Musk’s AI chip ambition: Turning cars into “artificial animals”


Tencent technology news on September 8, Tesla is famous for manufacturing electric vehicles, but now the company is also seeking to occupy a greater advantage in the field of artificial intelligence (AI) through independent research and development of chips. In last month’s “AI day” campaign, Tesla announced details of the custom AI chip, called D1, which is used to train machine learning algorithms behind its Autopilot. The event focused on Tesla’s AI efforts and demonstrated its planned humanoid robot.
Tesla is the latest non-traditional chip manufacturer with independently designed chips. As AI becomes more and more important and the deployment cost is higher and higher, other companies that invest heavily in this technology are now designing their own chips, including Google, Amazon and Microsoft. At the event, Elon Musk, CEO of Tesla, said that squeezing more performance from the computer system used to train the company’s neural network would be the key to the progress of automatic driving. “If the training time required for a model is reduced to a few hours rather than a few days, this may have a significant impact,” he said
After giving up NVIDIA hardware in 2019, Tesla has designed a chip that can interpret sensor data in its car. However, creating powerful and complex chips that can train AI algorithms is much more expensive and challenging. Chris Gerdes, director of Stanford University’s Automotive Research Center who participated in Tesla activities, said: “if you think the solution of automatic driving is to train a larger neural network, then the next is the vertical integration strategy you need.”
Many automobile companies use neural networks to identify objects on the road, but Tesla relies more on this technology. Its giant neural network called “transformer” can receive the input of eight cameras at the same time. Andrej karparis, head of Tesla AI business, said at the event in August: “we are actually making an artificial animal from scratch. The car can be imagined as an adult animal. It can move freely, perceive the environment and take independent action. ”
In recent years, the “transformer” model has made great progress in the field of language understanding, mainly from the data that makes the model bigger and more needed. Training the largest AI project requires millions of dollars worth of cloud computing capabilities. David Kanter, a chip analyst at real world technologies, said Musk’s bet was that by speeding up the training, “I can accelerate the progress of the whole machine (autopilot program) ahead of cruises and waymo.” He refers to Tesla’s two major competitors in the field of automatic driving.
Gerdes of Stanford University said Tesla’s strategy was built around its neural network. Unlike many automated driving companies, Tesla does not use lidar, which is a more expensive sensor that can see the 3D world. Instead, Tesla relies on interpreting the scene by using neural network algorithms to parse input from cameras and radar. This is more computationally demanding because the algorithm must reconstruct the map of its surrounding environment according to the camera feed, rather than relying on the sensor that can directly capture the photo.
But Tesla also collects more training data than other car companies. Of the more than 1 million Tesla cars driving on the road, each sent the video source captured by its eight cameras back to the company. Tesla said it hired 1000 people to mark these images, including paying attention to cars, trucks, traffic signs, lane markings and other features, to help train large neural networks. At the event in August, Tesla also said that it can automatically select which images to annotate first to improve processing efficiency.
Gedders said that one risk of Tesla’s method is that at some time, adding more data may not make the system better. “Is this just a problem that more data can solve?” he said? Or is the ability of neural networks stagnant, below the level we want? ” Either way, solving this problem may be costly.
The rise of large and expensive AI models has not only inspired many large companies to develop their own chips, but also spawned dozens of well funded start-ups specializing in chip research and development. The AI training chip market is currently dominated by NVIDIA, which initially produced game chips. When the situation became obvious, the company turned to supplying AI chips because its graphics processor (GPU) was more suitable for running large neural networks than the central processing unit (CPU) of the general-purpose computer core.
AI also promotes the diversification of chip design through clever recursion. Chip design usually requires deep technical expertise and judgment, but machine learning has proved to be very effective in automating the elements in the process. Google, Samsung and other companies are making chips partially designed by AI. Analyst Kanter said that there are still some technical problems with special chips such as Tesla D1, including how to connect them effectively and how the algorithm is split and distributed on different chips. “In a sense, you’re writing a big check for your software team,” Kanter said
Tesla did not respond to a request for comment. Huei Peng, a professor at the University of Michigan who focuses on automatic driving, said that if D1 eventually succeeds, musk may sell it to other carmakers, who will need to follow its technology leadership. However, it is not clear whether the method Tesla is taking will work financially and technically, but Huei Peng has learned not to bet with musk. “He did a lot of things that everyone said would not work, but the end result proved that he was right,” he said´╝ł Tencent Technology (reviser / Jinlu)