Smart Home and Automotive are the two hot topics at CES this year, with AI only really making any impression on the back of announcements regarding digital assistants. However, two silicon announcements are significant in both the enterprise and consumer markets – with CEVA unveiling its new NeuPro family of network-edge AI processors, and Ambarella launching its new CV1 computer vision range.
For CEVA, this is about snaring a share of the low-power devices that want to adopt AI-based processing – outside of the data centers. CEVA is pitching it as a deep learning option for edge-device vendors ‘looking for a streamlined way to quickly take advantage of the significant possibilities that deep neural network technologies offer.’ NeuPro is another tool in CEVA’s AI box, which also includes its XM4 and XM6 machine vision DSPs (Digital Signal Processors), and its CDNN software framework, which CEVA says is being used by dozens of customers in consumer, surveillance, and ADAS products.
On the topic of machine vision, Ambarella’s CV1 4K Stereovision Processor, which uses its CVflow Computer Vision software architecture. The first in a new family of new chips, the CV1 will be aimed at autonomous vehicle, ADAS, security cameras, and drones. Ambarella is a fabless semiconductor company that has evolved from video encoding chips into consumer video cameras, and is best known for powering GoPro’s cameras. In 2015, it acquired VisLab, a specialist in computer vision systems for autonomous vehicles, and the area has become one of increased interest for the company.
CEVA’s big claim is a notable increase in performance, from the dedicated AI NeuPro processors – ranging from 2 TOPS (Tera Operations per Second), to 12.5 TOPS in its most advanced configuration. Machine Vision seems to be the main application, but CEVA is also citing the NeuPro’s suitability for natural language processing, real-time translation, authentication, and workflow management – meaning the chips are likely going to appear inside Human-Machine Interfaces (HMIs), and smart home devices that house digital assistants.
In terms of the hardware, CEVA has aimed to provide a modular approach for developers – based around four processors (NP500/1000/2000/4000). The different tiers of processor house different amounts of MAC units (multiply-accumulate unit), with the NP500 aimed at IoT devices, wearables and cameras, and the NP1000 aimed at mid-range smartphones, industrial applications, and AR/VR headsets. The next step up, the NP2000, is aimed at premium smartphones, surveillance cameras, robotics, and drones, and finally the NP4000 (with its 4096 MAC units), is being pitched at enterprise surveillance applications and autonomous driving.
Using the NeuPro engine, the hardware implementation of CEVA’s neural network layers, and the NeuPro VPU, a vector DSP to manage the CDNN software, the new family supports both 8-bit and 16-bit neural networking applications. In addition, CEVA says it will be offering the NeuPro engine as an accelerator option for developer that want to add just that piece to their XM4 or XM6 platforms.
The new CEVA designs have been built for a 16nm manufacturing process, and will be available for license in Q2, and availability in Q3 2018. In terms of performance, CEVA says the NeuPro chips are 30x better than the XM4 and 20x better than the XM6, when running the ResNet-50 deep neural network.
The launch is the first non-DSP processor from CEVA, a fabless semiconductor designer comparable to ARM – in that it licenses its chip designs but doesn’t build them. Founded in 2002, its designs are used in LTE, WiFi, and Bluetooth processors, as well as vision and audio. CEVA says that 8bn chips containing its IP have shipped to date, and by addressing the emerging AI and edge-processing space, CEVA is hoping to expand into a lucrative new market.
Ilan Yona, VP and GM of the Vision Business Unit at CEVA, said, “it’s abundantly clear that AI applications are trending toward processing at the edge, rather than relying on services from the cloud. The computational power required along with the low power constraints for edge processing, calls for specialized processors rather than using CPUs, GPUs or DSPs. We designed the NeuPro processors to reduce the high barriers-to-entry into the AI space in terms of both architecture and software. Our customers now have an optimized and cost-effective standard AI platform that can be utilized for a multitude of AI-based workloads and applications.”
Turning now to Ambarella’s new CV1, CES attendees might have been able to see it in action inside a Embedded Vehicle Autonomy (EVA) car and a drone – in which the CV1 would have been providing stereo processing for the cameras. With support for 4K and 8K inputs, the chip is designed to provide object recognition at extended distances, with very low power consumption.
The CV1-powered SuperCam cameras provided a 360-degree short and long-range perception and self-location – very important to avoid crashing into things. The full list of features includes 4K stereovision perception to 150-meters, with stereo obstacle detection at 180-meters, with automatic calibration, stereo generic obstacle detection, terrain modeling, traffic light detection, 3D free space detection, lane detection, and CNN classification for vehicle/pedestrian/bicycle detection. The fully autonomous SuperDrone demo featured much the same, but for aerial platforms.
“We are delighted to introduce CV1, the first in a new family of 4K computer vision processors based on our CVflow architecture,” said Fermi Wang, CEO of Ambarella. “CV1 combines Ambarella’s traditional strength in high-resolution imaging for human vision with stereo and neural network processing for advanced computer vision. It will provide customers with a highly programmable and high-performance platform to develop the next generation of intelligent vehicles, drones and IP security cameras.”
As well as the new chip, Ambarella is providing a suite of developer software tools, designed to let them easily port their own neural networks into the CV1 chips. This is very good news, if it works as promised, as the training time for neural networks (in which the computers are left to churn through vast datasets, and then calibrated and sent out to do it again), represent a pretty significant development cost.
Ambarella will also be looking to diversify into these new applications because of the troubles at GoPro, one of its biggest customers, which has recently cut its Karma drone program after a pretty disastrous few quarters. Rumors of a GoPro acquisition have also abound, in the past week, with the CEO confirming that he was very open to selling the business to a larger parent.