To get autonomous vehicles from prototypes to production Arm believes we need a consortium of companies to come together on standards on computing a safety. That’s why Dipti Vachani, senior vice president of automotive and embedded at Arm, announced the Autonomous Vehicle Computing Consortium during a keynote at the Arm TechCon 2019 conference in San Jose, California.
The consortium includes General Motors, Nvidia, Denso, Toyota, Bosch, Arm, Continental, and NXP. It will focus on collaborative projects such as improving safety, security, computing power, and software. No single company can do all this, nor can it convince people that taking cars on the road is a safe thing to do.
Vachani said a lot of people — including, like herself, mothers of 16-year-old drivers — are worried about the safety of both human-driven and autonomous cars. She said that, as a mother, the stats about driving deaths mortify her. And Forrester Research’s data shows that autonomous driving experts are all worried about the same thing.
“As all of you already know, a mother’s instincts are always correct,” she said.
Everyone in the self-driving car ecosystem will have to optimize workloads for computing devices such as central processing units, graphics processing units, image sensor processors, and machine learning. The group will enable companies to do things such as pre-vet applications still in development. After her speech, I interviewed Vachani about how we’ll move from prototypes to production with self-driving cars.
Here’s an edited transcript of our interview.
VentureBeat: I was interested in the automotive alliance there, the consortium. Why was it those particular companies that joined in first?
Dipti Vachani: We did reach out to a network of folks. Those companies have seen the problem in which the solution that we have today is not built for autonomous. The power and performance and cost is way outside of–the solution most companies have today is not autonomous because for the power and performance, the cost is way too high. They recognize the platform is going to have to come to–each one of them can’t continue to invest at the levels they have independently.
Coming together, we can start to solve some of these problems. What’s a common OS? What’s a common hypervisor to use? What’s common hardware? They’re not going to, say, decide on a chip, but they’ll decide on a platform that they can then build off. It’s a recognition of maturity, understanding that each one of these platforms independently just costs way too much to develop.
VentureBeat: Was there enough of a critical mass here to announce the launch, then?
Vachani: Right. We need to create the by-laws and how the consortium works. We had to go through all the effort of ensuring that consortium is truly a consortium. We have a board of directors now and a chairman of the board. We have a governing body and by-laws in place. Everything is signed off. That’s the right time to announce it.
VentureBeat: It’s an interesting day when the car-makers care about chips this much.
Vachani: It is interesting, considering that–they’re starting to feel a bit like they’re going to have to get engaged at the level, to ensure that they continue to stay competitive.
VentureBeat: Are there more problems that are becoming obvious if this doesn’t get done? If there are separate silos of technology that get developed–are there examples you can already see?
Vachani: More so what we’re worried about is–the cost of building a chip of this magnitude, and writing software of this magnitude, and the volumes we see, they don’t add up. The equation doesn’t work. It’s going to take time before this sees some volume. We recognize that we’ll be stuck in this paradigm shift, this terrible circular logic. “I can’t make it, but then I can’t build it, but then I can’t write all the software.” It never gets to market.
VentureBeat: Just one prototype after another.
Vachani: We’re trying to break that. We’re going to have to look at this holistically, and we’re going to have to share the cost across the board. No one company can afford to take all of this on. Each company is looking at areas where they can differentiate. We recognize that everyone is going to have to find a solution that they can differentiate on. But there are some common elements, and if we don’t share the cost of that across the industry, it’s just not going to be effective.
VentureBeat: I would have thought this had been done by now. It might have happened at the beginning when people first started talking about autonomous cars, but we’re so far into it now.
Vachani: “Far into it” is a hard one. We’re very far into prototyping. Do you see anyone out there in mass production? Right. We’re very far into prototyping, and that prototyping is fundamental and essential. Don’t get me wrong. That’s valuable time we need to spend to develop code and understand. Any learning engine requires learning and learning means time. Learning means time on the road, understanding the different variables. All of that is happening today, which is great. But that isn’t going to solve the cost problem at scale.
VentureBeat: Is there a first problem you have to tackle? Are you still sorting out how the consortium will work?
Vachani: We’re in the early stages, but we have started to create working groups. One on hardware and distribution of workloads there, where these workloads sit, one at a system level, and one on software. We’re starting to build the right working groups, and then even after announcing I have had a whole slew of people approach me that want to join. I’m sure that we’ll start to have additions to help out in these working groups.
VentureBeat: Is there a kind of architecture that is, on a general level, the right way to go yet? Particularly when it comes to what you do in the car versus what you do in a data center, whether you want to rely on anything that goes out toward a cloud or not.
Vachani: No, we’re still in the early stages. Though I will tell you, our personal belief is that it’s going to be a combination. It’s not one or the other. It’s not all in the cloud nor is it all in the car. That’s not possible either. There will have to be some level of communication between the two. A lot of it’s going to depend on how quickly you can make decisions and what data you need to make decisions. Latency is going to be key in determining where the stuff goes. But we recognize that it’s a combination of the cloud and the car.
Then, in a micro-universe, it’s the same kind of thing. If you think of central compute versus what’s at the camera, the same thing. What goes on in the camera could be a latency problem. What independent decisions can we make? Or you need to balance that with the central compute that needs to know what’s going on. Those are the kinds of discussions that we’re starting to have. We recognize that neither one extreme makes sense.
VentureBeat: What about the level of autonomy you want to attack? Complete self-driving is going to be very interesting to get to, but driver assistance seems like it’s making great strides right now.
Vachani: Absolutely, it is. Driver assistance will continue to grow. Today ARM is already a significant player in that. 60 percent of ADAS systems are ARM-based today. We already have a good position there. We’ll continue to develop that. Those are the AE devices that are going into that, so the same technology is used. We’re very happy to have that traction. That will continue.
What we recognize and are saying is, and what I was trying to express in the keynote, the moment you remove the driver completely, the problem changes. It’s no longer incremental. Driver assists are incremental developments. That will get better and better. But the moment you remove the driver and the backup plan is gone, that’s a significant change in paradigm. That change has to be looked at differently from the ground up, and that’s what the AVCC is trying to tackle together.
VentureBeat: It sounds like you probably don’t believe that any of this stuff is happening very soon, then. Every time I go to CES it feels like it’s like self-driving cars are right around the corner.
Vachani: Tomorrow, yeah. You know, it’s hard to know. Several reports will say it’s never going to happen. Some will say it’s happening soon. Some will say it’s somewhere in the middle. If I could truly predict that, I’d probably be in some other job. We can’t.
What I talked about today is the technology challenges. There are also other challenges — government, society, people’s comfort level. You also have regional concerns. Each region around the world has its own solutions. There are too many variables for us to come up with an accurate date and say, “This is when we believe it’s going to move.” But we believe at ARM that if we can start to bring the industry together and start to solve some of these problems, that will work itself out. We can control the technology, and that’s what we’re working on. There are lots of other aspects to whether this will go into production or not that are outside of necessary technology concerns, of technology control.
VentureBeat: There was another interesting thread about custom instructions and the desire of the partners to have creative control and be as independent as they want to be. It almost sounds like it’s going in the opposite direction of what this consortium wants to do.
Vachani: Let’s think about this. These are two different things. They’re not conflicting, but I can see at face value why they may seem that way. Let’s talk about the consortium. The consortium is not hardware-specific. It’s trying to solve the whole system-level solution of autonomous cars. That includes software and how we distribute workloads and things like that. That’s the consortium.
Then there’s IoT and small IoT devices, where you may have something you want to do really fast and really optimal and you have a custom instruction to go do that. That’s a different problem. That’s a real hardware-specific problem. Well, let me be more clear. It’s a software/hardware mix. Your software is driving what maybe the hardware needs to do, and from the software development we now know that if we optimize these instructions we’ll get better power efficiency, better cost, and it’ll be a more optimal solution. We have very small power windows. We care about cost and power consumption.
That’s where custom instructions make sense, and that’s why it’s starting with our M portfolio. Often that’s in storage devices or small IoT devices. Maybe even deeply embedded, where you know exactly what you’re doing and it’s a fixed-function thing. That’s a whole different world from autonomous cars, as you can imagine from just the scope of technology and power.
VentureBeat: In that world, I guess those customers have an option. They could go off toward RISC-V. What’s interesting for ARM is to offer the same thing that RISC-V could do, but also be careful about how wide you want to open this door. Is that something to think about?
Vachani: It’s not that confusing to us. It’s very clear. We’ll always honor our software ecosystem. That’s the value we provide — our software tools and flows, everything just works. This is why you engage with Arm. This the value you see in Arm. If we can allow flexibility while still honoring our software ecosystem and the fact that our tools just work and our flows just work, we will do it. In this case we found a way to creatively do it with these custom instructions.
VentureBeat: Without leading toward those problems of fragmentation?
Vachani: Yeah. We’re not causing any of that. This is isolated. We can continue our flow. We will do it. The moment that it causes fragmentation, the moment it violates our tools and flows in our software, that’s our line. Our line is pretty black and white. We believe that it’s what the ecosystem needs. It’s what our customers tell us, and so we honor that quite highly. We respect that ecosystem.
VentureBeat: The Arm Cortex-M33 line. Was that the obvious [processor to use for custom instructions] a particular reason?
Vachani: It’s the next one in line. It’s easy. It’s already used so predominantly. It’s something we can easily give free to everyone to start looking at. It’s low power. It happens to be an IoT application. It made sense. From this point on, every M will have it. We’ve gotten rave reviews on it. Every customer that we’ve shared this with has been extremely impressed.
What we’re trying to solve for is this three-pronged approach. We want to honor our software ecosystem and not create fragmentation. We want to be able to add custom instructions for very fixed-function-like things, while also providing the verification and stability–60 percent of my resources are used in verification, because when you get Arm it just works. We have to honor that too. Playing with this triangle, we have to keep all of it equal and make sure we respect that, because that’s exactly the value created when all of that comes together.
VentureBeat: I have a 16-year-old daughter as well, and she’s very excited about driving.
Vachani: [Laughs] You’re scared for her and for everyone else on the road with you?
VentureBeat: I said, “Wait a minute, you don’t have to learn. You can wait for self-driving cars to come along.”
Vachani: I’m sure that did not fly. First of all, I thought my life would get much easier. She’s driving, so now I no longer have to drive her around. I’m always the taxi lady, and so is my husband. Between the two of us we’re constantly driving around. But it actually isn’t. Life isn’t any easier. [Laughs]