Tech News

Why Tesla designs chips to train itself in driving technology

Tesla makes cars. Now, it’s also the last company looking for an advantage Artificial intelligence making his own silicon chips.

At a promotional event last month Tesla He has revealed the details of a custom AI chip called the D1 to work on the automatic learning algorithm behind his Autopilot autofocus system. The event was based on Tesla’s AI work and she was acting like a human dancer humanoid robot the company intends to build.

Tesla is a traditional chip maker that designs its own silicon. As AI deployment becomes more important and expensive, other companies that invest in technology — including Google, Amazon, and Microsoft — are still designing their own chips.

At the ceremony, Tesla CEO Elon Musk he said the company used to train the company to get more performance out of it neural network it will be the key to progress in autonomous driving. “If it takes a couple of days for a model to train for a couple of hours, that’s a great thing,” he said.

Tesla has already designed chips that interpret sensor input in cars after it switched from using Nvidia hardware in 2019. But creating the powerful and complex type of chip needed to train AI algorithms is much more expensive and challenging.

“If you think the solution to autonomous driving is to train a large neural network, then what you needed to follow was the kind of vertically integrated strategy that you needed,” he says. Chris Gerdes, the director Stanford Center for Automotive Research, Who attended the Tesla event.

Many car companies use neural networks to identify road objects, but rely more on Tesla technology, known as a “transformer” that receives input from a single giant neural network through eight cameras.

“We are building a synthetic animal from the ground up,” Tesla AI head Andrej Karpathy said at the August ceremony. “The car can be considered an animal. It moves autonomously, it senses the environment and it acts autonomously. “

Transformer models have provided a great deal advances in areas such as language comprehension in recent years; gains have come from having larger models and more data hungry. Train the largest AI programs it needs several million dollars worth the power of the cloud computer.

David Kanter, a chip analyst at Real World Technologies, says Musk is committed to speeding up workouts, “so I can do all this machine (self-driving program) that will speed up the Cruise and the Waymos of the world.” in autonomous driving to two of Tesla’s rivals.

Gerdes, from Stanford, says Tesla’s strategy is built around its neural network. Unlike many car driving companies, Tesla doesn’t use leaders, it’s a more expensive type of sensor that can see the world in 3D. It is based on interpreting scenes using a neural network algorithm to analyze camera and radar input. This is more rigorous because the algorithm has to reconstruct the map around it from sources around the camera, rather than relying on sensors that can take that photo directly.

But Tesla collects more training data than other car companies. More than a million Tesla on the road each send their company eight video camera outputs. Tesla says it employs 1,000 people to label these images — cars, trucks, traffic signs, lane markings and other features — to help train a large transformer. At the August event, Tesla said it could automatically prioritize which image to make the label process more efficient.

Gerdes says it’s a risk from Tesla’s view that at some point, adding more data might not improve the system. “Is it an account of more data?” he says. “Or are the capabilities of neural networks at a lower level than you expect?”

Answering that question will probably be expensive.

The rise of large and expensive AI models has not only encouraged some large companies to develop their own chips; it has also created dozens of well-funded startups that work in specialized silicon.


Source link

Related Articles

Back to top button