For longtime residents of Pittsburgh, seeing built by Uber, Argo AI, and others is nothing new. The city’s history with robot cars goes back to the late 1980s, when students at Carnegie Mellon University caught the occasional glimpse of a strange vehicle lumbering across campus. The bright-blue Chevy panel van, chugging along at slower than a walking pace, may not have looked like much. But NavLab 1 was slowly—very slowly—pioneering the age of autonomous driving.
Why did the researchers at CMU’s Robotics Institute use the van instead of, say, a Prius? First, this was a decade before Toyota started making the hybrid. Second, the NavLab (that’s Navigational Laboratory) was one of the first autonomous vehicles to carry its computers with it. They needed space, and lots of it. For the four researchers monitoring computer workstations, with their bulky cathode ray monitors stretched across a workbench. For the on-board supercomputer, camera, giant laser scanner, and air-conditioner. And for the four-cylinder gasoline engine that did nothing but generate electricity to keep the kit running.
Thirty years on, the companies carrying that early research into reality have proven that cars can indeed drive themselves, and now they’re . Those include regulations, liability, security, business models, and turning prototypes into production vehicles, by miniaturizing the electronics and reducing that massive electricity draw.
Today’s self-drivers don’t need extra engines, but they still use terrific amounts of power to run their onboard sensors and do all the calculations needed to analyze the world and make driving decisions. And it’s becoming a problem.
Tour Navlab 1, a lumbering blue Ford panel van.
A production car you can buy today, with just cameras and radar, generates something like 6 gigabytes of data every 30 seconds. It’s even more for a self-driver, with additional sensors like lidar. All the data needs to be combined, sorted, and turned into a robot-friendly picture of the world, with instructions on how to move through it. That takes huge computing power, which means huge electricity demands. Prototypes use around 2,500 watts, enough to light 40 incandescent light bulbs.
“To put such a system into a combustion-engined car doesn’t make any sense, because the fuel consumption will go up tremendously,” says Wilko Stark, Mercedes-Benz’s vice president of strategy. Switch over to electric cars, and that draw translates to reduced range, because power from the battery goes to the computers instead of the motors.
At first, companies may be able to write off the lost range or fuel. “It’s not a huge problem for the early applications, where we expect them to be used,” says Chris Urmson, who ran Google’s self-driving program and is now CEO of Aurora, a self-driving startup that has partnered with Volkswagen, Hyundai, and Chinese automaker Byton. That’s because the first robocars will likely be city-bound fleets of electric shuttles, moseying along at low speeds and able to recharge often.
Buyers of regular cars aren’t likely to be impressed though. Maybe you’re old enough to have dealt with a parent who turned off the car’s AC to save gas. Now imagine having to turn off the self-driving abilities just to make it to your destination without running out of electrons.
The good news is that the folks who make the chips buried in the car’s computers are on the case. At CES last month, Nvidia put the spotlight on a new processor designed specifically for autonomous vehicles, called Xavier. It has an eight-core CPU and 512-core GPU, a deep learning accelerator, computer vision accelerators, and 8K video processors. The company says it’s the most complex system on a chip ever created. “We’re bringing supercomputing from the data center into the car,” says the company’s man in charge of automotive, Danny Shapiro. But what’s key is that Xavier does more work for less power. “We’re able to deliver 30 trillion operations per second, all on a single SOC, or system on chip, that consumes 30 watts of energy.”
Even that’s not good enough for full self-driving vehicles. Nvidia believes that a fully self-sufficient, no-steering-wheel-or-pedals kind of driverless car will need to run on a platform it’s calling Pegasus. With two Xavier chips and two more GPUs, this platform can crunch 320 trillion operations per second and keep power consumption to an acceptable 500 watts.
Nvidia’s competitors are chasing the same goals. Intel is developing low-power chips optimized for self-driving cars, , and Qualcomm is building the communications hardware they’ll need—all with low power and efficiency in mind.
Specialized automotive chips help with other practical problems. Open the trunk on one of the self-driving prototypes running around Phoenix or San Francisco and you’ll likely see racks of computer equipment. Some of that is for testing and development—designers want to capture and record every moment the car’s in motion—and a consumer-facing version will require less hardware. But having somewhere to throw your groceries isn’t a negotiable for regular car buyers.
If you’ve had your laptop burn your legs, you know computers also get hot when they work hard. That heat is wasted energy, and it’s also not something you want in your car on a hot day. Some robocar prototypes need water-cooling with hoses and radiators, which eat up even more space. So now, the race is on to compact all that prototype equipment down to something the size of a laptop and tuck it away behind the glovebox, where it can be reached for upgrades but mostly ignored. New chips, with their lower power requirements, help here too: They generate less heat, so can get away with a small fan for cooling, and smaller packaging.
The car industry has a role model in the consumer electronics business, where devices get ever smaller and more capable. “All of our customers are always saying more performance, lower power—we need to do that across all markets,” says John Ronco, VP of product marketing at ARM, which designs the basic architecture of chips you find in most modern smartphones, as well as Nvidia’s self-driving chips.
It’s the age-old refrain—more for less—but it’s crucial if you want your first roboride to be in something a little more comfortable than a lurching blue panel van.