Continental and NVIDIA’s computer cluster develops self-driving solutions in hours instead of weeks. During its design, a hybrid approach was chosen in order to expand computing and storage capacity with cloud-based solutions where necessary.
Driving support systems in today’s vehicles make decisions using algorithms based on artificial intelligence. The raw data is provided by environmental sensors such as radars and cameras. These are processed in real time by intelligent systems to create a transparent model of the vehicle environment and develop a strategy for interaction. Finally, the vehicle must also be steered to behave as planned. However, as systems become more complex,
traditional software development methods have reached their own ceilings,
and as a result, machine learning and simulations have become key to developing solutions.
Artificial neural networks are implemented to allow machine learning based on experience and new connect information with existing knowledge, more or less in the same way as the learning process takes place in the human brain. But while a child is able to recognize a car after being shown a few dozen different types of cars, thousands of hours of teaching and millions of images, or huge amounts of data, are needed to teach a neural network that will later help or even guide the driver on his own.
Since the beginning of this year, Continental’s artificial intelligence development center in Budapest has been served by a data center in Frankfurt am Main, Germany, with computing and storage capacity. This supercomputer is built from more than fifty NVIDIA DGX servers that use the NVIDIA Mellanox InfiniBand network for intra-cluster communication. The system is listed in the Top500 list of supercomputers as one of the largest systems in the automotive industry.
The data used to teach neural networks come mainly from Continental’s test vehicle fleet. These vehicles do 15,000 test kilometers a day while collecting about 100 terabytes of data, which would be equivalent to 50,000 hours of film. The recorded data can already be used to teach new systems by repetition, ie by simulating real test leads.
However, with the help of a supercomputer, the data can already be generated synthetically.
This can have a number of advantages during development, such as the need to record, store and search the data generated by the car fleet, because the scenarios to be used for learning can be created immediately by the system itself. At the same time, it accelerates development because virtual vehicles can do as many test kilometers in a few hours as a real car would only be able to do in several weeks. Third, the synthetic generation of data allows for the processing and response of changing and unpredictable traffic situations.