Tesla's new autopilot chip is ready

Tesla's MODEL S sedan, the flagship model of the car company, was first shown in 2009 and has been on sale since 2012. And in addition to a small change - the removal of the fake radiator guardrail at the front of the car, it looks almost like a decade. This persistence is unusual because most manufacturers redesign their cars every four to six years to stay fresh – so that buyers can continue to buy. For Tesla, technology upgrades are the selling point. The company launches several software updates each year, adding features such as summoning to allow a car to automatically enter and exit the garage without the need for a car, and a camper mode so that users can sleep in a heated car.

One of Tesla's biggest goals is that all the cars that the company is currently building can be fully automated. Therefore, when Elon Musk announced that Tesla was developing its own computer chip after the company's second-quarter earnings report was released, it represented a major statement from the car company. Musk's high-profile exposure (actually may not be news, because according to WIRED's report at the time, Musk had already talked about this at the artificial intelligence conference held in December last year), he boasted that the company has "the world's most advanced and dedicated operation." And the computer is designed."

Musk said that their new silicon is an order of magnitude faster than current automotive chips, a product developed by NVIDIA. NVIDIA is an industry leader and a supplier to more than 20 robot developers. Musk said that the NVIDIA chip can process 200 frames of video from a camera around the car every second. (Unlike most other autonomous car developers, Tesla insists that the camera is enough to perceive the world, and it does not require more expensive lidar sensors.) Tesla's chips can handle 2,000 frames per second and still have remaining capacity. Used for data redundancy and security.

This may be true, but this is not a comparison of Ruffles and Ruffles. Danny Shapiro, NVIDIA's director of automotive, said, "The new looks don't match the vehicle's internal equipment. These devices are three years ago." NVIDIA's latest chips are at least 10 times faster than the chips, which will be with Tesla. The chip is comparable. Tesla's chip development is the key to improving Autopilot. The term is currently a series of umbrellas for driving aids, but in October 2016, Elon Musk claimed that all new Tesla were equipped with a range of hardware including camera, ultrasonic sensors and car supercomputers. These will be used for autonomous driving. They only need to develop software.

It turns out that the computer is not strong enough. A year later, Tesla upgraded again, but it now says it has to upgrade all models again, with better digital computing power and new chips. All connectors on this computer are identical, behind the glove box in Model S and X. So the upgrade is just a new connector plugged into the service center. "This time the computer is super powerful," Musk said.

Building your own chip is the latest example of Tesla's push for full vertical integration, which relies on the company's internal design and manufacturing of all products. Most car manufacturers have chains of suppliers and sub-suppliers, but Tesla is self-sufficient in its unusual way. This allows it to create custom solutions using Musk's new approach to resetting.

For these chips, the team returned to the "bare metal level" instead of the layered GPU, which was originally designed to speed up the graphics processing of the gaming machine. Tesla designed the calculator and memory circuits from the ground up, allowing them to sit next to each other, allowing high-speed data transfer, a traditional blocking point in the past when the GPU was running an AI neural network in emulation mode.

"These chips are running and we have provided direct replacements for S, X and 3," said Peter Banon, who is responsible for Tesla's computer hardware team, who previously designed high-profile processors for Apple's iPhone.

Tesla's approach may be unconventional, but this is not uncommon for their company. Jeff Miller, a professor of engineering practice at the University of Southern California's Department of Computer Science, thinks this is a smart idea. "They don't use a generic graphics processing unit, but instead customize it because they know exactly what data is being entered and what data is expected to be returned from."

However, Shapiro said that NVIDIA's chips have also been heavily customized. "In the case of GPUs, although the name has not changed, the architecture is fundamentally different," he said. They now have deep learning accelerators and new data links to eliminate bottlenecks and blocking points, greatly speeding them up. Moreover, he believes that sometimes this requires more attention. NVIDIA's stack is designed to run other applications needed for automated programs, such as mapping and path planning, not just image processing.

Losing Tesla as a customer does not affect NVIDIA; the company says that sales in this segment are small and will not have a significant impact. But it does emphasize the importance of computing power in the autonomous automotive industry. These vehicles must run heavy supercomputers, and their builders must find a way to drive and cool these load-running and overheated machines. Making the best chips, hiding them, may be more important than designing a fancy look or a screaming look.

For Tesla owners, they still don't know when they will have a car that the company promises to drive automatically, but Musk does say that a full demonstration may be available by the end of the year. (He said that last year.)

Tesla announced that some improvements to the autopilot will be available in software V9.0 within a few months. This may mean introducing the company's so-called "change lane" driving, ie the car automatically processes the lane change on the highway. But it's mainly about security improvements, which may help to counter some of the criticisms of the feature's security record so far, especially since the recent crash has highlighted its shortcomings.

Stuart Bowers, Tesla's vice president of engineering, said: "I think we've started to see a new set of security features that make sense in a world where we have a very high level of understanding of the road conditions around the vehicle." He did not elaborate However, a car that can always pay attention to the surrounding car may move forward to avoid, for example, a rear-end collision; or refuse to open the door when the rider passes; of course, win back to better avoid the obstacles of the highway.

If these features are not as expected, Tesla will focus more on other businesses. Elon Musk said that the V9.0 update will bring classic Atari games to Tesla's giant screen, with computer control of the Pole Position car much easier than real cars.