Ambarella Shifts From GoPro to Robo

Source:Junko Yoshida
Share :

While GoPro suffers from the market saturation of mobile action cameras for sports enthusiasts, Ambarella, a very high-resolution image processor company that once generated as much as 30 percent of its revenue from GoPro, showcased at the Consumer Electronics Show its newly architected computer vision chip, CV1, primarily designed for highly automated vehicles.

On the eve of CES, GoPro announced plans to exit the drone business, cut 250 jobs and lower its fourth quarter revenue estimate. Fermi Wang, told EE Times that Ambarella has already experienced a decline in GoPro-based revenue. Last year, it was "10-plus percent," Wang said. He expects the company's GoPro revenue to sink to a "very low number" this year.

Making up for the lost revenue are the surveillance (professional and consumer) and auto OEM markets, he noted. Today, the company derives roughly 15 percent of the its revenue from the automotive sector.

Inside the CV1 SoC. Source: Ambarella
Click here for larger image

Inside the CV1 SoC.Source: Ambarella 

What Ambarella sees as its ace in the hole, though, is a new CVflow architecture that delivers stereovision processing and deep learning perception algorithms. Ambarella’s goal for CV1 and a series of new computer vision chips based on CVflow (to follow later) is to get a head start in the self-driving vehicle market while capturing other automotive applications, including ADAS, electronic mirror, and surround view.

Alberto Broggi

Alberto Broggi

In the summer of 2015, Ambarella acquired for $30 million VisLab, a startup spun from the University of Parma, Italy. A team led by Professor Alberto Broggi, a founder of VisLab, is the backbone of Ambarella’s AV software stacks for highly automated vehicles.

In Ambarella’s off-site demo in Las Vegas, Broggi showed off two cameras — a short-range monocular (up to a few meters), and a stereoscopic camera for views 150 meters. Both are based on CV1. By applying CNN, a monocular camera can detect and classify objects for known classes like pedestrian, vehicles, motorcycles. The stereoscopic camera detects generic objects —which the camera is not trained to classify — in 3D structures, much like the way a lidar sees things in point clouds.

Compared to a lidar that generates 2 million 3D points per second, Broggi said, the long-range stereoscopic camera captures "800 to 900 million 3D points per second."

Monocular cameras are detecting and classifying objects in yellow boxes. Source: Ambarella
Click here for larger image

Monocular cameras are detecting and classifying objects in yellow boxes.Source: Ambarella 

Stereo cameras are detecting untrained objects in green 3D structures. Source: Abarella
Click here for larger image

Stereo cameras are detecting untrained objects in green 3D structures.Source: Abarella 

Asked if Ambarella's computer-vision chip could replace a lidar, Broggi said that, although this is possible, it's not necessarily the company's intention.

Asked if Ambarella's computer-vision chip could replace a lidar, Broggi said that, although this is possible, it's not necessarily the company's intention.

The secret of Ambarella’s CV1 is its ability to bring in so much more information to computer vision, because CV1 supports computer-vision processing up to 4K or 8-megapixel resolution.

While the VisLab team brings advancements in deep learning to the CVflow architecture, Ambarella applies years of expertise in low-power HD and Ultra HD image processing to the CVflow.

Broggi told us, "There couldn’t have been a better union than VisLab and Ambarella." There is no overlap between what each team does. More important, CVflow exploits Ambarella’s image signaling pipelines for high-dynamic range (HDR) imaging, Ultra HD processing and automatic calibration in a stereo camera.

While not many companies talk about it, Broggi said stereo cameras need to be very stable. Calibration in stereo cameras can be a challenge, he said, especially in automotive applications, because cars vibrate and operate in a wide range of temperatures. With Ambarella’s new CV1 chip, "We do real-time auto calibration on the fly on the chip," he said.

There is no need for infrared cameras to process images in low light, either, he added. Asked about Foresight’s new quad-cam unit designed to fuse data coming from infrared (night vision) and day cameras, Broggi said, "We don’t need that. Our HDR can process images in very low-light conditions."

Ambarella's robocar. Source: Ambarella

Ambarella’s robocar.Source: Ambarella

But the clincher is that the CVflow architecture is fully-programmable and highly-efficient, providing significant computer vision performance with very low power consumption. CV1 runs at 4 watts, according to Broggi.

"You don’t need a powerful GPU to do all these things like CNN-based classification and stereovision processing for detecting generic objects without training," Broggi said. "We are doing all of it just on CV1."

The CV1 by itself is a very high-performance computer. Inside are Ambarella’s home-grown "engine" to power CNN and DNN, image DSP, two quad-core ARM Cortex-A53, and other accelerators, Broggi explained.

Asked about the price for CV1, Chris Day, vice president of marketing and business development at Ambarella, told us, "A lot cheaper than a GPU…below $50."

Ambarella already got its chip — fabricated by using a 14nm CMOS process technology — back from the foundry last May. The company is currently engaging with a number of customers in the automotive market, Day added.  

Key word:

Previous:PC Shipments Decline... Again

Next:Meltdown, Spectre Repeat Hard Security Lessons