Two AI Processors Push Computer Vision to the Next Level for IoT and Automotives
Hoping to create more power-efficient and highly functional AI processors, Syntiant and Intuitive are both aiming to create the next step in computer vision.
Between booming fields like autonomous vehicles and robotics, computer vision has become one of the hottest applications of artificial intelligence. Arguably more so than other AI applications, computer vision relies heavily on the underlying hardware, where software performance can be limited by the underlying processing units and imaging systems.
One of the biggest uses of computer vision is in autonomous vehicles. Image used courtesy of Sustained Quality
For this reason, EEs across the board are focused on pushing the boundaries of state-of-the-art and developing the best vision hardware possible. To go along with this push, multiple smaller-name companies, specifically Syntiant and Intuitive, have been aiming to make headlines.
This article will take a look at each company's latest advancement to get a feel for what's happening in the industry as a whole.
Syntiant’s Neural Decision Processor
Recently, California-based Syntiant made news with the release of its newest "Neural Decision Processor."
Designed for ultra-low-power applications, the new NDP200 is a special-purpose processor explicitly meant for deep learning. Its core architecture consists of a proprietary Syntiant core and an embedded Arm Cortex-M0 processor, which allows the chip to achieve operating frequencies up to 100 MHz.
Block diagram of the NDP200. Image used courtesy of Syntiant
The NDP200 was optimized specifically for power efficiency while running deep neural networks such as CNNs and RNNs, which are integral to computer vision applications.
According to Syntiant, the chip has been shown to perform vision processing with high inference accuracy while maintaining a power budget of under 1 mW. From a performance perspective, the chip could achieve an inference acceleration of over 6.4 GOP/s and support over 7 million parameters, allowing for edge computing of larger networks.
Syntiant envisions its chip to be used for battery-powered vision applications, including doorbells and security cameras.
This mesh between power efficiency and the ability to run deep neural networks has the potential to make this the next evolutionary step in creating better processors for computer vision.
Another company that is hoping to boost computer vision is Intuitive.
Intuitive Hardware Powers Electric Cart
Earlier this summer, Israeli company Intuitive received a boon when it announced that Fukushin Electronics would use its NU4000 edge AI processor in Fukushin's POLCAR, a new electric cart with a built-in obstacle detection unit.
Powering a sophisticated object detection unit in a battery-powered device like an electric cart requires a solid blend of performance and power efficiency, which is why Fukushin turned to the NU4000.
The NU4000 is a multi-core SoC that supports many on-chip applications, including 3D depth-sensing, simultaneous localization and mapping (SLAM), and computer vision.
Functional block diagram of the NU4000. Image used courtesy of Intuitive
To achieve these feats, the NU4000 integrates three Vector Cores which provide 500 GOPS. It also includes a dedicated CNN processor, which offers 2 TOPS, three CPUs, a dedicated depth processing engine, and a dedicated SLAM engine. This chip is built on a 12 nm process and includes an LPDDR4 interface and connectivity to up to six cameras and two displays.
Integrating all of this in a small form factor with low power makes the NU4000 sound like it has many key features that could be important for an application like Fukushin's POLCAR.
All in all, while many of the challenges of computer vision may be software challenges, the hardware that underlies these applications is equally as important. EEs are working hard to develop new, better computer vision hardware, and the news from both Syntiant and Intuitive are testaments to this point. With as much technology and innovation that these two companies are bringing to the table and into market applications, it will be interesting to see what will come out next in the way of computer vision.