In this presentation, you will learn the development flow and implementation considerations for moving from an academic CNN/deep learning graph to a commercial embedded vision design. The presentation will use practical examples that highlight the latest CNN graph mapping tool capabilities, including dispatched processing and pruning/compression. You will also learn about the cost vs. accuracy trade-offs of CNN bit width, balancing internal memory size and external memory bandwidth, and the importance of keeping data local to the CNN processor to improve bandwidth. Key deep CNN/learning benchmarks will be discussed including VGG16, Yolo, Denoiser, and more.
Dexian Li, Senior AE, Synopsys