We live and breathe car performance metrics like horsepower, lb-ft, zero-to-60 and lateral g’s. Terms such as kilowatt-hours—how much energy batteries can hold—is being added to the list as more electric vehicles are hitting the road. Now, with semi-autonomous cars starting to arrive, there’s a weird-sounding acronym we’ll need to know – FLOPS.
FLOPS? No, not those thin rubber soles for the beach. Rather, it represents how fast a car’s computer can calculate floating point operations per second (FLOPS). Instead of bantering with friends about acceleration and lap times, FLOPS will be the new performance metric to debate—the power to recognize and process the road environment for safe, self-driving.
OK, FLOPS doesn’t sound as exciting as horsepower. But it’s plenty interesting to me—an AI and machine learning geek heading out on a six-hour drive from L.A. to Silicon Valley to attend an automotive future-think conference, while behind the wheel of a 2018 Cadillac CT6 with Super Cruise.
Available as a $5,000 option, Super Cruise is a new adaptive cruise control technology that I’m hoping is going to steer, accelerate and brake for me, all on its own, for the next 400 miles. For reference, that’s just a fraction of the 130,000-plus highway miles in the U.S. and Canada that’s been mapped out by Ushr Inc. of Livonia, Michigan, who’ve assembled high-definition LiDAR-derived road maps upon which GM overlays its own GPS data.
Unlike Tesla’s Autopilot and Nissan’s ProPilot Assist – which rely mainly on forward-facing cameras to recognize road markings for lane-keeping – the Cadillac Super Cruise uses hi-def maps to locate the car’s position within four inches of accuracy. It also knows up to 1.5 miles worth of terrain ahead. The HD maps are updated quarterly over-the-air via GM’s OnStar system, but more frequent updates can be sent to include condition changes such as construction zones.
Although the Autopilot and ProPilot Assist require the driver’s hands remain on the wheel for their system to stay on, the Super Cruise will stay active only if the driver is paying attention. It determines this by continuously tracking driver face and eye gaze direction via a steering-column-mounted infrared camera.
Don’t keep your eyes looking ahead for more than five or six seconds – whether it be staring down at the infotainment screen, side mirrors, or any place but the road – and the Super Cruise’s overlords will intervene. First, by alerting you with a flashing green light bar on the steering wheel; ignore it and the bar flashes red, accompanied by seat pulses and beeps. Still playing dumb? It adds voice prompts, then slows the car to a safe stop if you still don’t respond. By that time—even if you finally did take over—you’re put in the penalty box and can’t reactivate Super Cruise again without an engine restart.
With my face illuminated by invisible infrared lights (located near the top of the steering wheel) I head north. The Cadillac Super Cruise website shows that Interstate 5 is mostly mapped and is also the fastest route.
Even if you’re trying Super Cruise for the first time, it’s easy to learn. You activate it by using two steering wheel buttons. First, turn on the adaptive cruise control. Once it has determined the car is on a limited-access highway with visible lane markings and a strong GPS signal, a grayed-out steering wheel symbol will appear to the right of the speedometer indicating it’s ready. Press the Super Cruise button, and the steering wheel symbol and light bar across the top of the steering wheel both turn green to show the car is now in command.
Let’s be clear: Super Cruise is not autonomous driving. Your eyes have to stay on the road for it to keep working. To pass slower traffic, you still have to look around and manually lane-change just like an ordinary car (unlike Autopilot which automatically steers you into the next lane after you’ve taken responsibility by flicking the turn signal). As you change lanes, Super Cruise’s green steering wheel lights turn blue until it reorients itself – whereupon it switches back to green, and the chauffeuring automatically resumes.
After about 150 miles, I’m deep into Central California where the roads are mostly straight with minimal on- and off-ramps. But Super Cruise follows the few sweeping bends with smooth steering adjustments. It also determines a safe speed for its radius; there isn’t the jitteriness we’ve seen with other camera-only lane-keeping technologies. Less to my liking is the noticeable latency of the system’s adaptive (radar-based) cruise control as it responds to a car ahead. Some may appreciate the gentleness in which the car varies the spacing in front, but to me it seems to react in slow motion. Call me impatient.
Nearing San Jose, the setting afternoon sun had me putting on my sunglasses. But Super Cruise’s steering-wheel-mounted infrared light emitters illuminate right through them to follow my eyes, regardless of my sunglasses or the time of day.
Although the Super Cruise performed admirably, it is not perfect. There are plenty of instances where a big tractor-trailer or an aggressive car will suddenly slice into my lane, and I’d have to take control long before Super Cruise would seem to realize that anything was awry. If your eyes are looking forward, but you aren’t paying attention to nearby drivers, this can really catch you out. This shortcoming highlights the need for the driver in any assistive driving technology cars to still be at the ready to react to unforeseen situations, especially in context of recent, fatal Uber autonomous vehicle and Tesla Model X AutoPilot incidents.
These nascent systems underline the importance of maintaining driver situational awareness. And the Super Cruise’s eye-tracking is definitely an improvement toward that goal. With many lane-keeping and self-steering technologies being introduced into the market and working seemingly so well under ideal conditions, the average Joe consumer can be easily lulled to overestimate the car’s self-driving capabilities. For me, it is already tempting after a few hours in Super Cruise on the Interstate 5 to think that I can grab some food and eat on-the-go, and just let the Cadillac do all the driving.
The AI and machine learning conference in San Jose is hosted by Nvidia – a company originally known for its supercharged computer graphics components, but for many years now has been investing heavily in powerful autonomous vehicle hardware and software. Within the McEnery Convention Center’s Main Hall, Nvidia’s CEO Jensen Huang, dressed in a black leather jacket and looking like a rock star, paces the huge stage as he delivers his keynote speech to conference’s 8,500 attendees.
Even though most of new features unveiled sound more like Star Trek’s Scotty describing to Captain Kirk how he can give the Enterprise more power, all I can catch (once in a while) is the term FLOPS. Nvidia has a new computer processing system that can reach two PetaFLOPS (that’s two thousand trillion of them). For reference, that’s a little over 2 percent of the speed of the world’s fastest supercomputer to-date (China National Research Center’s Sunway TaihuLight machine) and about 6,700 times faster than an iPhone 7.
What makes FLOPS the new horsepower? According to Danny Shapiro, Nvidia’s senior director of automotive, its SAE Level 5 full autonomous Pegasus in-car computer system can deliver 320 trillion operations per second, or 1,000 times the iPhone 7’s speed. Loosely translated, you need at least 0.003 PetaFLOPS of computing power for a self-driving car to shuttle you around.
Besides the than new high-speed FLOPS hardware, Nvidia also introduced its Drive Constellation Simulation System. This consists of two computer servers. The first is the Drive SIM that simulates how vehicle sensors like the camera, LiDAR and radar might see the road; the second is that same in-car Pegasus hardware processing all the information and doing the actual driving. By connecting the two, companies can add simulated miles to their actual self-driven test ones to raise the learning curve.
On the return back to Los Angeles, the more scenic California Highway 101 is my pick to see how Super Cruise performs on a less-mapped route. Threading through the various cities’ multiple highway interchanges, the Cadillac often indicates that it is ready to self-steer (with adaptive cruise control operating) but then immediately disengages when I press the button. At other times, such as through hilly parts of 101 (when I think it would be impossible for Super Cruise to engage) it would easily take over and self-steer. Not knowing exactly where the system would (or would not) work is very frustrating. Note to Cadillac: Perhaps overlay the Super Cruise’s availability on the car’s GPS map display, with specific reasons regarding the system’s inability to activate.
Nearing Los Angeles, I am feeling pretty confident with less than 100 miles left to go. But my mood soured quickly after spending nearly three hours creeping through traffic the rest of the way home. On one hand, it would have been worse without Super Cruise taking over most of the mundane stop-and-go duties. On the other, the brain power saved is now spent anxiously watching adjacent lanes for cars that may dive-bomb in front of me at any time.
Out of 1,000 miles total in the car, perhaps 25 percent of the time Super Cruise was fully active. But it was never one continuous drive with the system on – maybe a maximum of 20 to 25 minutes at a time, because something would happen and I needed to take over. Sometimes the system wanted me to drive; others, I wanted to pass slower traffic or take preemptive action.
There is a good argument to be made that any production-ready self-driving vehicle should carry Level 5 autonomy—no steering wheel, or accelerator or brake pedals—and nothing less. This way the driver can be completely stress-free and not worry about needing to take over control at any time. But it’ll be a few years before we see semi- or full-autonomous cars on the road. Until then, let the FLOPS wars begin.