Analysis Of Tensor Processors Application In Machine Learning Tasks On The Example Of GOOGLE TPU
N. Kravets, Артем Ховрат, Nataliia Saichyshyna · Bionics of Intelligence · 2020
A detailed analysis of the tensor processor by Google was carried out, its mathematical basis, structural components,and key work stages for use in solving problems related to machine learning were considered. Methods for speeding upthe neural network training process without loss of quality implemented in TPU are considered: quantization, parallelprocessing, systolic array, and the mechanism for encapsulating calculations in neural networks. The analysis of thelimitations and advantages of this type of the processor as a whole and in comparison, with the graphics and central processors. The competitive advantages of this tensor processor with analogues offered by other companies are considered.The interaction with the Google cloud platform and with the TensorFlow software library is described.