公开(公告)号：US11200948B1

公开(公告)日：2021-12-14

申请号：US17005219

申请日：2020-08-27

Applicant: HEWLETT PACKARD ENTERPRISE DEVELOPMENT LP

Inventor： Joao Claudio Ambrosi ,  Arthur Carvalho Walraven Da Cunha ,  Jefferson Rodrigo Alves Cavalcante

IPC: G11C13/00 ,  G06N3/02 ,  H01L45/00 ,  G11C5/02 ,  H01L27/24

Abstract: Systems are provided for implementing a hybrid resistor-memristor crossbar array, which allows for flexible conductance to be used in implementing the weight matrix of a neural network. The hybrid resistor-memristor crossbar array may include resistor crossbars, each resistor having a static conductance value. The hybrid resistor-memristor crossbar array may also have a memristor coupled to an output line associated with the resistor crossbar array, wherein the memristor has a variable conductance value, and further wherein the static conductance values and the variable conductance value are set to calculate a matrix-vector multiplication associated with a weight matrix of a neural network. An expected range of coefficients for a weight matrix of a neural network can be given by the Discrete Transform Cosine (DCT). Accordingly, the static conductance values of the resistors in the resistors crossbar array are set to values equal to known coefficients of the DCT.

公开(公告)号：US12254395B2

公开(公告)日：2025-03-18

申请号：US17027628

申请日：2020-09-21

Applicant: HEWLETT PACKARD ENTERPRISE DEVELOPMENT LP

Inventor： Glaucimar Da Silva 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: G06F15/78 ,  G06F17/16 ,  G06N3/063

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.