Tesla Says Its New Self-Driving Chip Is Finally Baked
Tesla’s Model S sedan, the car company’s flagship vehicle, was first shown as a prototype in 2009, has been on sale since 2012, and, barring one small change to remove the fake grille at the front, has looked exactly the same for nearly a decade. This is notable because most manufacturers fully redesign their cars every four to six years to keep them fresh—and to keep buyers buying. For Tesla, tech upgrades are the selling point. The company pushes software updates several times a year, adding features like summon, where a car pulls in and out of a garage with nobody inside, or camper-mode, for sleeping in the car with the heating on.
Tesla’s biggest claim is that one day, all the cars it’s currently building will be capable of full-self driving. So when Elon Musk announced after the company’s second-quarter earnings report that Tesla is developing its own computer chips, it was a momentous claim by the car company. Musk’s grand reveal (which maybe actually isn’t news given Musk said this at an AI conference in December, a fact WIRED reported on then) led him to boast the company has “the world’s most advanced computer designed specifically for autonomous operation.”
Musk says the new silicon is an order of magnitude faster than the chips in their cars now, a product developed by NVIDIA, an industry leader and supplier to at least 20 other robo-car developers. The NVIDIA chips, Musk says, can deal with 200 frames of video per second, from the cameras around the car. (Unlike most other autonomous vehicle developers, Tesla is determined that cameras are enough to perceive the world, and it doesn’t need the more expensive lidar sensors.) Tesla’s chip can handle 2,000 frames per second, with some spare capacity left for redundancy and safety.
That might be true, but it’s not a Ruffles to Ruffles comparison. “The performance claims are against what they have in the vehicle today, which are three years old,” says NVIDIA’s director of automotive Danny Shapiro. NVIDIA’s latest silicon is at least 10 times faster than that, which would put it on a par with Tesla’s chip.
Tesla’s chip development is key to improving Autopilot. Right now the term is an umbrella for a collection of driver assistance tools, but in October 2016 Elon Musk claimed that all new Teslas now come with the hardware—cameras, ultrasonic sensors, and an on-board supercomputer—for self driving. They just needed to develop the software.
It turns out that computer wasn’t powerful enough. Tesla upgraded it once, a year later, but now says it’ll have to upgrade all the cars again, with even more number crunching power—and the new chip. All the connectors on the computer, which sits behind the glove box in the Model S and X, are exactly the same, so it’s just a case of service centers plugging in a new one. “It’s super kick-ass,” says Musk.
Building its own chips is the latest example of Tesla’s drive for full vertical integration, where it designs and makes everything in-house. Most automakers have chains of suppliers and sub-suppliers, but Tesla famously, and unusually, even builds its own seats. That allows it to create custom solutions, using the ground-up redesign approach that Musk prefers.
For the chips, the team went back to a “bare metal level” and instead of layering up GPUs, which were originally designed to speed up graphics processing for gaming machines, Tesla designed the calculator and memory circuits from scratch, sitting right next to each other, so they can pass data at high speed, which has been a traditional choke point when a GPU is used in emulation mode to run an AI neural net.
“The chips are up and working, and we have drop-in replacements for S, X and 3,” says Peter Banon, who leads Tesla’s computer hardware team, and who previously designed high-spec processors for Apple’s iPhones.
Tesla’s approach might be unconventional, but that’s hardly unusual for the company. And it’s a clever idea, says Jeff Miller, engineering practice professor in the computer science department at USC. “Instead of using a generic graphics processing unit, they’re able to customize it, because they know exactly what the data coming in is, and what they’re expecting to get back out of it.”
But Shapiro says NVIDIA’s chips are heavily customized too. “With a GPU, even though the name hasn’t changed, the architecture is fundamentally different,” he says. They now have deep learning accelerators, and new data links to remove bottlenecks and choke points, speeding them up dramatically. And, he says, sometimes a broader focus is desirable. NVIDIA’s stacks are designed to run the other applications self-drivers will need, like mapping and path planning, not just image processing.
Losing Tesla as a customer isn’t going to impact NVIDIA; the company says sales were so small that it won’t have a material impact. But it does highlight the importance of computing power in the self-driving car industry. These vehicles will have to run heavy duty supercomputers, and their builders will have to find a way to power, and cool, these machines, which run hot and hungry. Making the best chip, hidden away inside, could prove more important than fancy looks or screaming performance.
For Tesla owners, there’s still no timeline for those promised full self-driving capabilities, but Musk did say a coast-to-coast demonstration drive might be possible by the end of the year. (Something he also said last year.)
Tesla says some improvements to in Autopilot will come in a couple of months, with software V9.0. That could mean the introduction of what the company calls “on-ramp to off ramp” driving, where the car handles lane changes on the freeway automatically. But it’s mostly going to be about safety improvements, which might help counter some of the criticism of the feature’s safety record so far, with recent crashes highlighting its deficiencies.
“I think we’re starting to see a new set of safety features that really only make sense in this world where we have an extremely high understanding of what’s happening around the vehicle,” says Stuart Bowers, Tesla’s VP of engineering. He didn’t give specifics, but a car that can see all around itself, all the time, could nudge forwards to avoid a rear-end collision for example, or refuse to open doors when a cyclist is passing, and certainly should be better at avoiding highway barriers.
And if the features don’t live up to expectations, Tesla’s ready with a distraction. Elon Musk says the V9.0 update will bring classic Atari games to Tesla’s giant screens. Computer control of a Pole Position car is a lot easier than a real one.