Like always, Nvidia kicked its CES routine traditionally with a conference this morning at Las Vegas Convention and World Trade Center. When the company’s CEO, Jen-Hsun Huang announced: “We’re gonna talk about self-driving cars today,” the immediate understanding of the announcement of a cool GPU for handling self-driving operations was delivered and as expected the company unveiled its most powerful ‘computer’ and they are calling it the Drive PX2 which is followed by its last year’s Drive CX.
After numerous slides with flowcharts and diagrams describing the new technology, the real deal (specs) were revealed and Drive PX2 is reportedly next to a ‘supercomputer’ with a processing power equivalent to 150 MacBook Pros. When we uncover the analogy, it can deliver up to 8 Teraflops of throughput and 24 deep learning tera operations per second. The machine has 12 CPU cores with latest Pascal GPUs and all packed in the size of a child’s lunchbox.
The main point behind this heavily loaded piece of hardware is that Nvidia believes that the concept of self-driving in its roots depends on handling and processing of huge amount of information in small time so that it can handle all the sensors, learning and the control it needs to actually be fully autonomous with very few or no dependencies. Drive PX2 is so powerful that it controls 12 video cameras, lidar, radar and ultrasonic sensors at the same time. The hardware requires water cooling mechanism. Well, we don’t want it to melt while driving on a highway, right?
With this announcement, the Huang announced that the company is already working with Volvo, its first partner, to use the Drive PX2 in some of their test self-driving vehicles. Apart from Volvo, BMW and Audi are also in the list which will be deploying this lunch-boxed powerhouse into their future self-driving vehicles.
Huang said, “Self-driving is hard [because] the world is complex and unpredictable” and what comes to the rescue? A deep neural network, huge amount of data and processing power. The processing powerhouse being the Drive PX2, Nvidia also announced its deep neural network called NvidiaDrivenet which is composed of 37 million neurons spreading across nine inception layers and three convolutional layers which carry out 40 billion operations for detecting single class or multi-class objects.
The demos were pretty promising,even recognizing the cars in heavy snowfall which a human eye can easily overlook. Let’s see what Nvidia has given birth to comes out to be. Is it going to be as cool as the CEO’s leather jacket he wore today?