Some people will think that there are many deep learning engine frameworks on the market, such as tensorflow, pytorch, caffe, etc. Why do you need to accelerate it? It is very simple. Although the deep learning technology is developing rapidly, there are also many big cows, but the deep learning framework is in Technically there is currently no convergence. This kind of thing is like big data analysis hadoop. Before, there was no mature framework for processing distributed data. After many years of precipitation, now hadoop has become an industry standard, and later memory computing uses spark. The same is true for deep learning frameworks. Because deep learning algorithms are not yet fully mature and are different from traditional hardware computing power (there is a demand for new chip computing power), deep learning frameworks have not really converged. Let's take a look at what Yuan Jinhui can do to optimize the deep learning architecture.
Mr. Yuan Jinhui first analyzed the advantages and disadvantages of the existing frameworks, and then shared his views on how to speed up the deep learning framework (data parallelism, model parallelism, pipeline parallelism).
Computing power requirements for deep learning
The State of Deep Learning Engines
Why Existing Deep Learning Engines Can't Run Efficiently
Optimal Architecture of Deep Learning Engines
Summarize
The video is reproduced from Tencent Video
Finally, reply " Neural Networks and Deep Learning Chinese Version " in the background and send you a book for free.
Swipe lightly and welcome to pay attention~