RECONFIGURABLE OPERATION ACCELERATOR OF LOGICAL 3D STRUCTURE FOR CONVOLUTIONAL NEURAL NETWORK

摘要

According to an embodiment of the present invention, as a reconfigurable operation accelerator of a logical 3D structure for a convolutional neural network, provided is the operation accelerator for a convolutional neural network, including: a shape adapter unit for reading a 3D array given as an input and outputting a specific 2D array; an input-reuse network unit for shifting the 2D array outputted from the shape adapter and outputting the 2D array required in the next step for a specific cycle; a parallel calculation unit for performing a convolutional operation with a weight value by using the 2D array obtained from the output of the input-reuse network unit as an input; an on-chip buffer unit for storing some or all of an input array, a weight array, and an output array; and a configuration value selection unit for determining an optimal parameter combination of a calculation tile and a data tile. Accordingly, the present invention can increase an operation speed and operation resource utilization in the convolutional neural network.;COPYRIGHT KIPO 2018