Nvidia revealed its brand new Tegra X1 mobile SoC at CES, alongside an X1 powered in-vehicle infotainment (IVI) center called the Drive CX, and a vision of autonomous cars powered by two X1s, on the Drive PX computer vision and self-learning system aimed squarely at vehicular autopilot use.
A year after revealing the Kepler K1 chip, which unified Nvidia’s PC and mobile architecture, and then 4 months after unveiling the new Maxwell architecture in its desktop video cards, CEO Jen-Hsun Huang declared that the X1 is “unquestionably the most advanced GPU we have ever built. This little tiny thing you see here is a mobile super chip. The world’s first mobile chip that can do 4K video.”
The X1, which derives its product name from its architecture, brings desktop-capable processing to mobile platforms – and any other platform in which the chip is installed (such as automotive, but more on that later).
To illustrate the point, Huang noted that the X1 can run the same technical demo that the Xbox One used at its unveiling, the Unreal Engine 4.0 Elemental Demo. The console uses around 100W to run the benchmarking tool, while the X1 manages the same test using its 10W power envelope – using HDR lighting (which requires each pixel to have a floating point RGB value), and all the physics and particle engines found on the larger device versions.
The new chip has a 256-core GPU, an 8 core 64-bit ARM CPU, which can handle 4K video with a 10-bit color depth at 60fps, with support for both HEVC and VP9 codecs. Nvidia claims that the X1 solidly outperforms Apple’s A8X chip, which is on par with the K1, according to Nvidia’s figures. It is cut in 20nm, and cast by TSMC.
The X1 architecture is the first Nvidia GPU to use a new format called FP16 (16-bit Floating Point). When all 256 GPU cores are combined, the X1 can achieve 1 teraflops of computational throughput, which Huang notes makes it the world’s first teraflop capable mobile chip. However, Nvidia better hope that the visible jerkiness seen on its Unreal test video, that all watching the CES video would have seen, was down to an explanation other than the chip not being powerful enough – as that rather defeats the claims it and Nvidia make.
Huang then explained to the CES audience what Nvidia planned to do with all this horsepower and introduced Nvidia Drive CX, a Digital Cockpit Computer that is powered by the X1. Able to process 16.6 megapixels, enough for two 4K displays or 4 1080p screens, with the ability to run each screen as a separate virtual machine to simplify programming. Nvidia wants the unit to serve as the heart of next-gen in-vehicle infotainment (IVI) systems.
Moving into the automotive sector is an attempt to diversify Nvidia’s portfolio. After losing out to AMD in the new Xbox and PlayStation (and likely the next Nintendo console), Nvidia still hasn’t managed to gain much of a foothold in the mobile sphere, although it’s unlikely any single company stands to win much ground against the heavily entrenched Qualcomm and its Snapdragons – which are also found in many IVI systems.
The Drive CX supports BlackBerry’s QNX, the leader in the IVI market, as well as Linux and Android Auto. Huang seemed to catch himself before mentioning Apple, which was conspicuously absent from the presentation, but he did go on to explain the Nvidia Drive Studio software package, which allows OEMs to import designs into the Drive CX rendering environment – for dashboards, UIs, and navigation systems, all powered by the X1’s GPU grunt.
The Nvidia CEO said that the company believes the cars of the future will be the most advanced computers in the world, with more computing horsepower, sensors and rich displays than anything currently existing. Huang added that the number of displays inside cars is due to grow rapidly, from a combined resolution of 380,000 pixels in 2010 to well over 20 million pixels by 2020, largely thanks to consumers becoming more used to high-res displays in their other devices. Nvidia’s core business is providing the silicon that drives these displays, but it is slowly branching out into new areas, on the back of its GPUs.
Huang then introduced the audience to Nvidia’s Drive PX – a computing module that Nvidia hopes will act as the brains of a car outfitted with up to 12 Full HD 60Hz camera inputs, and able to process 1.3 billion pixels per second (1.3 gigapixels/second). The board features two X1 processors, which can work in parallel or in redundancy modes, for up to 2.3 teraflops of computing throughput.
Huang said that the PX introduces two revolutionary new technologies: a deep learning system and a computer vision system, collectively called the Deep Learning Neural Network Computer Vision system. The combination of hardware and software is being badged as the Drive PX Auto Pilot Computer.
The 1.3 gigapixels per second capacity for image capture requires a hefty piece of silicon to analyze, which Nvidia hopes will be the X1. Huang said that neural networks had shot up from 74% image recognition accuracy to well over 90%, while cutting the batch-processing times down from weeks and months to days and hours, thanks to GPU-acceleration.
While Nvidia can provide the brute force needed to analyze the images, it has so far not mentioned how it plans to build the cloud infrastructure for the automotive companies or trade bodies that would be in charge of the self-driving vehicles on a road. This is an ambitious project, but given the power available in both the new chips like the X1 and in cloud instances, it looks like bureaucracy and contract negotiation would be the more likely stumbling blocks, rather than the capabilities of hardware or software. Whether societies are receptive is also a major hurdle, but data and test cases should assuage any fear of robo-cars.
The region-specific visual data can be pre-loaded into the PX so that each car doesn’t have to learn the roads once it has left the showroom, and similarly the databases can be updated as time progresses. This process can lead to a crowd-sourced database, generated from multiple cars, which can be updated on the fly – allowing cars to query the datacenter and receive answers to images it does not recognize. In Nvidia’s demo, the car readily recognized crosswalk signs and pedestrians, although we did spot it confuse a Chrysler PT Cruiser with a police sedan momentarily.
Because these cars will need to be able to learn and adapt, they need to be able to update their software and image databases on the go, according to Huang, who said “the car of the future is going to become a software-defined car.” This need for connectivity is a potential new market for cellular providers, who are better placed than WiFi to provide high-bandwidth streams to fast-moving vehicles – data which will be used for OTA updates as well as the more typical entertainment uses.
AT&T is still bringing new partners into its Georgia-based Drive Studio project, and will no doubt be enquiring about the naming of Nvidia’s Drive Studio software package, but many other cellular players have already entered the market – with both operators such as AT&T and Verizon, but also the established M2M businesses such as Cohda Wireless and recent Vodafone acquisition Cobra.
But returning to the issue of self-learning vehicles, Nvidia then brought Ricky Hudi from Audi on stage to discuss Audi’s vision for the Nvidia product and autonomous, who then revealed that a self-driving car was currently driving from Silicon Valley to Las Vegas.
Nicknamed Jack, the Audi A7 sedan has now successfully completed its 2-leg trip, accompanied by five journalists, using 3 cameras, 6 radar units and two light detection and ranging (LIDAR) units – with the prototype technology due to appear in Audi’s cars within three to five years. Nvidia and Audi used CES 2014 to announce their partnership on the zentrale Fahrerassistenzsteuergerät (zFAS) computer that manages Jack’s systems. Nvidia will be hoping that OEMs like Audi add more cameras to their vehicles – although it should be noted that Audi never actually said it had adopted the Nvidia system or planned to, while on stage.
The final part of the key note saw the culmination of the above technologies. Called Surround Vision, and demonstrated in a photorealistic simulation, the system used simulated cameras on the car to navigate a parking garage and find an empty parking spot – without the need for the driver to be in the vehicle. This feature is currently being called Auto Valet, but it highlights the ways in which vehicles will very quickly be able to operate without a human at the wheel. Now we just have to see legislative frameworks, insurance bodies and consumers get onboard.