Energy-Efficient Hardware for Deep Convolutional Neural Networks
This talk will describe methods to enable energy-efficient processing of deep convolutional neural networks (CNN), which are the cornerstone of many deep-learning algorithms. While CNNs deliver record-breaking accuracy for many computer vision tasks, they require significant computation resources due to the size of the networks (e.g. hundreds of megabytes for filter weights storage and 30k-600k operations per input pixel). In this talk, we will discuss how to efficiently manage these large networks using a new CNN dataflow, called row stationary, that maximizes data reuse to minimize data movement both on- and off-chip for optimal energy-efficiency and throughput. We will present our reconfigurable accelerator named Eyeriss that supports the energy-efficient row-stationary dataflow, as well as exploits data statistics, making it 10x more energy-efficient than most mobile GPUs. The chip has been integrated into a real-time image recognition system and can be configured to efficiently support a wide range of CNNs including AlexNet and VGG-16.
|Energy-Efficient Hardware for Deep Convolutional Neural Networks|
Tuesday, Jan. 10, 2017, 4 p.m.–5 p.m. ET
Durham, NC, United States