CES 2016: NVIDIA Drive PX 2 supercomputer for self-driving cars like having 150 MacBook Pros in your trunk

NVIDIA plans to put a supercomputer and deep-learning neural network in the truck of every self-driving car.

Although it's best known for making graphics chips, NVIDIA has been increasingly focused on the automotive market. At CES 2015, CEO Jen-Hsun Huang introduced Drive PX, a computer that would help automakers develop autonomous vehicles. At NVIDIA's CES 2016 press conference, Huang followed up by announcing Drive PX 2, an "AI supercomputer" for self-driving cars.

NVIDIA CEO Jen-Hsun Huang and DRIVE PX 2
NVIDIA CEO Jen-Hsun Huang announces DRIVE PX 2 at CES 2016.
Bill Detwiler/TechRepublic

Drive PX 2 is about the size of a lunchbox and has some impressive hardware specs:

  • Two next-generation Tegra processors (12 CPU cores)
  • Two discreet Pascal GPUs
  • Liquid cooling

All this hardware lets the Drive PX 2 process up to 8 trillion floating-point operations per second (8 TFLOP/S) or 24 trillion deep-learning operations per second (24 DL TOP/S). With this much power, the Drive PX 2 can churn through 2,800 images every second. This is over six times the number of images a GeForce GTX TITAN X (NVIDIA's highest performance graphics card at time of publication) can process. Huang equated using the Drive PX 2 to putting a 150 MacBook Pros in your trunk.

NVIDIA Drivenet compared to 150 MacBook Pros
DRIVE PX 2's processing power is equivilant to that of 150 MacBook Pros
Bill Detwiler/TechRepublic

Why the need for so much power? Driving is hard. Autonomous vehicles must collect, process and act upon millions of data points each second. The Drive PX 2 can process data from up to 12 video cameras as well as information from radar, lidar and ultrasonic sensors. The computer uses all this data to detect and identify objects (such as other vehicles, people, road signs, etc.), to determine the car's position relative to its environment, and then compute the best and safest route.

NVIDIA Drivenet being used to detect and identify vehicles while driving
Bill Detwiler/TechRepublic

Although impressive in its own right, the Drive PX 2 is only one part of NVIDIA's plans for automotive AI. Huang outlined the company's three-pronged strategy for autonomous vehicles cars:

  1. To ensure that NVIDIA GPUs are able to accelerate all deep learning and neural network frameworks
  2. Create the computing platforms to allow deep learning everywhere
  3. Create an end-to-end platform for deep learning

NVIDIA Drivenet deep-learning neural network for self-driving cars
NVIDIA Drivenet deep-learning neural network for self-driving cars
Bill Detwiler/TechRepublic

As part of this strategy, the company also announced NVIDIA Drivenet, its own deep-learning neural network for self-driving cars. Drivenet has 37 million neurons and it takes 40 billion operations to run through the whole network one time. Running on a GeForce GTX TITAN X, Drivenet can run at 50 frames per second. During the press conference, Mike Houston, head of development for NVIDIA's neural network, showed examples of Drivenet being used to detect objects on crowded city streets and highways, during clear and inclement weather.

Volvo will be the first automaker to test the Drive PX 2, installing the systems it in a fleet of a hundred self-driving cars.


Bill Detwiler is Managing Editor of TechRepublic and Tech Pro Research and the host of Cracking Open, CNET and TechRepublic's popular online show. Prior to joining TechRepublic in 2000, Bill was an IT manager, database administrator, and desktop supp...

Editor's Picks