Home  /  Products  /  CEVA's DSP Cores  /  CEVA-XM Family  /  CEVA Deep Neural Network (CDNN)

CEVA Deep Neural Network (CDNN)Print this page

A Deep Learning Toolkit for the CEVA-XM Family

CDNN is a comprehensive toolkit that simplifies the development and deployment of deep learning systems for mass-market embedded devices. Tailored and optimized for the CEVA-XM family of imaging and vision DSPs, the CDNN toolkit includes the CEVA network generator, the CDNN software framework and a CDNN hardware accelerator that work in tandem to deliver superior performance while ensuring flexibility to stay up to date with the constantly evolving domain of machine learning. Separately, each component of the CDNN toolkit is a powerful enabler of imaging & vision use-cases on embedded platforms. Combined, these pieces deliver all the critical components to support new imaging algorithms, network structures, and changing layer types of deep neural networks.

The CDNN toolkit streamlines implementations of deep learning in embedded systems by:

  • Automatically converting offline pre-trained neural networks to real-time embedded-ready networks for CEVA-XM cores, utilizing the CEVA network generator, a PC offline tool
  • Enabling real-time high quality image classification, object recognition and vision analytics
  • Delivering the flexibility to support various neural network structures, including any number and type of layers
  • Utilizing powerful hardware accelerators to maximize the processing throughput.
    Utilizing specific configurations of the CDNN toolkit along with the CEVA-XM DSP enables deep learning tasks to perform >4x faster and >25x more power efficiently than the leading GPU-based system, while requiring significantly less memory bandwidth for any network, including those generated using Alexnet and GoogLeNet.

Utilizing specific configurations of the CDNN toolkit along with the CEVA-XM DSP enables deep learning tasks to perform >4x faster and >25x more power efficiently than the leading GPU-based system, while requiring significantly less memory bandwidth for any network, including those generated using Alexnet and GoogLeNet.

Figure 1 CEVA Deep Neural Network Toolkit

CDNN2 Software Framework

A 2nd Generation Neural Network Software Framework to Accelerate Machine Learning Deployment in Low-Power Embedded Systems
CDNN2 Software Framework component is a real-time neural network software framework for embedded systems that harnesses the processing power of the CEVA-XM family of intelligent vision processors to implement deep learning applications.

CDNN2 enables localized, deep learning-based video analytics on camera enabled devices in real time. Coupled with the CEVA-XM family of intelligent vision processors, CDNN2 offers significant time-to-market and power advantages for implementing machine learning in embedded systems.

Real-time classification with pre-trained networks

CDNN2 receives a pre-trained network model and the relevant network weights as inputs from the offline training (e.g. via “Caffe” and “TensorFlow” frameworks) to the proprietary CEVA Network Generator which then automatically converts the network into a real-time network model. Convolutions Neural Network (CNN) based applications can then utilize this real-time network model via calling the CDNN2 real-time libraries and running them on the CEVA-XM family DSP cores.

Figure 2 CDNN2 Usage Flow with Caffe & TensorFlow

CDNN Hardware Accelerator

The CDNN HWA is designed to work together with the CEVA-XM family of processors as part of the CDNN toolkit to enable extremely high performance and very low power consumption on deep learning algorithms. The CDNN hardware accelerator delivers 512 MACs/cycle, ensuring best-in-class performance to handle today’s most complex neural networks. The CDNN accelerator relieves the DSP core of the heavy load of MAC operations, freeing the DSP native MAC units to perform additional parallel tasks.

Figure 3 CDNN2 SW Framework and CDNN HWA

The CDNN software framework and the CDNN hardware accelerator works in tandem to deliver superior performance while ensuring flexibility to stay up to date with the constantly evolving domain of machine learning. CDNN hardware accelerator supports 16bit precision, which most CNN applications demand as an input and output data from neural network layer. Using lower MAC with lower precision, will cause severe degradation in classification accuracy.

CEVA-XM family of Intelligent vision processors support additional hardware accelerators such as image de-warping and a variety of third party accelerators.
Learn more about CEVA Hardware accelerators

Deep Learning development for embedded systems using CDNN

CEVA supplies a full development platform for partners and developers based on the CEVA-XM family of DSP cores to enable the development of deep learning with the CDNN for embedded systems.

Target Algorithms and Applications

CDNN is intended to be used for object and scene recognition, automotive advanced driver assistance systems (ADAS), Artificial intelligence (AI), video analytics, augmented reality (AR), virtual reality (VR) and similar computer vision applications.