Nvidia flops out TERAFLOP X1 for self-aware cars
New Tegra chip boasts 256 GPU cores, 8 CPU cores, and is baked into car kits
CES 2015 Nvidia revealed its new Tegra X1 system-on-chip processor, and hardware to power autonomous cars, at “day zero” CES press conference.
The new silicon features the Maxwell graphics engine, juiced up with eight CPU cores and 256 GPU cores, all available while consuming ten watts. Formerly known as “Project Erista”, the Tegra X1 is said to possess a teraflop of computing power.
The 64-bit ARMv8 CPU cores follow ARM's big.LITTLE idea: four high-performance Cortex-A57s, and four power-efficient Cortex-A53s, lashed together using Nvidia's own interconnect. The A57s share a 2MB L2 cache, while the A53s share a 512K cache.
Nvidia thinks this chip may find a home in some mobile devices, and also launched Drive CX, an in-car computer with a Tegra X1 on board.
The company's CEO Jen-Hsun Huang mentioned research suggesting the number of screens in cars will soar between now and 2020, when about 20 megapixels worth of display will be found behind the wheel. Huang said he expects mirrors could eventually become smart displays and pillars could house curved displays to create an illusion of transparency.
Whatever the future holds, Huang wants to be there: the Drive CX is therefore accompanied by a runtime called Drive Studio that will allow developers to create virtual machines that run on a Drive CX and take care of different tasks.
A demo showed off a dashboard and in-car entertainment system, each running in their own virtual machine in a single Drive CX.
The Tegra X1 can already drive to 4K displays at 60 hertz, or four full HD screens, and Huang feels that's enough grunt to give auto-makers reasons to get excited about Nvidia's offerings.
Huang's also betting that Tegra X1's grunt will help to make cameras, rather than sensor technologies like radar, car-makers' preferred method for helping cars to detect events on the road.
Today, he said, each sensor in a car is a standalone system. Huang hopes that cars of the future “see and understand what is happening around it” and can learn about the environment continually. To do that, cars need to be easily upgradable.
“We imagine all these cameras being connected to a platform,” Huang said on Sunday, then unveiled one called Drive PX, a dual Tegra X1 board boasting 2.3 teraflops capable of handling 12 cameras at a time and processing 1.3 billion pixels a second.
Drive PX bakes in “deep neural network” technology to help cars make decisions like classifying objects on the road and Huang said the board correctly identifies the things it sees 93 per cent of the time and is even able to learn the difference between a truck and a fire engine, then change a car's behavior accordingly.
Drive PX and CX put Nvidia squarely in competition with Intel, which last May revealed a car kit it said bridged the gap between in-car automation and autonomous vehicles.
Nvidia looks to have leapfrogged that idea and given car-makers something rather more substantial to contemplate. Developers for other applications may also be interested: Nvidia didn't mention the hypervisor it will use (we're guessing Xen's embedded version aimed at cars), but did say it can run QNX, Android or Linux in the Drive CX. ®