Nvidia Corp., the graphics-chip maker that has been seeking to expand in the automotive market, temporarily suspended its self-driving vehicle testing. The shares tumbled as much as 7.7 per cent.
The company, which provides technology to Uber Technologies Inc., stopped its self-driving test program on public roads in the aftermath of a March 18 fatal accident involving an Uber vehicle in Tempe, Arizona. It will continue to use vehicles with drivers that gather data.
Ultimately, autonomous vehicles “will be far safer than human drivers, so this important work needs to continue,” the Santa Clara, California-based company said in an email. The suspension was reported earlier by Reuters.
Chief Executive Officer Jen-Hsun Huang called the challenge of perfecting self-driving car systems “probably the hardest computing problem that the world has ever encountered.”
“We know so much is at stake and we have the opportunity to save so many lives in the future if we do it right,” he said Tuesday at a company event in San Jose, California. “This is the ultimate AI problem and we think we can solve this.”
Nvidia’s stock fell 6.6 per cent to US$228.31 at 2:44 p.m. in New York after earlier falling as low as US$225.77. The shares gained 26 per cent this year through Monday’s close.
Nvidia is trying to parlay its dominance of graphics chips, most commonly found in computer gaming machines, into new markets where their ability to perform multiple calculations in parallel can help with artificial intelligence work. Automotive is a key focus because Nvidia says its chips are ideal for running image-recognition software needed to turn camera pictures into driving decisions more quickly than the blink of an eye.
While the self-driving testing will be halted, Nvidia on Tuesday introduced a new virtual-reality product that will simulate the conditions automated systems will face on the roads.
The company said its computer system will help provide self-driving hardware with enough virtual experience in dangerous and difficult situations to become proficient in the real world. The product, Drive Constellation, is made up of two computers -- one simulates all of the inputs from cameras and other sensors, and the other uses that information to make smarter road maneuvers in the virtual environment. The two operate in a loop, Huang said at the event.
Huang explained that the numbers are against on-the-road testing.
A fleet of 20 vehicles will rack up about 1 million miles a year, he said. Worldwide human beings drive about 10 trillion miles a year.
“How confident are you when your fleet of test vehicles have driven about 1 million miles?” he said. “The amount of scenario coverage is just not possible in real life.”
The promise of growth in the automotive market is one reason investors have poured money into Nvidia’s stock. Revenue from Nvidia’s automotive business has been under pressure as sales of infotainment-system chips slow, leading the company to turn to what it calls a bright future in self-driving related parts.
As with any artificial intelligence system, the more experience a self-driving car has, the better it’ll become. That’s particularly a problem for vehicles because those systems require millions of miles of driving around public roads. Most of that driving is uneventful, contributing little educational value for the system. But in a virtual environment, the systems can be tested in all scenarios -- for example, lighting conditions can be set to be at their worst, such as the glare of a low sun at sunset, said Danny Shapiro, senior director of Nvidia’s automotive business. Dangerous scenarios can be played out over and over again with no risk.
“We can train at sunset for 24 hours, creating these blinding conditions that would only happen for a few minutes,” Shapiro said. The new system can “generate billions of miles to statistically show how safe these cars are,” he said.
Nvidia said the Drive Constellation system will be available starting in the third quarter this year. Alphabet Inc.’s Waymo, the leading driverless car company, developed its own virtual-testing environment years ago.