of power, which creates significant scaling problems. “If you look at the scaling that’s required, those chips will have to scale by a matter of 4 to 12X in performance, which means the associated scaling of power. Seventy-five watts is about as hot as you can run a chip before you actively need to cool it, and automakers have signaled they have no desire to actively cool electronics in cars because it’s expensive and it adds weight.”
Power constraints Even if it were possible to scale current solutions to 3,000 TOPS, he said, doing so would consume up to 1% of the battery on one of today’s electric cars. Karazuba said one workaround is putting multiple chips next to one another, but doing so creates both space and cost issues.
“With today’s silicon, you’re talking multiple hundreds of dollars — probably approaching $1,000 per chip — and you’d have four of those in a car, or maybe 12 of those in a car. You also can’t take a chip that’s already pretty darn big and grow it by 4X or 12X. It’s going to have to go to desegregated processing, either moving processing to multiple places within the car or moving to something like a chiplet model.”
That’s only part of the cost. All of these new features and capabilities require power, and in electric vehicles that power comes from batteries. But battery technology is only improving at a rate of about 5% to 6% per year, while demands for processing more data faster are exploding, creating a gap that widens as more autonomy is added into vehicles.
The batteries themselves are big variable. They are heavy, they take time to charge, and they have finite energy storage. So the best solution available today is to improve the efficiency of the electronics, and to simply add more batteries to increase range. That adds cost, as well as complexity, because batteries need to be cooled and properly managed, and they need to be protected and robust enough to withstand a decade or more of harsh road and environmental conditions. And while it may seem a long way from Level 5 driving, they remain one of the limiting factors for how much intelligence can be added into a vehicle — and another area where there is no obvious best choice.
“Battery modules are getting larger and larger, so you try to have higher density cells in the battery pack because they take up less space,” said Felix Weidner, senior staff engineer at Infineon. “The prismatic cell used to to be the easiest to package, and especially at the beginning, prismatic cells were used a lot. But we’re also seeing pouch cells a lot, and Tesla is using cylindrical cells. Everybody is trying to optimize this volumetric energy. But there are tradeoffs. A prismatic cell comes with a proper housing and you don’t need that much setup. You need much more rugged housing with pouch cells. So now do you save on the housing with a prismatic cell, or do you continue to push the limit on energy density? On top of that, the chemistry inside these cells is also constantly changing. The first BMW i3 battery had an NMC battery (a type of lithium-ion battery) with nickel, manganese and cobalt. Now we’re talking about LiFePO (lithium-iron-phosphate). And then they’re trying to put less electrolyte in there without reducing the conductivity too much, and they are trying to adjust the anode by blending some silicon inside to increase the surface area of the anode. All of this is constantly changing, and that affects how you make your package.”
Drawing too much power limits range of any vehicle. But a fully autonomous vehicle needs even more power because it needs to process massive amounts of data. “The number of complex SoCs that are going into vehicles today is growing really quickly,” said Paul Graykowski, senior technical marketing manager at Arteris IP. “We’re seeing an average of 23 now, and that’s likely to go up to 26 in the near future. That’s for Level 2 or Level 3. I can’t even imagine what it’s going to take to get to Level 5. It wouldn’t surprise me if the number of SoCs doubles. That means we’ll have to be very careful with power consumption, because power consumption is everything.”
Integration issues Another thing the industry will have to figure out, according to Graykowski, is car-to-car communication, so vehicles on the road can communicate with one another about position and speed and make adjustments accordingly.
David Fritz, vice president of hybrid and virtual systems at Siemens PLM, believes the industry will get closer to Level 5 around 2027 or 2028. “To justify the development and production costs for relatively low volumes — 100 million a year — you’re looking to get the biggest bang for the buck. We’ve been talking for years about ECU consolidation. Now they’re saying, ‘We’re going to build one chip. However, it’s going to replace 12 boards, 47 discrete integrated circuits, 17 SoCs, and it’s all going into one.’ Now we’re at that point where it starts to make good financial sense. And you’re eliminating having to work with 12 different suppliers for software, 12 for the hardware, so all those headaches just go away. That’s when it starts to become beneficial in terms of methodology, in terms of development, team size, and capability. And this moves toward vertical integration.”
But vertical integration and ECU consolidation won’t happen overnight. In the meantime, the number of ECUs likely will continue to rise as these activities evolve. For a model year 2024 vehicle, the number of ECUs is expected to be the same as today, or slightly higher, Fritz said.
“The OEMs need to almost take a half step back,” he noted. “They need to say, ‘I’m going to add this functionality, which is going to require yet another ECU. But the follow-on step is to integrate four of these other ECUs into that one.’ It’s evolutionary approach, because people are so afraid of being so radically different from what they’re used to. How do they get there in just a step or two and not lose further ground to the leaders?”
Even realizing Level 4 autonomy likely will require a “massive restructure of current technologies,” according to a June white paper from Rambus and Siemens. Full autonomy likely will involve automotive Ethernet, vehicle-to-everything (V2X) connectivity, and domain controller units. The chips that power these systems and others must be safe, secure, and reliable over the lifespan of the vehicle.
But creating such a chip is easier said than done. According to the paper, “compared to other commercial silicon designs, chips for automotive use cases face much higher environmental challenges, like temperature, moisture, and physical abuse due to vibrations. Therefore, robust, fail-safe, and/or fail-operational systems are paramount in automotive designs. Safety and security are achieved through multiple engineering activities and practices.” The authors contend that engineers must assume the silicon will fail, and therefore must design and verify products that will still function even when those failures occur. One possible solution in some cases is using CMOS nodes made specifically for the auto environment.
Of course, no expert can predict the future, and it is always possible that a major technological innovation in the near future could shift the timeline significantly. Such timing also depends on how one defines Level 5 autonomy within the guidelines set out by SAE. For example, does a vehicle that theoretically has the capacity to drive itself on any road mapped by a human, but only operates within a major city, qualify as a Level 5 vehicle? Or is that really a Level 4 vehicle?
Bumpy road ahead Regardless of how such technology is categorized, there are certain technical challenges involved with vehicles that must operate beyond trial areas. “For autonomous vehicles to be commonplace, and not just in initial trial cities, it will require much more sensor technology at an affordable price point, and more advanced software and computing capabilities,” said Robert Day, director of automotive partnerships for Arm’s automotive and IoT line of business. “One of the major challenges the industry faces when it comes to bringing autonomous driving to the mass market is delivering the compute performance required within the power, cost, and thermal constraints of the vehicle. In addition, as vehicle architectures evolve, developers face increasing code complexity.”
Much of this will be determined by the importance of different systems to the functioning of the vehicle. “As those computing systems become more centralized, there needs to be additional protection of the different criticality levels of the software that runs on these systems,” Day said. “There is also a component that helps with the deployment of fully autonomous vehicles, and that is restricting the operational design domain (ODD) that vehicles operate under. By restricting the geography, environmental conditions (rain, snow, fog, etc.), and the speed that the vehicle operates under, it reduces the complexity of circumstances the vehicle has to deal with, and hence can allow for a faster deployment of driverless vehicles on our roads.”
Frank Schirrmeister, senior group director of solutions and ecosystems at Cadence, said he expects a vehicle that operates according to the guidelines set forth by the Carmack and Atwood bet to appear in the next few years using existing technology. He predicts the infrastructure for such cars, however, won’t be available until 2030, and that consumers may be reluctant to adopt this type of autonomous technology until several years after that. Schirrmeister pointed to a Cadence report, which shows that while consumers generally believe hyperconnectivity will positively impact their lives, many dislike certain features associated with fully autonomous driving and some lower levels of autonomy.
“There is some hesitancy in consumer adoption when it becomes about control,” said Schirrmeister. “People want autonomy to be able to be disabled. They really just want advice. They do not want the car to do things for them completely. So mainstream adoption will probably take longer, and the numbers I’ve seen that look realistic put it at 2030 or 2040.”
Mainstream adoption also will require solving infrastructure issues. Arteris IP’s Graykowski said extremely “smart” cars will require equally smart cities to provide those cars with necessary information. “It’s going to be much more than a car thinking on its own. It’s got to be the whole city talking to the car, and the car talking back to it.”
Conclusion In the end, significant changes will be required for vehicles to truly become autonomous.
“You’re going to need chips that are very powerful in terms of computing, and in terms of what they enable you to do, such as AI support, and other advanced functionalities,” said Synopsys’ Serughetti. “That will require changes in the software, because now you have all those computing elements and you want to leverage that computing power, such that you may have a zonal architecture, for example. You say, ‘I have a different zone controller, and I’m trying to do a calculation. Why can’t I use the computing power provided by one of those zone architectures? And if I don’t have enough out of that, why can’t I use another part of the computing power in the car?’”
Even then, the timeline for autonomy remains fluid. A complex system is difficult enough to build, but when the functionality of that system needs to be integrated and interdependent on other systems, it becomes exponentially more difficult.
— Additional reporting by Ann Steffora Mutschler and Ed Sperling
Name* (Note: This name will be displayed publicly)
Email* (This will not be displayed publicly)
AMD CTO Mark Papermaster talks about why heterogeneous architectures will be needed to achieve improvements in PPA.
Companies are speeding ahead to identify the most production-worthy processes for 3D chip stacking.
New capacity planned for 2024, but production will depend on equipment availability.
Number of options is growing, but so is the list of tradeoffs.
Where neural networks are being deployed in semiconductor manufacturing and how well they’re performing.
Increased transistor density and utilization are creating memory performance issues.
Suppliers are investing new 300mm capacity, but it’s probably not enough. And despite burgeoning 200mm demand, only Okmetic and new players in China are adding capacity.
The industry reached an inflection point where analog is getting a fresh look, but digital will not cede ground readily.
100% inspection, more data, and traceability will reduce assembly defects plaguing automotive customer returns.
Engineers are finding ways to effectively thermally dissipate heat from complex modules.
Some of the less common considerations for assessing the suitability of a system for high-performance workloads.
Different interconnect standards and packaging options being readied for mass chiplet adoption.
Disaggregation and the wind-down of Moore’s Law have changed everything.