Nvidia CEO Sees a Long, Simulated Road to AVs

The fatal accident involving a self-driving Uber car earlier this month will lead to increased investment in autonomous driving research and development, Nvidia CEO Jensen Huang said Tuesday.Companies working on autonomous cars will double down on refining the technology and improving safety, with much of the testing taking place on simulation systems like Nvidia’s new Constellation platform, Huang noted at his company’s annual GTC conference.

“Anybody who thought they could get by without supercomputers, and simulators, and just vast amounts of data collection, and all those engineers dedicated to making sure that this product is as safe as possible … the sensibility has completely changed,” Huang told reporters at a question-and-answer session on March 27. “The investment’s going to go up.”

Nvidia supplies silicon and software platforms for many of the autonomous systems emerging from carmakers and startups, including Uber test vehicles. Like Uber, Nvidia has halted its own real-world tests until it can find out more about why the accident on Sunday, March 18, took place, Huang said. “We should give ourselves the opportunity to learn from that incident,” he said. Then, “we’ll self-assess and decide what to to.”

It hasn’t been a good week for autonomous cars at Uber, one of the most high-profile players in this fledgling field. On Monday, the governor of Arizona said the state would seek to suspendUber’s tests there. On Tuesday, Uber said it would stop testing autonomous cars in California.

But Huang is bullish on the technology, saying ultimately it will lead to fewer accidents.

“The best way to remind people all over the world of the importance of this work is not to discourage it but to bring awareness to the importance of it,” he said.

In a keynote address earlier in the day, he tamped down expectations that driverless cars will be on every road soon.

“We are far from finished,” Huang said. “There are several thousand engineers right now working on autonomous vehicles. We are going to work on this for another two or three years before we ship in volume cars.”

As an example of where that work stands, Nvidia’s systems in current self-driving vehicles have ten neural networks for artificial intelligence functions to constantly refine their driving. The mass-market cars coming in two or three years will probably have 20 to 30 neural networks, he said.

Another telling detail: There are 1,500 people at Nvidia labeling images of things like street signs and potholes to train self-driving systems to recognize them. (AI does a lot of that labeling, too.)

“We’re dedicating ourselves to this problem, the grandest of computer problems,” Huang said.

After unveiling next-generation hardware for cars earlier in the year, Nvidia focused on simulation testing in its GTC announcements on Tuesday.

The company’s Drive Constellation platform and its Drive Sim software are designed to accelerate virtual testing of autonomous car systems. They will put those systems through their paces within photorealistic simulations, generating more miles of testing with more rare, challenging situations than cars in the real world could go through.

Using 10,000 Constellation systems, testers could simulate 3 billion miles of driving per year, Huang said. By contrast, it would take a fleet of 20 real-world cars to rack up even 1 million test miles, he said. Nvidia will both use Constellation itself and sell it to other companies.

— Stephen Lawson is a freelance writer based in San Francisco. Follow him on Twitter @sdlawsonmedia.

Leave a comment

Your email address will not be published. Required fields are marked *