AI chips can be divided into two types by function: Training(training) and Inference(inference) :
Training usually requires a large amount of data input, or reinforcement learning and other unsupervised learning methods to train a complex deep neural network model. Because the training process involves massive training data and complex deep neural network structure, it requires a huge amount of computation and a huge computing scale, which requires high performance such as computing power, accuracy and scalability of the processor. At present, the training session is mainly completed by NVIDIA GPU cluster, and the ASIC chip TPU2.0 independently developed by Google also supports the deep network acceleration of the training session.
Inference link refers to the use of trained models and the use of new data to \"reason\" a variety of conclusions, such as video surveillance equipment through the background of the deep neural network model to determine whether a captured face belongs to the blacklist. Although the calculation amount of Inference is much less than Training, it still involves a lot of matrix operations. In the reasoning process, Gpus, FPgas and ASics have many application values.