Nvidia outs Pascal GPU in new “supercomputer” for self-driving cars

As with last year's CES, Nvidia devoted its keynote this year to machine learning and self-driving cars, unveiling a new version of its Drive PX system. But while Drive PX 2—which Nvidia describes as a "self driving supercomputer that fits nicely in your trunk," with the "processing power of 150 MacBook Pros"—will no doubt help partner Volvo with its first public trial of autonomous driving, it's what's under the bonnet that's exciting.

Drive PX 2, which consists of two SoCs paired to two graphics cards, is the first device from Nvidia to feature its new Pascal GPU and the first to be fabricated on TSMC's 16nm FinFET process. Not only that, but the new SoC is also quite exciting, mashing together two of the company's custom Denver CPU cores (as used in the Nexus 9 tablet) with four ARM Cortex A57 cores. The two GPUs are connected to the SoCs via a pair of discrete MXM (Mobile PCI Express Module) cards.

This configuration is a marked departure from the original Drive PX, which sported dual Tegra X1 SoCs with four Cortex A57s cores and four Cortex A53s cores along with an integrated Maxwell GPU.

On paper Drive PX 2 is an extremely powerful (and power-hungry) system. Unfortunately, Nvidia is keeping quiet on the full specifications of the Pascal GPUs, but it claims that the whole Drive PX 2 system will be capable of 8 teraflops of FP32 processing power, with a combined TDP of 250W. The latter is high for an ARM-based system, but presumably the twin GPUs account for most of the power draw. Nvidia, somewhat intriguingly for a computer that sits in the boot of your car, is cooling the Drive PX 2 with a custom liquid cooling system.

The claim that the PX 2 is "150 times faster than a MacBook Pro" is probably spurious. We've e-mailed Nvidia to ask how it arrived at such a figure, but it's probably something like a single MacBook Pro CPU against two Pascal GPUs.

As for why you'd need so much processing power in a car, Drive PX 2's first home will be in 100 autonomous Volvo cars, which the company plans to deploy on Sweden's public roads in 2017. Instead of hard-coding driving rules for the cars to follow, they'll use deep learning and computer vision to identify objects fed to them by a 360-degree camera array. The Drive PX 2 will be available to early access partners in Q2, with general availability in Q4.

For now, Pascal is only part of Drive PX 2, but with AMD having recently revealed its new GPU architecture Polaris, it's only a matter of time until Nvidia brings its new architecture to desktop GPUs. Current rumours point to the full Pascal unveil occurring in mid-2016.