NVIDIA's sequel to the Drive PX in-car computer it debuted last year is a liquid-cooled beast with the power equivalent to 150 MacBook Pros. Say hello to the Drive PX 2. It sports 12 CPU cores and has 8 teraflops worth of processing power -- similar to about 6 Titan X video cards. NVIDIA also claims that it can achieve 24 trillion operations a second, which should make it particularly useful for powering self-driving cars. Jen-Hsun Huang, NVIDIA's CEO, says it's the first supercomputer made for cars -- it's certainly the first we've seen with such insane specs.
So why the need for all that horsepower? Huang says it's all necessary for building capable self-driving cars. In particular, deep learning is going to be essential. It doesn't matter how well we program a car's maps and sensors -- it also needs to know how to deal with changes on the fly, like when a child jumps out on the road. Deep learning will allow self-driving cars to train themselves over time for all sorts of unexpected scenarios.
"We aren't realize the full potential of our vision unless we can solve city driving," Huang said. "Bikers are in the same road you are... People are sometimes following the rules and most times not... It's very chaotic and very hard."
With the Drive PX2, NVIDIA is doubling-down on the vision it had last year: Instead of just enabling high-end PC gaming, NVIDIA's GPUs can also be used to perform the crazy levels of computation necessary for self-driving cars.
NVIDIA also announced DIGITS, a deep learning platform that it's already testing with its own self-driving cars. It's basically a way for self-driving cars to take everything their learn and share it with a cloud-based network, which can then be sent to other cars, DIGITS is meant to make life easier for car companies as they start testing autonomous driving. Huang assured us that car makers will still own their own deep neural networks -- it's basically just getting the process started.
DIGITS helped NVIDIA produce its own deep neural network, dubbed Drivenet. It features nine "inception layers," which Huang described as nine separate neural networks embedded with in each other. Running information through the network just once takes 40 billion operations -- so yah, there's a lot of computing power behind it.
Drivenet is able to identify five different classes of objects, including pedestrians and motorcyclists (you can see how they're color coded above). That sort of recognition is going to be particularly important for autonomous driving. Huang noted that Audi used Drivenet to analyze visual data taken from a snowstorm, and after just one night they were able to detect data that the human eye can't see.
Ultimately, all the computing power in the world doesn't matter if nobody actually adopts your technology. Huang revealed that Volvo is the first company to adopt the PX2 officially, which is a decent start. But come next CES, NVIDIA will need more than just new hardware: It'll need to prove that it can actually land plenty of willing customers.
Roberto Baldwin contributed to this report.