Tesla uses artificial intelligence to improve autonomous driving

Tesla said at an investor conference last week that its self-driving capabilities will be significantly improved once the Dojo supercomputer joins its high-performance computing infrastructure.

Ashok Elluswamy, director of Tesla Autopilot software, said during a speech at the Investor Day meeting that running FSD (Full Self-Driving) Software-based Tesla vehicles (currently around 400,000 customers) will be able to make smarter self-driving decisions through hardware upgrades, which will improve overall artificial intelligence (AI) capabilities.

The company currently has an artificial intelligence system that collects visual data from eight cameras on the vehicle in real time and generates a 3D output that identifies obstacles and their movement, vehicles, roads and traffic traffic lights and modeling tasks that help cars make decisions.

Tesla mines its car network for more visual data and feeds it into training models. Training models to continuously learn to solve new problems helps AI better understand patterns on the road. Through FSD software upgrades, new knowledge is input into the car.

“If we run and repeat this process, it gets better and better,” Elluswamy said. “The solution to scalable FSD is to get the architecture, data and computation just right, We have assembled a world-class team to execute this work. They are bringing these three efforts to the forefront.”

It’s not all plain sailing for FSD, software glitch forces Tesla recall purchased more than 360,000 vehicles. The company provided a software fix via an over-the-air update. Tesla customers can purchase FSD starting at $99 per month. Some customers with older Tesla models also need to pay extra to install FSD computers. Elluswamy claims that Teslas with FSD are still five to six times safer than the U.S. national average.

#Elluswamy said: “As we improve the safety, reliability and comfort of our systems, they can unlock driverless operation, enabling new ways of using cars and beyond today’s How to use it."

Today, the company runs its AI systems on 14,000 GPUs in its data centers and can take advantage of 30 petabytes of video cache, which is growing to 200 petabytes. About 4,000 GPUs are used for automatic labeling, and the remaining 10,000 GPUs are used for artificial intelligence data training.

“Once we bring Dojo (our training computer) into this space, all of that will increase significantly,” Elluswamy said.

The Dojo system is based on Tesla’s self-developed D1 chip, which can provide 22.6 trillion FP32 performance. It has 50 billion transistors and 10TBps of on-chip bandwidth, as well as 4TBps of off-chip bandwidth.

A set of D1 chips will be housed in a high-density ExaPOD cabinet, which will deliver 1.1 exabytes of BFP16 and CFP8 performance. Tesla’s on-board FSB computer can provide 150 teraflops of performance and is mainly used for inference.

Ganesh Venkataraman, senior director of hardware at Tesla, said in a speech at last year’s Hot Chips conference that Tesla made the D1 chip because of the existing GPU and CPU expansion capabilities. defect.

Venkataraman said: "We noticed a lot of bottlenecks. First on the inference side, which is why we do FSD computers. Then we started noticing similar training scale issues, understanding the work After measuring... we can optimize our system based on output needs."

In the early days, Tesla's AI system relied on a single camera and single frame video, and then in the autonomous car Splicing is done in post-processing of the planning system.

"It was very fragile and didn't lead to significant success," Elluswamy said.

######### Over the past few years, Tesla has transformed into a "multi-camera video world." Each vehicle has eight cameras that feed visual information into the AI system, which then generates a 3D output space. AI makes decisions about the presence of obstacles, their movement, lanes, roads and traffic lights, etc. ######

Task modeling goes beyond computer vision and uses techniques used in artificial intelligence systems such as ChatGPT, including transformers, attention modules, and autoregression of tokens Modeling.

#Elluswamy said: "With such an end-to-end solution to the perception system, we really remove the fragile post-processing step and provide high-quality output to the planning system. Even if Even the planning system is not static. It is now starting to use more and more artificial intelligence systems to solve this problem."

Autonomous vehicles need to respond quickly to make smooth, real-time decisions. Safe decision-making. Elluswamy gave the example of a 50 millisecond response time, where a self-driving car can make driving decisions after interacting with the surrounding environment (including pedestrians and traffic lights).

This is a lot of data, and in traditional computing, "each piece of data requires 10 milliseconds of calculation time, which can easily exceed 1,000 milliseconds. This is unacceptable." "But with AI, we package all of this into 50 milliseconds of computation so it can run in real time," Elluswamy said.

Tesla is augmenting its raw data by collecting vehicle data on different road conditions and traffic trends around the world. Tesla uses algorithms to reconstruct lanes, road boundaries, curbs, crosswalks and other images, which are then used as a basis to help the car navigate.

#This is achieved by collecting various fragments of different cars in the fleet and combining all the fragments into a unified representation of the world around the car," Elluswamy said.

As more data is input into the system, the training model is continuously reconstructed. To train the networks, Tesla built a complex automated labeling pipeline on the collected data, ran computational algorithms on it, and then generated labels to train these networks.

##############################################################"Once we have completed rebuilding the foundation, we can build various simulations on top of the foundation to generate an infinite variety of data for training." Tesla has powerful Simulators that can synthesize adversarial weather, lighting conditions, and even the motion of other objects. "Every time we add data, performance improves."######

The above is the detailed content of Tesla uses artificial intelligence to improve autonomous driving. For more information, please follow other related articles on the PHP Chinese website!