(video) Co-Design of Deep Neural Nets and Neural Net Accelerators for Embedded Vision Applications

Prof. Kurt Keutzer’s presentation at the 2018 CVPR Workshop on Efficient Deep Learning for Computer Vision

1. Co-design of DNNs and accelerators requires detailed layer-by-layer analysis of contemporary DNNs
2. Key to efficient DNN computation is data reuse
3. Different CNN/DNN layers have different patterns of data reuse
4. Different NN accelerator architectures favor different types of reuse (outputs vs weights)
5. We present a NN accelerator design, the Squeezelerator, that supports Weight-Stationary (WS) and Output-Stationary (OS) dataflow and matches the data reuse in the accelerator to the characteristics of the layer
6. Careful attention to layer-by-layer speed/energy of mobile DNNs resulted in up to 6x speedup over other accelerator architectures