Tesla is an American vehicle manufacturing company founded by Elon Musk in 2003.
The company is best known for its electric cars and for specializing in solar panels and lithium-ion battery energy storage.
Tesla cars come with a lot of revolutionary features including super-charging, keycard access, and an autopilot mode.
The autopilot mode has been possible because of ideas from Artificial Intelligence (AI) and Tesla’s advanced Neural Network architecture.
Let’s discuss the Tesla Neural Network architecture in detail.
What are Neural Networks?
Neural Networks, or NNs, are a series of algorithms modeled after the biological activity of the human brain. Neural Networks consist of nodes, also called neurons. A collection of vertical nodes are known as layers.
Each layer consists of nodes, also called neurons, where the calculations take place. The nodes of one layer are connected to the next layer through transmission lines as seen below.
In the following diagram, the circles represent the nodes and the vertical collection of nodes represent the layers. There are three layers in this model.
How do they learn?
Data is fed to the model one entity at a time along with a label. The data is broken down into chunks and passed through each node of the model.
Nodes carry out mathematical operations on these chunks. After a series of calculations in one layer, data passes onto the next layer and so on.
Once completed, our model predicts the data label at the output layer. The model then proceeds to compare this predicted value with that of the actual label value.
If the values match, our model will take the next input but if the values differ the model will calculate the difference between both values, called loss, and adjust node calculations to produce matching labels next time.
Tesla’s Neural Network Architecture
Tesla uses cutting-edge research to train deep neural networks on problems ranging from perception to control.
Tesla’s per-camera networks analyze raw images to perform semantic segmentation, object detection, and monocular depth estimation.
The Neural Networks are trained on raw images that are extracted from videos taken from birds-eye-view network cameras that output the road layout, static infrastructure, and 3D objects directly in the top-down view.
Data images are unlabeled and cover a lot of diverse scenarios around the world and consist of one million vehicles in real time.
How does it work?
The network consists of 70,000 Graphical Processing Units (GPUs), that train 48 deep learning models.
The hardware components of the car including cameras and sensors, provide unsupervised data that is passed through the network of these models.
The car learns about possible objects in an environment, like a pedestrian, tree etc. from the given data.
The architecture also consists of two AI chips that use the principles of deep learning. These chips help make real-time decisions for the car, like when and how to turn, while driving.
The Neural Network architecture includes many powerful devices and concepts that contribute to its workings, including:
Full Self-Driving (FSD) chips are AI inference chips that run Tesla’s autopilot software. These chips have been designed with micro-architectural improvements that squeeze the maximum silicon performance-per-watt.
FSDs implement floor-planning, timing and power analysis while writing robust tests and scoreboards to verify AI’s functionality and performance.
Dojo Chips and Systems
Dojo is Tesla’s super computer system that solves hard problems with advanced technology for high-power delivery and cooling.
Dojo Chips include the AI that powers these systems and are designed for maximum performance, throughput and bandwidth at every granularity.
Together, the chips and systems are used to optimize power and performance for Tesla’s NN.
Autonomy algorithms are the core algorithms that drive the car by creating a high-fidelity representation of the world and planning out trajectories in a given space.
To train neural networks to predict such representations, Tesla algorithmically creates accurate and large-scale ground-truth data by combining information from the car’s sensors across space and time.
These algorithms use advanced techniques to build a robust planning and decision-making system that operates in complicated real-world situations under uncertainty.
Tesla’s evaluation infrastructure includes open-loop, closed-loop and hardware-in-the-loop evaluation tools and infrastructure at scale.
This infrastructure allows for AI to track performance improvements and prevent regressions.
Key Features of Tesla’s NN
- Cameras, ultrasonic sensors, and radar perceive the environment
- A radar measures the distance around the car
- Ultraviolet techniques measure proximity and passive video recognizes objects around the car
- Uses two AI chips built on principles of deep neural networks
- AI chips making up of 6 billion transistors
- 21 times faster than Nvidia chips
- AI chips have 32 megabytes of high-speed SRAM memory
- Consists of 48 Deep Learning models
- Contains 70,000 Graphical Processing Units (GPUs)
- Outputs 1000 distinct tensors (predictions) at each timestep
Tesla’s cutting-edge Neural Networks and AI architecture has made the idea of self-driving cars a reality.
This success of the leading AI-based automobile manufacturer is a result of its advanced FSD chips, Dojo chips, autonomy algorithms, evaluation infrastructure, and more.
If you want to learn more about AI, Deep Learning and the latest technology trends, check out our other interesting articles.