Skip to main content

Tesla’s new AI chip isn’t a silver bullet for self-driving cars

Tesla’s new AI chip isn’t a silver bullet for self-driving cars

/

Processing power is important, but building chips could be an expensive distraction for Tesla

Share this story

Tesla’s new self-driving computer contains the “best chip in the world,” according to Elon Musk.
Tesla’s new self-driving computer contains the “best chip in the world,” according to Elon Musk.
Credit: Tesla

At Tesla’s Autonomy Day event for investors on Monday, CEO Elon Musk made a characteristically big deal out of the company’s new microchips. Describing the new FSD computer (it stands for “full self-driving”) that will power Tesla’s vehicles, Musk said they contained what was “objectively” the “best chip in the world.” And that’s not best by a little, but “by a huge margin.”

Experts and rivals beg to differ. They say this claim should be taken with a pinch of salt, and that while Tesla’s new hardware is impressive, it doesn’t provide an insurmountable advantage for the company, nor will it solve the challenges facing self-driving cars.

“Any company that Elon Musk is affiliated with, they make big claims,” says Patrick Moorhead, a semiconductor analyst and industry veteran. “And sometimes things go through, and the rocket lands on a pedestal in the middle of the ocean, and sometimes ... it blows up in mid-air.”

“it’s a huge distraction and a waste of money.”

The claims being made about Tesla’s new self-driving chips fit this pattern, says Moorhead. It could be a fantastic coup for the company, but it could also create unnecessary problems. “I’ve been doing this for 30 years, and I’ve never seen a great first chip — out of anybody,” he tells The Verge. “To me, it’s a huge distraction and a waste of money.”

But Tesla certainly thinks its chips are worth shouting about, giving its new hardware a significant amount of airtime on Monday. Pete Bannon, the company’s lead on the project, opened the event with a detailed rundown of the FSD’s technical specs. Musk was sitting beside him as the hype man, ready to translate technical specs into sound bites and boasts.

Bannon is a respected figure in the industry, with stints designing chips at Intel and Apple before he joined Tesla in 2016. He explained that the FSD computer will be the brain of Tesla’s cars, processing input from eight cameras, 12 ultrasonic sensors, a front-facing radar, and GPS and mapping data. This data will be used to steer the car on the road. Tesla even claims that the FSD computer (which it started putting in cars a month ago) offers all the necessary processing grunt work to power full self-driving cars — if and when the software catches up.

Each FSD computer contains two separate chips with their own power lines, a backup in case the first chip fails.
Each FSD computer contains two separate chips with their own power lines, a backup in case the first chip fails.

Each FSD computer contains a number of components, most important of which are of Tesla’s custom-designed chips, which are built by Samsung. Each FSD contains two chips, and each chip has two accelerators that are specially designed to run neural networks, the AI components Tesla’s cars use to read the road. Each chip performs up to 72 trillion operations per second (or TOPS), and the system, as a whole, is capable of analyzing 2,100 frames of video each second, which is 21 times faster than previous-generation hardware, says Tesla.

These are impressive numbers, but some experts take exception to how they were presented. Nvidia’s Danny Shapiro, the company’s senior director of automotive, tells The Verge that the comparisons Tesla made against its own chips were simply “not accurate.” Tesla compared the processing power of two of its chips to just one chip made by Nvidia, says Shapiro, and even then, they understated its power.

“They inaccurately stated the performance as 21 TOPS, or trillion operations per second, and we’re at 30,” says Shapiro. “They cut 30 percent off our actual performance.” Nvidia previously provided the chips for Tesla’s cars, and it still supplies many other companies in the industry.

Tesla’s comparisons were “not accurate” and “deceptive”

Shapiro notes that Tesla also compared its new chips to Nvidia’s older hardware, specifically the company’s Xavier computer, which was unveiled back in 2016. A fairer comparison would be with Nvidia’s new Pegasus computer, says Shapiro, which has a processing speed of 320 TOPS. That’s more than double the output of Tesla’s new FSD hardware, though it comes with a corresponding greater power usage.

This aspect of the presentation was “deceptive,” says Mike Demler, a microprocessor expert and analyst with The Linley Group. “It wasn’t an apples-to-apples comparison at all,” he tells The Verge. “On pure technical grounds, [Tesla has built] a significant chip. It’s just not the best thing since sliced bread, as Musk claims.”

When asked if Nvidia’s hardware could replicate the 2,100 frames-per-second processing speed of the FSD computer, Shapiro said Tesla’s presentation hadn’t offered enough detail to answer that. “I don’t know what that benchmark is, so it’s hard for us to comment on what our performance would be,” he said.

Shapiro praised Tesla’s work on the FSD, though, and said that, like Nvidia, the company realized that huge amounts of processing power were necessary to make self-driving cars a reality. “The savvy automakers get that, [but] Tesla has a big lead,” says Shapiro. “I drive a Tesla every day, and it’s amazing.”

Tesla’s new chips are already in its latest cars, and Musk claims they’re powerful enough to power full self-driving.
Tesla’s new chips are already in its latest cars, and Musk claims they’re powerful enough to power full self-driving.

Putting aside technical comparisons, though, analysts suggest the bigger challenge for the automaker might be the expense and inflexibility that comes with committing to an in-house chip design. The field of artificial intelligence is fast-moving, and as new approaches for autonomous driving are developed, Tesla could find it’s backed the wrong horse, chip-wise.

“There’s a danger in making a chip too specific,” says Demler. “Did they make it flexible enough that, as these new algorithms are developed, they can adapt?”

Moorhead agrees, saying that while the field of self-driving is in flux, it would have been wiser to buy and customize chips from another company. Tesla has a history of bringing production in-house in order to get things right, and it’s an approach that doesn’t always pay off. Back in 2017, for example, the company bought automated equipment-maker Perbix, only for Musk to admit the following year that it had relied too much on robots for car assembly, causing it to miss production targets. “I feel like the company is painting itself into a corner,” says Moorhead.

But all this discussion of chips and TOPS is, to a certain degree, a distraction. Processing power is important for self-driving cars, no doubt, but achieving full autonomy has significant challenges that processors alone can’t solve. Tesla, for all its promises of a self-driving future, still only provides what’s known as Level 2 autonomy: a driver-assisted mode where vehicles can accelerate, brake, and steer on their own but only in controlled settings with drivers ready to take control.

Tesla has been promising full self-driving cars for years

During the event on Monday, Musk presented the integration of its new chips as an important piece in achieving full autonomy. “All cars being produced all have the hardware necessary, computer and otherwise, for full self-driving,” he said. “All you need to do is improve the software.” He added that, by 2020, Tesla would have “over a million” fully autonomous cars on the road.

These claims, combined with Bannon’s detailed and impressive rundown of the FSD’s technical specs, might lead you to think self-driving cars are just around the corner. But it’s worth remembering that Musk said exactly the same thing about Tesla’s cars in 2016. That same year, he promised that a Tesla would drive autonomously from LA to New York by the end of 2017. Three years later, it still hasn’t happened.