Nvidia has been best known for building chips powerful enough for 3-D graphics and serious gaming, which still accounts for about half of the company's business. One of the newest (and fastest growing) focus areas for the Silicon Valley chipmaker is powering the artificial intelligence needed for self-driving cars.
We caught up with Danny Shapiro, senior director of Automotive at Nvidia, before the company talks to the Senate Commerce Committee this week about getting automated vehicles on the roads faster.
Here are a few excerpts from our conversation:
How did Nvidia make the shift from graphics and gaming to self-driving cars?
Graphics is a computing-intensive operation and it turns out the types of algorithms used for AI — more specifically deep learning — are very closely aligned with the types of math used for 3-D graphics.... We've been involved with the auto industry for two decades. First, automotive designers used our technologies to design 3-D models of cars. As processors became more powerful, we had 3-D models to simulate certain conditions, like a virtual wind tunnel and crash tests.
Over the last 10 years we've been adding our technology inside the vehicle, like the infotainment screens powered by mobile processors that you see in Tesla, Audi, Lamborghini, Rolls Royce, Honda and Mini…. Then the big bang of AI came about due to access to data, new algorithms and GPUs with the horse power to be able to train and execute these neural networks so AI can take place in real time. We've been working with automakers and suppliers and sensor systems to build the systems that replace the driver.
How does your relationship with carmakers work?
We've built an AI super-computer specifically for the car — our DRIVE PX product. We are developing a car computer platform — the brain — that ties to the cloud because it needs to be able to train in the data center first and then be able to run the results in real time in the car. The automaker is developing the application layer and user interface, and the ultimate decision of how the car will drive is up to them.
The brain takes in massive amounts of information and performs trillions of operations per second and ultimately outputs two numbers: Am I accelerating or braking? Am I turning left or right? It's able to continually train the vehicle. It learns from the data, it gets smarter, can drive at higher speeds, can drive on different kinds of roads that it might not have been able to do when you first bought the car.
When will we see widespread adoption of self-driving vehicles?
The timeline is really a function of regulation. Right now you're allowed to have autopilot like Tesla and Audi is rolling out Level 3 autopilot (where a safety driver is present to intervene if necessary) this summer, but there is now where where it is legal to have Level 4 or Level 5 technology (with no human is present) on the road. In California, where we are doing a log of testing, there always has to be a driver behind the wheel and be able to take over at any time. That's not what a Level 4 is, that's certainly not what a Level 5 (fully autonomous) is, so I think we need to see new regulation.
I think what we're going to see is an accelerated timeline. We have a lot of partners talking about 2020, for example, Audi announced Level 4 automated driving based on Nvidia AI. We announced with Toyota (to release) within a few years.
What else should policymakers do to help all this along?
The obvious one is the need for unified standards across our country as opposed to 50 sets of rules or regs. That's going to be a nightmare and make it very challenging for anyone who wants to cross state lines in an autonomous vehicle.
How much will consumer safety concerns impact adoption?
We are focused very heavily on safety and cybersecurity. Cars on the roads today that have been hacked have basically nothing put in them to prevent hacking. We're building it in at the hardware level to ensure cars are as tamper-proof as they physically can be.
There's an opportunity to put standards in place — certification and validation is very important — we can use both real road tested miles and miles on private proving grounds. And simulated miles I think will be very key, which ironically almost comes back to the whole graphics side where it's almost like you're putting cars into a video game to stimulate hazardous situations that you wouldn't want to subject humans to — cars running red lights, children running in front of a speeding vehicle — and the use of simulation as a way to both train neural networks and test and validate them before deploying vehicles.