Why Tesla designs chips to train its self-driving technology


Tesla makes cars. Now it is also the latest company to seek advantages artificial intelligence By making your own Silicon chip.

exist Promotions last month, Tesla It revealed the details of a custom AI chip called D1, which is used to train the machine learning algorithm behind its Autopilot autopilot system.The focus of this event is Tesla’s artificial intelligence work, and there is a dancing human posing Humanoid robot The company plans to build.

Tesla is the latest non-traditional chip manufacturer to design its own chips. As artificial intelligence becomes more important and costly to deploy, other companies that have invested heavily in the technology, such as Google, Amazon, and Microsoft, are now designing their own chips.

Event site, Tesla CEO Elon Musk Said to squeeze more performance from the computer system used to train the company Neural Networks Will become the key to the advancement of autonomous driving. “If a model takes days to train instead of hours, it will be a big deal,” he said.

After switching to Nvidia hardware in 2019, Tesla has designed a chip that can interpret sensor input in its cars. However, creating the powerful and complex chips required to train AI algorithms is much more expensive and challenging.

“If you think that the solution for autonomous driving is to train a large neural network, then the next step is the kind of vertical integration strategy you need,” said Chris Geddes, Director Stanford University Automotive Research Center, Who participated in the Tesla event.

Many car companies use neural networks to identify objects on the road, but Tesla relies more on this technology. A huge neural network called a “transformer” receives input from eight cameras at a time.

“We are effectively building a synthetic animal from the ground up,” Tesla’s head of artificial intelligence Andre Capaci said at the August event. “The car can be thought of as an animal. It moves autonomously, perceives the environment and acts autonomously.”

The transformer model provides a large Advances in language comprehension and other fields In recent years; gains have come from making models larger and requiring more data.Train the largest artificial intelligence program Need millions of dollars Worth the cloud computing power.

According to David Kanter, a chip analyst at Real World Technologies, Musk is betting that by speeding up training, “then I can make the whole machine—the self-driving program—accelerate before the Cruises and Waymos in the world. ,” He was referring to Tesla’s two competitors in the field of autonomous driving.

Gerdes of Stanford University said that Tesla’s strategy is built around its neural network. Unlike many self-driving car companies, Tesla does not use lidar, which is a more expensive sensor that can observe the world in 3D. It relies on interpreting the scene by parsing input from its cameras and radars using neural network algorithms. This is more computationally demanding because the algorithm must reconstruct a map of the surrounding environment from the camera feed, rather than relying on a sensor that can directly capture the picture.

But Tesla also collected more training data than other car companies. Each of the more than 1 million Tesla vehicles on the road feeds video from its eight cameras back to the company. Tesla said it hired 1,000 employees to mark these images — pay attention to cars, trucks, traffic signs, lane markings, and other features — to help train large transformers. At the August event, Tesla also stated that it can automatically select which images to prioritize to improve process efficiency.

Gerdes said that one risk of Tesla’s approach is that at some point, adding more data may not make the system better. “Is it just a matter of more data?” he said. “Or is the capability of the neural network lower than you hoped?”

Either way, answering this question can be costly.

The rise of large, expensive AI models has not only inspired some large companies to develop their own chips; it has also spawned dozens of well-funded startups dedicated to specialized silicon chips.


Source link