TECH REPORT
How parallel processing is key to AI learning to drive
The day looms closer when autonomous vehicles will be let loose on our roads, as the legal implications, safety concerns, and technical considerations get ironed out. Movement on the last of those has been swift, with progress from companies such as Google, Tesla, General Motors, and MIT’s Computer Science and Artificial Intelligence Laboratory. Some vehicles, though, are powered by a name familiar to anyone who’s played a game on PC in the past 20 years. Nvidia’s entry into the autonomous vehicle market might seem to be an odd choice for a company more commonly associated with texture fill rates and polygons-per-second, but it’s the massive parallelism of its processors, as well as their relatively low power requirements, that makes them suitable.
Autonomous vehicle technology is divided into levels, with Level One being something as simple as cruise control, and Level Two taking that a step further with automatic parking or adaptive cruise control that stays a set distance from the car in front. Level Five is a fully robotic car that may not even have a steering wheel, and which can drive itself in areas that don’t have hyperdetailed mapping or any kind of geofencing to help it out. Cars currently available are around Level Two, with anything higher being the domain of Tesla, high-end Audi, or Mercedes, and those used purely for testing.
There is, after all, quite a lot for an autonomous vehicle to think about, with one estimate putting the computational needs of a driverless car at 50 to 100 times what’s required in today’s most advanced vehicles. It’s receiving data from multiple sources—cameras, radar, even lidar on some. Then there’s GPS information, and the requests of passengers for radio stations and other in-car entertainment, which could be managed by an AI that’s separate from the one used to drive. Whatever system is managing the car needs to keep up.
Parallel lines
Nvidia’s solution isn’t exactly like the GPU you may have in your PC, but it uses similar properties.The most important of these is the GPU’s capacity to perform mass calculations. Nvidia’s senior director of automotive, Danny Shapiro explains this in terms of cores, which you can think of as lanes where data can travel in order to be processed. A GPU can have thousands of cores (a GTX 1080Ti has 3,584) and thus perform massive, repetitive calculations. That’s useful for gaming (and cryptocurrency mining), but it’s also really well suited to the kinds of calculations the vehicle AI need to perform. Wherever you find AI and neural nets, you also find Nvidia’s specialism; highly parallel processing.
“Nvidia has moved beyond 3D graphics for gaming,” Shapiro says. “We have products that go into the data center, GPUs that can do all types of number-crunching. Google,
Facebook, Netflix, you name it, their data centers are full of Nvidia GPUs. Whenever you say ‘okay Google’ or ask Siri something, it’s Nvidia GPUs that are doing the processing, trying to understand what is said, what the context is, and apply what is called ‘deep learning’ to figure that out and return the results.”
The likes of Siri have an easier time, switching your lights on and off and telling you jokes, than a vehicle AI would. There’s a clear parallel between gaming and autonomous vehicles, as the AI uses the input from its sensors to build a virtual environment that reflects the world around it, deciding for itself where it’s safe to drive. AI neural nets are trained in virtual reality for this, unable to distinguish between the VR input and the readings of their own sensors.
Nvidia has been involved with the car industry for longer than you may think: Its products have been used for vehicle design for 20 years. “The same technology that’s used in PCs for gaming is used in workstations for 3D graphics and CAD,” says Shapiro. “Pretty much every stage of vehicle design, development, engineering, testing, marketing, and sales, Nvidia tech is in all those different departments. It was probably about ten years ago that we started working with the auto makers to actually bring our technology into the car, and that first foray was taking our mobile graphics processors … [they] were going into the cars to drive the screens. So what we did was to develop an onboard computer, bringing a consumer electronics level of interaction into the car.”
Power demands also need to be factored in. A GTX 1080 Ti pulls in over 270 watts while playing a game, and needs suitable cooling. Any power drawn by the hardware is power that’s not being used to move the car, so efficiency is key, especially as internal combustion engines give way to batteries. “Inside of a car, there are different operating conditions than for your home PC,” says Shapiro. “So we go through a different manufacturing process and testing and QA process to ensure that our automotive products will work in freezing temperatures in the middle of winter, on a hot summer day in the desert, as well as in a harsh operating environment with the shock and vibration and dust.”
Level Five is alive
The systems Nvidia is putting into this environment go by the name Nvidia Drive PX, and range from a single, passively cooled chip that draws no more than 10 watts of power, to Xavier, a system-on-a-chip that marries an eight-core ARM-derived CPU with 512-core Volta GPUs, and finally to Pegasus, which straps two Xaviers together, adds two more GPUs, and is capable of 320 trillion operations per second. This platform is designed for Level Five operations, and Nvidia claims more than 25 companies are developing robo-taxis based on the tech.
Level Five vehicles are a decade away at least: The technology is not yet perfected, and the slowest part of all may be governments passing the relevant legislation to allow them to operate, but the days of driving completely manually are definitely numbered. Ian Evenden
Any power drawn by the hardware is power that’s not being used to move the car