Ambarella Rolls 4K Edge AI SoC Boasting High AI Performance-per-Watt
Seeking to enable better on-camera AI for the security realm, Ambarella’s new edge AI SoC promises high performance per Watt, improved image quality, and sensor fusion capabilities
In advance of the ISC West (International Security Conference West) event, running next week March 28-31 in Las Vegas, Ambarella today has unveiled its new CV72S chip. The device is a 4K, 5 nm AI vision system-on-chip (SoC) designed for mainstream professional security cameras.
The SoC is built on the company’s latest CVflow 3.0 AI architecture. The CV72S SoC was crafted to provide what the company claims is the security industry’s highest artificial intelligence (AI) performance-per-Watt SoC. It has dedicated AI hardware for running the latest transformer neural networks (NNs). Amberella says the SoC provides 6x greater AI performance than its predecessor, the CV22S.
The CV72S SoC offers 2x faster encoding and 6x higher AI performance than its predecessor the CV22S.
In this article, we review the key features of the new CV72S SoC, and we share perspectives from our interview with Jerome Gigot, Sr. Director of AIoT at Ambarella.
Demand for Better Camera AI and More
According to Gigot, the security market is hungry for advances in camera technology. “The security market is very big,” he says. “It's one of the biggest markets for cameras outside of cellphone and tablets, with a billion installed base of security cameras in the world.”
He says the CV72S offers a slew of features in demand in the security industry camera space, including higher AI performance, improved image quality, support for fisheye and multi-image cameras, and fusion of radar and visual sensing.
“What’s happening is that the market is transforming in terms of how they use AI,“ says Gigot. “People used to run AI on servers for the last couple of years. Now they're running AI on the camera directly for faster, better accuracy.”
Gigot says there are multiple reasons that security camera design engineers want to embed AI processing hardware directly into cameras. On the one hand, they want to run the latest and greatest neural networking to get better accuracy and better performance, while also using AI to increase the image quality.
To illustrate these new trends, Gigot cites the example of an airport parking lot, or any big parking lot. “That's a big open space and it's really hard to cover with traditional cameras,” he says. “You need to install a lot of cameras everywhere and somehow link them. Or, you can just have one AI-based camera installed and cover everything. That’s about people starting to look at fusion and what you can do if you add a radar next to a visual sensor.”
Today’s security camera designs are asking for more AI, better image quality, fisheye camera support, and advanced sensor fusion. (Click image to enlarge)
An example of that kind of sensor fusion can be seen in the “Sensor Fusion” image on the right in the above image. It shows the parking lot of a commercial building. If you look at the bottom, you see what the camera sees visually at night.
But the camera feed on top shows blue and purple blobs. Those are people, identified using 4D radar. “You actually have four people moving in different directions and the radar can pick that up and then feed that into the security system and the camera,” says Gigot. “That's where we see a lot of interest in technology that supports vision and radar sensor fusion.”
Different On-chip Engines With Different Jobs
At the heart of the CV72S is Ambarella’s latest CVflow 3.0 AI architecture. A key feature of this architecture is natively supported AI hardware that can efficiently run the latest transformer neural networks (NNs). Made famous these days by natural language processing technologies such as ChatGPT, transformer NNs are said to be able to outperform convolutional NNs in many vision tasks.
Another key technology on the CV72S is its ability to run Ambarella’s AISP technology. AISP is a blend of traditional ISP (image processing) and AI processing. It enables neural network-enhanced 4K, long-range color night vision and HDR at very low lux levels with minimal noise and no external illumination, while leaving plenty of headroom for additional, concurrent NNs—for person tracking and mask detection.
Key features and block diagram of the CV72S SoC. (Click image to enlarge)
The chip's high performance per Watt is a result of the CVflow architecture’s efficiencies, combined with using Samsung’s 5 nm process technology. This enables the CV72S to consume less than 3 W of power, says the company.
Fisheye and Multi-imager Camera Support
The CV72S supports advanced 16MP30 fisheye dewarping and 4x 5MP30 multi-imager AI capabilities. For single-imager cameras, the CV72S supports 4KP60 encoding for AVC and HEVC. Those features make the SoC well suited for mainstream security cameras, including smart city and traffic applications, as well as monitoring crowded areas, such as retail stores, malls and stadiums. It can also do advanced video analytics such as long-distance object detection and license plate recognition.
Ambarella’s Oculii virtual aperture imaging (VAI) AI radar technology was leveraged for embedding hardware acceleration into the CV72S for fusion with camera data, along with dynamic load switching between radar and cameras. According to Gigot, radar capabilities are critical for system designs like perimeter security and nighttime monitoring. They also make a huge difference when operating in rain, snow or fog.
Control functions on the CV72S are done by dual Arm Cortex-A76 1.6GHz cores. The SoC features high DRAM bandwidth, with support for 32-bit LPDDR4x/LPDDR5/LPDDR5x DRAMs. The SoC also embeds hardware security to prevent hacking, including secure boot, OTP, and also provides Arm TrustZone technology. Connectivity for the CV72S is enabled by high-speed PCIe and USB 3.2 interfaces.
Ambarella says the CV72S SoC is available now for sampling. It will be demonstrated during ISC West 2023 in Las Vegas next week.
All image used courtesy of Ambarella