VB: As far as ARM’s role in some of these things — Nvidia announced a relationship with Bosch yesterday. They have their car supercomputer chip, which is ARM-based. It seems like you guys are going to have a relationship in there, but do you need to be part of this ecosystem in some way?
York: More and more, yes, for the simple reason that the car-makers need a longer-term view of where electronics and software and technology and security and safety are going. If all they do is rely on their supply chain to tell them that, they get a very restricted view of what’s possible. They get a very short-term view as well, based on what they want to sell them directly. Whereas we can pay much more of an independent advisory role for the car-makers. “This is what’s possible. This is what other people have done.”
A great example of what’s possible — these things solve so many problems that the auto industry is grappling with. Security, for example. These things aren’t unhackable, I would say, but they’re very hard to hack nowadays. Compare that to a car. Cars are scarily easy to hack in practice. We can directly, with the car-makers, give them a view of what the rest of the electronics industry does and how that solves the problems the car industry has. We’re in a good position to do that because we’re independent.
VB: People are coming up with different adjectives for the internet of things. “Internet of stupid things” is one of them. Somebody’s going to show a smart toothpick soon.
York: The internet of pointless things.
VB: “The internet of expensive things.” “Why do we need the internet in that thing?” A $200 pet feeder replaces a dog dish. “The internet of hackable things.” Cars come to mind there. Charlie Miller gave that nice talk at the ARM convention about hacking the CAN bus. I wonder what that says about the state we’re in, the progress we’re making.
York: It does vary, I think, depending on the car-maker again. Some of them are grappling with these things up front. Some are in denial. Vehicle security is a good example. Some of them are really saying, “Root and branch, how do we secure the whole vehicle? How do we make sure there’s no easy way in? How do we make sure that, if someone manages to hack it, how do we quickly update the software?” Others are just saying, “Oh, well, maybe we’ll deal with it somehow eventually.” It varies a great deal.
Don’t forget, a lot of the car-makers, the senior management at the car-makers, don’t come from an electronics background. They’re from a mechanical background, because cars have historically been mechanical systems. They come from a world of managing production. Cars are the most complex object you’ll ever buy in your lifetime. 30,000 parts in a typical vehicle, roughly. They come from that world of managing complex production and making sure that one missing screw doesn’t stop the whole production line.
That’s a very different set of problems compared to understanding how electronics and software can change your product, make your product more flexible and innovative. They have to change the way they think and make sure people in senior management understand electronics, rather than leaving that to their suppliers.
VB: Is there a group effort that has to happen to get the hacking situation under control?
York: Absolutely. The auto industry needs to settle on some standards, and settle on them quickly. Security isn’t differentiating. You don’t typically sell a car based on the fact that it’s less hackable than the next one. It’s just an expectation. If you say you’re less hackable now, does that mean you were more hackable before? You can imagine the conversation. Consumers don’t want to buy products based on the security level in general. They just expect things to not be stealable. It’s a must-do.
One thing we’re trying to encourage the auto industry to do is stop ignoring the problem, start talking to each other, and come with consistent standards that aren’t manufacturer-specific or region-specific. The problem of securing a car in North America isn’t any different to securing it in China or Japan. The same problems — a large distributed system, a lot of ECUs, a lot of communications buses. The way you secure those, and the problems you have to solve, are no different. They should stop trying to do things their own special way and start agreeing to do things in some standard ways.
We’re contributing to some of the current open discussions, like NHTSA. NHTSA did a consultation a few months ago on automotive security. We made some submissions to that. We just think that they need to start requiring that cars be tested for security. We need to encourage them to stop thinking about solving these problems in a proprietary way. A lot of the technology isn’t differentiating, as I say. Security certainly isn’t. It’s just a must-have.
We’re in a good position because we can secure microcontrollers. We can secure the telematics units. We can secure the IVI units, the powertrain units. We have security technology for the entire vehicle, if people would embrace the idea of having roots of trust and all the basic stuff that other electronics have had for years.
VB: I’m curious where you see some of the startup activity here. Nvidia’s CEO just said they deal with 1,500 AI startups now. That’s exploded in the last few years. It seems like auto logically would have a lot as well.
York: There is. In many senses they’re leading it, because of the amount of money that they’re investing. Toyota has talked about a billion dollars in its Toyota Research Institute. There’s an awful lot of money going into machine learning, computer vision, and sensor fusion directly from the auto industry. If you look at computer vision, that has a lot to thank the auto industry for. It’s gone from a university research activity to the mainstream. It’s pushed from being something that’s reasonably reliable to something that’s increasingly bulletproof. The auto industry is kickstarting a lot of machine learning research in many ways. Obviously the Googles of the world are doing a lot in a particular way, but the auto industry is driving it as well in its own way.
Probably one of the most common discussions I have with the car-makers is about machine learning and computer vision, how we can help them deploy systems with reasonable cost and reasonable power when they finally get the systems ready to go into production. It’s not going to be tomorrow. It’s going to be 2022, 2024, 2025. But they need to plan ahead. That’s a common conversation we’re having right now. How can they deliver the enormous amounts of compute they need at a sensible cost point, with all the safety and reliability and security aspects?
VB: Is there an analogy where a house-builder has a general contractor, and he has all these subcontractors? Somebody has to deal with all those startups. I’m not sure who in this whole chain gets that job.
York: The car-makers are taking a very hands-on approach to this. Some more than others, but some are very hands-on. It’s going to define how their products look in a decade. They don’t want and can’t afford to leave that to their suppliers. It’s quite hard to specify what you want. So they’re taking a very direct approach.
Some of them are investing quite heavily as well. You look at venture funds, most of the auto industry, the automakers, have venture funds. They’re very busy investing in lots of these little machine learning startups. There’s a company that’s set up in the U.K. called FiveAI. It looks very interesting, and it’s being invested in by some of these venture funds. It’s a way of hedging your bets. It’s such a complicated and tricky space, making vehicles drive themselves, that you need to back a lot of different horses until you work out which ones are going to be successful.
VB: What do you think is new for ARM in automotive this year, compared to a year ago?
York: In some senses it’s a very slow-moving space. But the thing that really seems to have changed in the last few months is the car-makers’ interest in electrification. We’ve done fairly well in terms of being designed into some of the early hybrid and electric vehicles. We had quite good motor control solutions through some of our silicon partners. But there does seem to be a bit of an acceleration there.
Electric vehicles and hybrids are tricky, because they’re more expensive to build than a conventional car. The car industry is very cost-sensitive. A lot of people will buy a car very much based on the price tag. Having a hybrid vehicle is quite nice, but if it doesn’t pay for itself, that’s tricky. What’s changing is the car industry is realizing it needs to invest more. In some cities they already say they’re going to start banning diesel and internal combustion engines in city centers. The car industry is realizing that if it doesn’t start finding ways of building electric and hybrid vehicles at the same cost point as a conventional car, they’ll be in trouble. They’ll be forced to sell these things, but they can’t sell them at the same price point. They’ll make less money or they’ll have to subsidize it, which the car industry can’t afford.
That definitely seems to have picked up in the last few months. But the auto industry never changes that quick. The decisions it makes today won’t enter production until 2020 at the earliest. It takes that long.
VB: The high end of the space for ARM, with the biggest chipmakers — I wonder how much of it is ARM-optimized designs versus more custom versions. Like the Qualcomm 835. I’d heard that their cores they came up with more like ARM-optimized cores, as opposed to completely ground-up custom Qualcomm-engineered cores. It made me wonder what sort of interplay is happening there. One mentioned benefit is that if it’s ARM-optimized, the customer of that big company gets the benefit of all the architectural changes ARM is making that come in faster. If it’s a one-off customized version of ARM, that won’t necessarily get passed through as quickly.
York: It swings around to this, certainly. One of the nice things about having companies like Qualcomm build their own implementations is it brings in a healthy level of diversity. That’s been a very positive thing about companies like Qualcomm having architecture licenses and doing their own implementations. Nvidia is no different. They build very high-end ARM CPUs. In that sense, particularly when you’re building for very specific markets, that extra flexibility is useful for them.
Often those companies would combine their own stuff and our stuff into the same chips. They get the best of both worlds. They get to pick our implementations where they don’t feel like they can add any extra value by doing it themselves, and they can combine that with their own stuff. There’s a new middle ground where companies can take our product and add some targeted enhancements. It gives them the ability to, as you say, keep up with the latest architectural stuff we do, but still make some adjustments for a particular end product they want to build.
VB: Something more like semi-custom?
York: I can’t remember what my CPU colleagues call it exactly, or how we’ve branded it. But it’s that middle ground. People have asked to do this for years, to be honest. It’s taken us a while to work out the best way of allowing people to make modifications without breaking things. For us, the compatibility is sacrosanct. The moment you get something that’s incompatible in the market, our ecosystem starts to break up. We’ve always been a bit cautious about letting people modify.
We’ve ended up finding a nice framework where they can make modifications in sensible areas, without risk of breaking things. I quite like this new mid-point. In the end we have to be as relevant as we can to our semiconductor partners. That’s why we did architecture licenses in the first place, because there were some that said, “Look. You’re not delivering quite what we need. We’d love to use your architecture, but you need to allow us to do our own thing.” OK, we’ll do that, because we can’t do everything.
That’s the whole premise of the ARM architecture. We can’t do everything. We started off because we couldn’t do our own manufacturing. We licensed. This is an extension of that. We can’t build every possible variant of every possible ARM CPU, so we want to enable our architecture licensees to do that.