Tesla CEO Elon Musk showed investors on Monday a custom-designed artificial intelligence chip his company recently began installing in its vehicles. He claimed it was powerful enough to allow Teslas to pilot themselves without supervision by the end of this year. “All Tesla cars being produced right now have everything necessary for full self-driving,” Musk said. “All you need to do is improve the software.”
That’s a big claim, given that no one has yet demonstrated a self-driving vehicle that’s ready for market. Because chip design is a complex business, Tesla began designing the chip three years ago when self-driving technology was less advanced. So how did the company decide how much computing power would be sufficient? “A thumb in the wind,” said Pete Bannon, the veteran chip architect leading Tesla’s project.
On Monday, Musk said his estimate proved to be spot-on. He predicted that a year from now, more than 1 million Tesla vehicles will be capable of driving themselves while a person sleeps in the driver’s seat, and that the company will help customers make money when they are not using their cars by renting them out as “robotaxis”
But Bannon’s comment is a reminder that many aspects of producing self-driving vehicles are “thumb in the wind” exercises. Tesla is trying to get the technology working without the use of 3D sensors called lidars, which most of its competitors assert are essential. Chip and robotics experts say it’s unclear what amount of computing power is sufficient for self-driving. “I don’t think anyone knows what hardware is needed,” said Filip Piekniewski, a robot perception specialist at startup Accel Robotics.
That Tesla hasn’t fully proved out its technology became clear later Monday, when the company offered test rides to some attendees at its investor event. “I wouldn’t say it is perfect,” said Trip Chowdhry, managing director at Global Equities Research. At one point, the car didn’t recognize a turn arrow on a traffic light, and it was notably hesitant at a four-way stop, he adds, although most of the time the driving was impressive.
That experience sounds on par for the herky-jerky world of self-driving car prototypes. Some front-runners in the field are downplaying earlier suggestions that a robot car revolution is near.
Alphabet division Waymo is widely seen as the technological leader, but its launch of a driverless taxi service in Phoenix last year was muted. The vehicles have safety drivers sitting at the wheel, and are not available to the general public. Ford’s CEO said this month that the industry had “overestimated the arrival of autonomous vehicles.”
Even Musk’s bold claims Monday were a retrenchment. He first announced that Tesla vehicles had the hardware needed for full autonomy in October 2016, after the company added new cameras and a more powerful computer with chips from Nvidia. The company declined to make anyone available to comment Tuesday.
On Monday, Tesla made that computer, which shipped in hundreds of thousands of Tesla vehicles, into a punching bag. “It’s night and day,” Musk said, after Bannon shared stats showing his new hardware outperformed Nvidia’s. Customers will be able to pay Tesla to pull out the Nvidia-based computers from older cars, and retrofit new computers with the carmaker’s own chip, according to Musk.
Tesla’s claims were quickly contested by Nvidia, which called them inaccurate.
One of Bannon’s slides said Tesla’s computer, which contains two new chips and is installed behind the glove box, can handle data at almost seven times the rate of Nvidia’s technology—144 trillion operations per second, compared with 21 trillion for Nvidia.
Danny Shapiro, Nvidia’s senior director for automotive, said that was misleading. Tesla compared its two-chip computer with an Nvidia chip, called Xavier, that was released several years before, and is not marketed as a way to power full self-driving. Shapiro said a fairer comparison is Nvidia’s Drive AGX Pegasus computer, used by companies including Daimler, which he said can deliver more than twice as many operations per second as Tesla’s new computer.
As for a car, top speed isn’t the only consideration when choosing a computer chip. Bannon also said Tesla’s design uses power very efficiently, allowing it to be retrofitted into existing models, and reducing the drain on the vehicles’ battery and hence driving range.
Tesla’s foray into chip design follows a strategy used by Apple and Google. The iPhone maker has long designed its own processors for mobile devices to maximize battery life and performance, while Google makes server chips customized for its AI software. Tesla revealed its own chip project at the end of 2017, although the executive leading it stepped down months later, leaving Bannon to take charge.
All those projects are motivated by gaining efficiency from specially designed chips, compared with off-the-shelf models. General purpose chips pay a penalty for offering a range of features designed for multiple customers, said Kevin Krewell, a semiconductor analyst at Tirias Research. Tesla said Monday that it had crafted its chip to support the artificial neural networks it uses to process data from the eight cameras on each of its vehicles.
Designing your own silicon comes at a steep price, however. Krewell estimated that Tesla’s three-year project cost at least $60 million. “I’m not sure Tesla makes enough cars to have made the custom chip design economically worthwhile,” he said. Tesla ships cars in the hundreds of thousands per year, compared with the millions of units that phone or chipmakers produce annually, helping to spread out the costs.
Chowdhry at Global Equities Research said that’s not a problem for Tesla because it sells its products at higher prices, and is making a longer-term bet on autonomous driving. “When you think about Tesla, you can’t think in terms of quarter to quarter,” he added. “The technology is a driver for demand for years into the future.”