There was a major emphasis on Tegra Automotive at the GPU Technology Conference (GTC) 2015. NVIDIA announced a new product, NVIDIA Drive PX, a $10,000 automotive development kit. The Drive PX has dual Tegra X1 processors which provides 2 TeraFLOPS of processing power. There are 12 (!) camera inputs. Looky here:
Another demo that was in the NVIDIA booth was the view from a set of cameras on a self parking vehicle and how the Drive PX builds a 3D map of nearby objects in real time. The idea is to enter a parking lot, hit a “Park Vehicle” button and let the car park itself. Very nice, I’m sure it’s better at parking cars than I am. My usual strategy is to step on the gas, close my eyes and wait until the car comes to a rest. Hopefully near a parking space of some sort.
Also in the NVIDIA booth was the self-piloted Audi A7 that was shown at CES in January. You’ll remember that the Audi had driven from San Francisco, California to Las Vegas, Nevada autonomously. The vehicle looks surprisingly stock, there are only a couple of visual cues that give away the nature of the vehicle. Maybe the back end gives it away?
However, open up the trunk and you’ll find that the electronic genies have made themselves at home:
There were several other cars out and about the exhibition hall that use Tegra as part of their electronic systems. There was a BMW i8, a Lamborghini Aventador (“Angry called, it wants its car back”), a Tesla sedan, a Local Motors’ 3D-printed car “Strati”. Also being showcased, a Renovo Motors coupe. On the Renovo, I liked how the batteries were disguised as an “engine block”.
Note: One of the things that I learned at the show is that I’m not very good at taking pictures of vehicles. So I’ll leave it to you to go out and find the automotive PRoN to supplement this article. I need to work on my photographing skillz.
Outside of the convention center, there was a line for test drives of several different Tegra equipped cars, including Teslas and BMW Minis.
As a major focus of the show, there were several good sessions from industry experts about some of the challenges facing integration of computing devices into the automotive fabric. In the GTC Keynote, NVIDIA CEO and co-founder Jen-Hsun Huang spoke about ‘training’ cars to drive by using Deep Learning techniques. This is a very interesting idea, but at the same time traditional computer vision techniques now have enough computing horsepower on board the vehicle to perform a lot of the tasks that are needed for Advanced Driver Assistance Systems (ADAS). I would say an equally, if not more, challenging problem is integration of different sensor types.
It’s pretty easy to imagine when image processing is not sufficient to guide a car. Imagine a heavy fog, or a particularly bad day for air pollution such as smoke or haze. Traditional cameras don’t give much information in these cases, where as radars and other sensors may provide useful information. How do you marry the two, and who do you “trust”? Let’s say you enter a fog bank, the image processors can’t make heads or tails out of what’s going on. Do they pass control back to the driver and say “Good luck!”?. Does the system use some other sensor type and try to synthesize a view of what’s going on up ahead? Thinking as a computer nerd, there are a lot of edge cases that have to be taken into account, and just talking with an engineer for 10 minutes requires that you think about such problems multi-dimensionally in terms of hardware and software integration.
One of the great things at the show were the Sessions. Industry experts giving technical talks. For the automotive track, here’s the recorded sessions:
I especially liked:
On the latter one, kids you may want to check with Dad before you start hacking away on the full size BMW out in your driveway. Just sayin’. I’m sure he’s cool with it, and will probably want to help, but check just to be sure. And, don’t mess with the airbags, those things can give you a hurting.
At the poster session, here was one I particularly liked:
which tested the computational performance of the Tegra K1 against a high-performance workstation.