Jen-Hsun Huang, President and Chief Executive Officer of GTC Japan 2016, held an in-house seminar in Tokyo on Oct. 6 to introduce Nvidia's role in autopilot Initiatives.
Huang said transportation-related industries, including passenger cars, trucks and taxis, are expected to create a large-scale market of 1,000 trillion yen, driven by artificial intelligence (AI). Among them, the major changes for the industry is based on AI's automatic driving technology.
He said, AI in the field of autopilot use "around the recognition (sensing)", "self-positioning (inference)", "driving (action)" 3 kinds. As the execution environment for these three AI processes, NVIDIA prepared the vehicle development platform "DRIVE PX2" and the software foundation (OS) "DRIVEWORKS ALPHA 1" running on the platform.
DRIVE PX2 is a car board, using NVIDIA's in-depth learning to optimize the processing of the GPU "Pascal". Autopilot is very broad in scope, including "automatic cruise" on the highway, "automatic private driver" driving to the destination automatically, and "fully automatic driving", which is completely unattended. For a wide range of autopilot, the Drive PX 2 has the scalability to support it with the same architecture. Combine multiple GPUs to improve processing performance. For the purpose of automatic cruise use NVIDIA ready for a GPU with a car board for the use of automatic private drivers prepared with 2 GPU car board. In addition, the need for advanced processing of full-automatic travel can be used by a number of vehicles equipped with two GPUs board implementation.
The DRIVEWORKS ALPHA 1 is an autopilot OS running on the DRIVE PX2, with the basic functions required for autopilot. The general mechanism is as follows. The object is identified by an AI technique called "DriveNet", and the available space (free space) is identified by an AI technique called "OpenROADNET". And then the results obtained with the map data, infer their own position, and build a car around the situation that the data structure "to occupy the grid (Occupancy Grid)". Finally, using AI technology called "PilotNet" to predict the traffic situation based on the occupied grid, plan walking routes, control the steering wheel, throttle and brake.
In addition, Jen-Hsun Huang also introduced a new generation of car-based SoC (Sytem on a Chip) "Xavier" being developed by NVIDIA. This is an autopilot-oriented SoC that uses Pascal's next-generation GPU "Volta", with 20 trillion operations per second. By using the SoC, you can achieve the same DRIVE PX2 (power 80W) has the same performance but more compact and more power (20W power consumption) of the vehicle board.