公开(公告)号：US20220092393A1

公开(公告)日：2022-03-24

申请号：US17027628

申请日：2020-09-21

Applicant: HEWLETT PACKARD ENTERPRISE DEVELOPMENT LP

Inventor： GLAUCIMAR DA SIKVA AGUIAR ,  FRANCISCO PLÍNIO OLIVEIRA SILVEIRA ,  EUN SUB LEE ,  RODRIGO JOSE DA ROSA ANTUNES ,  JOAQUIM GOMES DA COSTA EULALIO DE SOUZA ,  MARTIN FOLTIN ,  JEFFERSON RODRIGO ALVES CAVALCANTE ,  LUCAS LEITE ,  ARTHUR CARVALHO WALRAVEN DA CUNHA ,  MONYCKY VASCONCELOS FRAZAO ,  ALEX FERREIRA RAMIRES TRAJANO

IPC: G06N3/063 ,  G06F17/16 ,  G06F15/78

Abstract: Systems and methods are provided to improve traditional chip processing. Using crossbar computations, the convolution layer can be flattened into vectors, and the vectors can be grouped into a matrix where each row or column is a flattened filter. Each submatrix of the input corresponding to a position of a convolution window is also flattened into a vector. The convolution is computed as the dot product of each input vector and the filter matrix. Using intra-crossbar computations, the unused space of the crossbars is used to store replicas of the filters matrices and the unused space in XIN is used to store more elements of the input. In inter-crossbar computations, the unused crossbars are used to store replicas of the filters matrices and the unused XINs are used to store more elements of the input. Then, the method performs multiple convolution iterations in a single step.