CEVA launches deep neural network framework to accelerate the application of machine learning technology in low-power embedded systems

CEVA, the world’s leading licensor of cellular communications, multimedia and connectivity DSP IP platforms, announced the launch of the real-time neural network software framework CEVA Deep Neural Network (CDNN) to simplify the deployment of machine learning in low-power embedded systems . By utilizing the processing power of CEVA-XM4 image and vision DSP, CDNN enables embedded systems to perform deep learning tasks 3 times faster than leading GPU-based systems, while reducing power consumption by 30 times and required storage bandwidth by 15 times ( Note). For example, running a deep neural network (DNN)-based pedestrian detection algorithm on a 1080p video stream at 30 frames per second in a 28nm process requires less than 30mW of power.

The key to CDNN’s high performance, low power, and low storage bandwidth is the CEVA Network Generator. This proprietary automation technology can convert the customer’s network structure and weights into real-time, lightweight ones. Customize the network model to achieve a faster network model that can significantly reduce power consumption and storage bandwidth. Compared with the original network, its accuracy degradation is less than 1%. Once this custom embedded-ready network is generated, it can be run on the CEVA-XM4 image and vision DSP using a fully optimized Convolutional Neural Network (CNN) layer, software library, and API.

Phi Algorithm Solutions, a member company of CEVA’s CEVAnet Partner Program, has used CDNN to implement a CNN-based Universal Object Detector algorithm for CEVA-XM4 DSP. Now application developers and OEM manufacturers can use this algorithm for a variety of applications, including pedestrian detection and face detection for safety, advanced driver assistance systems (ADAS) and other embedded devices based on low-power camera functional systems.

Steven Hanna, president and co-founder of Phi Algorithm Solutions, said: “The CEVA deep neural network framework provides our convolutional neural network-based algorithms with a fast and smooth path from offline training to real-time detection, enabling us to obtain the information in just a few days. Optimized and unique target detection network implementation scheme, and the power consumption is significantly lower than other platforms. CEVA-XM4 image and vision DSP combined with the CDNN framework is an ideal choice for embedded vision devices, and is an artificial intelligence device using deep learning technology in the future Years of progress have laid the foundation.”

Eran Briman, vice president of marketing at CEVA, said: “So far, we have won more than 20 designs, continue to lead the industry in the field of embedded vision processors, and continue to improve our vision IP product portfolio to help customers push products faster. To market and minimize risks. Our new deep neural network framework for CEVA-XM4 is the first product of its kind in the embedded technology industry, and is for the development of feasible deep learning algorithms in power-constrained embedded systems. The personnel has made great progress.”

The CDNN software framework is provided as source code, which extends the existing application development kit (ADK) of CEVA-XM. It has flexible and modular features, can support complete CNN implementations or specific layers, and can be shared with various networks and structures, such as networks developed using Caffe, Torch or Theano training frameworks, or proprietary networks. CDNN includes real-time example models for image classification, positioning and target recognition, for target and scene recognition, advanced driver assistance systems (ADAS), artificial intelligence (AI), video analysis, augmented reality (AR), virtual reality (VR) ) And similar computer vision applications.

CEVA will host a real-time webinar on implementing embedded system machine vision applications on November 12, including an in-depth discussion of CDNN. To learn more and register for the web seminar, please visit the webpage http://www.linleygroup.com/events/event.php?num=35.

Note: Compare the operation on the most commonly used deep neural network AlexNet

About CEVA

CEVA is a leading authorized manufacturer of cellular communications, multimedia and wireless connection technologies to semiconductor companies and OEM manufacturers in the mobile communications, consumer electronics, automotive and Internet of Things markets. Our DSP IP product line includes a variety of comprehensive platforms, multi-mode 2G/3G/LTE/LTE-A baseband processing for terminals and base stations, computer vision and computational image processing for any device with a camera, and multi-oriented Advanced audio/voice processing and ultra-low power Always-on/sensing applications for the Internet of Things market. In the field of connectivity, we provide the industry’s broadest IP coverage for Bluetooth (Bluetooth Smart and Bluetooth Smart Ready), Wi-Fi (802.11 b/g/n/ac and even 4×4), and serial storage (SATA and SAS) field. CEVA’s IP is used in one-third of the world’s mobile phones, from top mobile phone OEMs such as Samsung, Huawei, Xiaomi, Lenovo, HTC, LG, Coolpad, ZTE, Micromax and Meizu.