Abstract:
A realistic numerical model of a photonic tensor core based on the crossbar architecture with absorbing GeSbTe chalcogenide glass films as weight elements of a photonic matrix has been developed. The performance of the model for the matrix–vector multiplication has been demonstrated. The possibility of using the tensor core based on the implemented architecture in convolutional neural networks for image recognition tasks has been shown. Numerical simulations have been used for the first time to estimate the potential performance and energy efficiency of a photonic hardware accelerator, taking into account the modern experimental element base.
Fulltext:
PDF file (947 kB)
First page:
PDF file