Neural network processor designed for edge devices

Overview

Key Features

• High energy efficiency
• All series reach higher than 1.5 TOPS/W.
• Support mainstream deep learning frameworks
• Caffe, Keras, TensorFlow, and ONNX.
• Low power consumption
• Under 0.5W and can be less than 5 mW for specific applications.
• An integrated AI solution
• Include hardware IP, compiler, and model compression.
• Deep compression technology
• Compresses not only models but also data and coefficients during computing to reduce memory use.
• Filter decomposition and convolution acceleration
• Divides a large-scale convolutional block into a number of smaller ones for parallel computing and then integrates and accelerates the results.
• CNN model support optimization
• Supports diverse CNN models, including Vgg16, Resnet, GoogleNet, YOLO, Tiny YOLO, Lenet, MobileNet, and Densenet with model specific performance optimization.
• Interleaving computation architecture
• Enable parallel convolution computing and pooling to improve overall performance. The convolution layer can support both 8- and 16-bits fixed points concurrently.
• Adaptive data structure
• Adjusts the data structure dynamically to improve MAC efficiency depending on changing demand.
• Dynamic storage resource configuration
• Allows more efficient resource allocation between shared memory and operational memory. Increases storage resource utilization without affecting performance.

Technical Specifications