At its CES keynote this evening, Nvidia discussed the strategy it's using to realize the dream of the self-driving car. The company believes autonomous vehicles will increase road safety, create new mobility services, save time, and reshape urban centers for the better.
The first element of that strategy is the Drive PX 2 hardware, an update of last year's Drive PX. The Drive PX 2 system uses two Pascal GPUs and two Nvidia Tegra SoCs to provide an impressive amount of computing resources. Drive PX 2 is claimed to offer eight teraflops and 24 "deep learning tera-ops" of raw power. Those Pascal chips are fabricated on a 16-nm FinFET process.
CEO Jen-Hsun Huang used the company's Titan X graphics card as a handy unit of measurement when describing Drive PX 2's capabilities. This system's deep-learning throughput is claimed to be over six times that of a single Titan X. If you prefer Huang's oft-invoked "CEO math," he describes the performance of this system as being equivalent to "150 MacBook Pros."
Nvidia continues to emphasize the importance of accelerating and applying deep learning techniques to the challenge of the self-driving car and more. Huang says the company is ensuring its products can work with every commonly used deep-learning platform in the world, and that it has a single binary-compatible hardware platform for deep learning in every device where a programmer might want to exploit neural networks.
The company also envisions a world where cars from a given manufacturer will drive off the lot pre-programmed to handle the chaotic tasks of safe driving. Over time, those cars will share novel driving experiences with a central neural network that's then retrained and redistributed to every car. That sharing means every self-driving car can benefit from the experiences of every other vehicle in this hypothetical "internet of cars."
To demonstrate this vision, Nvidia has developed a reference neural network it calls Drivenet. This network springs from another Nvidia tool for training neural networks called Digits. Drivenet can be deployed on Drive PX 2, and improvements to the network can be made using Digits.
Nvidia is wrapping neural networks and the software required to use them into a package it calls Driveworks. This tool kit provides ways to run critical self-driving tasks like perception (understanding what's around the vehicle), localization (figuring out where the vehicle is), and path planning (figuring out what driving actions to take). The company demonstrated Driveworks using a simulated car with inputs from six cameras and lidar sensors.
All together, this hardware and software stack lets self-driving cars build models of their environment, spot other cars on the road, and take the appropriate actions to drive safely and comfortably. That information could be presented to a driver or rider using Nvidia's Drive CX in-car multimedia platform.
Volvo will be the first car maker to begin testing Drive PX 2. Nvidia says the Swedish automaker will deploy 100 autonomous vehicles in Gothenburg next year, and it expects other manufacturers to begin incorporating Drive PX 2 into their autonomous vehicle research programs, as well.