NVIDIA CEO Huang Renxun: Autopilot is an AI calculation problem

"Three key technologies for achieving unmanned driving: AI computers with super-capacity, powerful AI algorithms, and the ability to draw high-precision maps in real time. No company in the world currently has the above three technologies."

Huang Renxun believes that driverless driving has brought enormous challenges to the computing field. Real-time perception of the environment and determination of car location, development of driving plans, and safe driving based on the surrounding environment must be based on the ability of the driverless system to respond quickly. Deep learning and artificial intelligence can cope with the various situations in which unmanned driving takes place.

The following content is excerpted from the speech of the NVIDIA CEO Huang Renxun on GTC China 2016 on the unmanned part on September 13th. Lei Feng.com (search for “Lei Feng Wang” public number concern) finishing (with deletion):

Transportation is a $10 trillion industry and one of the largest in the world. 2 trillion automotive industry, 5 trillion including taxi transportation service industry, and truck industry. These add up to the 10 trillion transportation industry.

There is a lot of what we can do here, and one area is very exciting, that is, driverless cars. With AI, cars can think for themselves, always observe the surrounding environment, focus forever, don't feel tired, and the body is full of sensors. Therefore, the car has superhuman ability to ensure your safety, increase the efficiency of the taxi team, reduce accidents, and reduce insurance costs.

We had no way of knowing how to produce a driverless car until we had a GPU, so that we really know that cars with AI can truly achieve driverlessness.

The AI ​​car is not a car with a camera or radar. Radars and cameras are tools for discovering objects. We need more capabilities to truly be unmanned. In addition to finding obstacles in front of the car, other things like perception, you have to see and hear. And to understand the surrounding environment, we must judge how to act according to the information we see, whether to continue driving, changing lanes, or stopping. It is a very complicated problem for a car to judge countless states.

After judging the surrounding environment, it is necessary to make a decision on what action to take. This is an end-to-end, point-to-point AI problem that is an issue with AI calculations.

AI computing has already been available, and we can now imagine the dream of a driverless car from point to point. I hope this is a huge change for the $10 trillion industry.

Our BB8 AI car is not programmed. It has no driving skills. It is learned how to drive by observing the human driver. With AI computing supercomputer use in the automotive field, there are enough algorithms and training, we can teach these smart terminal devices to complete the task, which is not possible through software. This is why we believe that with AI, we will see a revolution in the field of smart devices.

There will be a lot of innovation in this area, but driverlessness is not a simple matter, and it is far from reaching an era of full or complete driverlessness.

Driverless is not an end product, not just to find a car in front of the car to stop. Unmanned driving will bring different values ​​to the society and make driving on the highway safer. We call it automatic cruising. On the other hand, it is completely unmanned, which means that there is no driver in the car. If there is no driver, it is basically impossible for the car to operate on its own. Unless the error rate is very low, we hope to achieve 99.9999% accuracy. rate.

Network accuracy and network computing are related. If you want to achieve high precision, you need more floating point calculations. If you need accuracy, you need a large model. If you have a large model and want to run it in real time, you need More floating point calculations; if you want more cameras around the car, you need more floating point calculations.

A completely driverless car requires a lot of supercomputers to achieve such a vision. People ask why so many calculations are needed, because driverlessness is not an ADAS. Driverlessness is an AI calculation problem, depending on how much you want to achieve, depending on how much calculation you need.

The DRIVE PX 2 architecture we developed is the world's first driverless car computing architecture. It is scalable, with very high performance and ability for perceptual, reasoning AI. AI supercomputers have a lot of AI algorithms, and all cars in the future will contain a lot of software, which will be the most complicated computing system.

Because we don't have a way to know what people's needs are in the future, each company produces cars that are different. There are three possible:

Auto-cruise, from one location to the destination, just tell the car where you want to go, such as taking me home, the car will take you home on the road;

C6-10D159 Engine 5 series:power range 414KWm-1050KWm

Engine.414KWm-1050KWm

CCSN POWER GENERATION INC.(Engine is a subsidiary of CCSN) , https://www.ccsnengine.com