By the time autonomous cars using the "super energy-efficient" Pegasus go into production, the module will be certified to ISO 26262 ASIL-D, the highest integrity level of functional safety for road vehicles, said Huang
This involves "a rigour you have never seen before" and "a gigantic investment on our part," he said. "This is a big deal. It is a huge deal."
The company has been working towards ASIL-D for five years. "It's an endeavour," he said. "You have to have scale, you have to have commitment."
Autonomous vehicles will need multiple Pegasus systems, so Nvidia is already working on Orin, which will put all the functionality of Pegasus onto a single chip, just as Xavier did for the earlier Drive PX2 system.
Energy efficiency is important not so much to reduce the load on a vehicle's batteries, but because it allows designers to include redundant computers within a given power budget.
Diversity is important in addition to redundancy, explained Huang. For instance, Nvidia's software actively recognises free space separately from detecting objects, rather than working on the assumption that if it hasn't identified any objects then that piece of road must be empty. Similarly, it uses scene detection to recognise junctions as well as detecting relevant road markings. And information from HD maps is used in conjunction with the camera and lidar data.
"We're going to work on this for another two or three years before we ship [it] in cars," said Huang.
But he pointed out that we don't have to wait for full autonomy to improve safety. For example, cars could take advantage of deep learning technology to recognise via an in-cabin camera that their drivers are falling asleep or using their phones.
Asked whether some companies are moving ahead too quickly with automated vehicles - presumably a reference to the recent fatality involving an Uber self-driving car - Huang did not answer directly, but pointed out that "our strategy is to create an open platform" allowing car makers and other customers to buy a fully-functioning system from Nvidia, or to replace or augment any the modules with their own code.
"Uber does not use Nvidia's Drive technology," he said later.
A lot of technological issues still need to be solved, Huang observed. That affects some basic components such as sensors.
Nvidia has halted its own testing on public roads until Uber has determined what went wrong in Tempe, Arizona. The company responded with "extreme caution," he said. "We take safety to extreme levels" when testing: "Our friends are in the car."
"As soon as the news became clear, we stopped [open road testing]," he stressed. "Everybody in the industry should pause."
The company's Drive Constellation system for the testing of autonomous vehicle software in a simulated environment will become a bigger part of testing, he suggested. "You're going to see a lot more simulated miles."
Autonomous vehicle technology is being developed to save lives, Huang said, and therefore he expects the incident will increase investment in the area.
Disclosure: The writer attended Nvidia's GPU Technology Conference as a guest of the company.