Nvidia’s current Pascal GPUs are generating a lot of enthusiasm, but there’s even more to be excited about with their successor, the Volta GPU architecture, on its way next year.
Nvidia unveiled Volta in a new Xavier supercomputer chip designed for self-driving cars. The small surprise came Wednesday at the company’s GPU Technology Conference in Amsterdam.
Xavier provides unprecedented computing horsepower, allowing cars to recognize images, analyze on-road situations, and take action. Much of the processing will happen on an integrated 512-core Volta GPU.
Xavier will power the successor to Nvidia’s current self-driving car computer, called Drive PX 2.
The upcoming chip hints at Volta’s graphics capabilities. It has dual 8K HDR video recording, suggesting the new GPU architecture will explore areas beyond 4K graphics.
Nvidia in the past has said Volta will solve memory bandwidth problems and will be much smaller than predecessors. However, Nvidia hasn’t revealed further details about the GPU architecture.
Volta will also have a speedier NVLink 2.0 interconnect. The GPU will be paired with IBM’s upcoming Power9 chips in the U.S. Department of Energy’s Summit supercomputer in 2018. IBM has said that each NVLink 2.0 lane will communicate at 25Gbps (gigabits per second), which would be 7 to 10 times the speed of PCI-Express 3.0.
Xavier also has a custom eight-core CPU and a new computer vision accelerator. It’s the most advanced chip ever built by Nvidia, the company said in a blog post.
Xavier delivers 20 TOPS (trillion operations per second) of computing power while consuming only 20 watts of power. It has seven billion transistors and is made using the 16-nm chip process. Those specs hint big gains in performance and power efficiency for the Volta GPU.
“A single Xavier AI processor will be able to replace today’s Drive PX 2 configured with dual mobile SoCs and dual discrete GPUs—at a fraction of the power consumption,” Nvidia said.
The Drive PX 2, which was introduced at CES in January, has 12 CPU cores and two Pascal GPUs. It was the equivalent of having “150 MacBook Pros in your trunk,” said Nvidia CEO Jen-Hsun Huang.
Samples of the Xavier will ship to car makers in the fourth quarter of next year, along with technology companies, research institutions, and other automotive companies. Volta-based GPUs could be released earlier than Xavier.
Computers built with Xavier will allow self-driving cars to cruise the streets safely, Nvidia says. The company hopes Xavier will make cars smarter by recognizing what drivers encounter regularly—signs, signals, lanes, objects, and pedestrians—to avoid collisions and incidents with the traffic police.
But Nvidia has recognized there are challenges to overcome. For example, self-driving cars may not be effective after a heavy snowfall, where lanes may not be clearly visible and driving conditions unsafe. Xavier could train a car to do a better job of reacting to such situations.
Safety is a big concern for self-driving cars. The recent tragedy involving a Tesla Model S, where the driver was killed while the car was in partly autonomous AutoPilot mode, highlighted the importance of human attention even when a car is helping.
Computers based on Xavier won’t make safe driving decisions off the bat because they first need to be trained. The training happens as the computer absorbs images and locations and cross-references them with large repositories of images and data in the cloud. As the Xavier computer learns more about its surroundings, it can make smarter driving decisions.
Nvidia has equipped the Xavier its new computer vision accelerator. The current Pascal chips already have machine learning instructions, and the GPUs combine parallel processing with algorithms to help cars safely navigate streets.
Nvidia is doing well in the self-driving car market and betting a big part of its future on it. Volvo is developing an autonomous car based on Drive PX 2. Nvidia is also working with companies like Audi and Ford on self-driving cars. The company competes with Intel in the space.