Neuromorphic device
A neuromorphic device includes a plurality of cell tiles, each of the plurality of cell tiles including a cell array including a plurality of memory cells storing weights of a neural network, a row driver connected to the plurality of memory cells through a plurality of row lines, and cell analog-digital converters (ADCs) connected to the plurality of memory cells through a plurality of column lines, and a controller configured to select, form the plurality of cell tiles, a plurality of valid cell tiles storing the weights, execute a neural network-based arithmetic operation based on the plurality of valid cell tiles, and redundantly store weights of a first layer among a plurality of layers included in the neural network in a plurality of first valid cell tiles that are divided into a plurality of first tile groups.
1 . A neuromorphic device comprising:
a plurality of cell tiles, each of the plurality of cell tiles comprising:
a cell array comprising a plurality of memory cells storing weights of a neural network,
a row driver connected to the plurality of memory cells through a plurality of row lines, and
cell analog-to-digital converters (ADCs) connected to the plurality of memory cells through a plurality of column lines; and
a controller configured to:
select, form the plurality of cell tiles, a plurality of valid cell tiles storing the weights,
execute a neural network-based arithmetic operation based on the plurality of valid cell tiles,
redundantly store weights of a first layer among a plurality of layers included in the neural network in a plurality of first valid cell tiles that are divided into a plurality of first tile groups,
input a lower bit of input data input to the first layer to a first tile group of the plurality of first tile groups, and
input an upper bit of the input data to a second tile group of the plurality of first tile groups during the neural network-based arithmetic operation.
2 . The neuromorphic device of claim 1 , wherein a number of cell tiles included in the first tile group of the plurality of first tile groups is equal to a number of cell tiles included in the second tile group of the plurality of first tile groups.
3 . The neuromorphic device of claim 1 , wherein the neural network comprises a second layer connected to the first layer, and
wherein the controller is further configured to:
redundantly store weights of the second layer in a plurality of second valid cell tiles that are divided into a plurality of second tile groups,
input a lower bit of input data input to the second layer to a first tile group of the plurality of second tile groups, and
input an upper bit of the input data to a second tile group of the plurality of second tile groups.
4 . The neuromorphic device of claim 3 , wherein a number of the plurality of first valid cell tiles is different from a number of the plurality of second valid cell tiles.
5 . The neuromorphic device of claim 1 , wherein at least two of the plurality of memory cells included in at least one of the plurality of first valid cell tiles in each of the plurality of first tile groups are zero-padded.
6 . The neuromorphic device of claim 1 , further comprising:
a plurality of adder trees configured to add digital cell data output from at least two valid cell tiles corresponding to one of the plurality of layers.
7 . The neuromorphic device of claim 6 , further comprising:
a plurality of offset shifters connected between the plurality of cell tiles and the plurality of adder trees and configured to execute a multiplication operation on outputs of the cell ADCs based on a number of bit digits of the input data input to the plurality of cell tiles.
8 . The neuromorphic device of claim 7 , wherein at least two of the plurality of offset shifters connected to the second tile group of the plurality of first tile groups are configured execute the multiplication operation on the outputs of the cell ADCs.
9 . The neuromorphic device of claim 6 , further comprising:
a shifter configured to execute a multiplication operation on outputs of at least some of the plurality of adder trees; and
an accumulator configured to accumulate an output of the shifter.
10 . The neuromorphic device of claim 1 , further comprising:
a layer buffer configured to store matching information between the plurality of layers included in the neural network and the plurality of valid cell tiles.
11 . The neuromorphic device of claim 10 , wherein the controller is further configured to:
select the plurality of valid cell tiles based on the matching information.
12 . The neuromorphic device of claim 10 , wherein weights of at least two of the plurality of layers are digitally converted with different precisions and are stored in at least two of the plurality of valid cell tiles, and
wherein the layer buffer is further configured to store precision information applied to each of the plurality of layers.
13 . The neuromorphic device of claim 10 , wherein the layer buffer is further configured to store, as the matching information, position information of each of the plurality of first tile groups, and offset information based on a number of bit digits of the input data input to each of the plurality of first tile groups.
14 . The neuromorphic device of claim 1 , wherein a number of the plurality of first tile groups is equal to or less than a number of bit digits of the input data.
15 . A neuromorphic device comprising:
a plurality of cell tiles, each of the plurality of cell tiles comprising:
a cell array comprising a plurality of memory cells,
a row driver connected to the plurality of memory cells through a plurality of row lines, and
cell analog-to-digital converters ADCs connected to the plurality of memory cells through a plurality of column lines and configured to convert cell currents read through the plurality of column lines into digital cell data;
a controller configured to:
define a plurality of first tile groups respectively comprising at least two of the plurality of cell tiles, and
store, in each of the plurality of first tile groups, weights of a first layer included in a trained neural network based on the weights of the first layer being received;
a layer buffer configured to receive, from the controller, and store matching information of the plurality of first tile groups, precision information of input data input to the first layer, and precision information of the weights of the first layer; and
a plurality of offset shifters configured to execute a multiplication operation on the digital cell data output from each of the plurality of cell tiles.
16 . The neuromorphic device of claim 15 , wherein the input data input to the first layer is N-bit data, N being a natural number of 2 or more, and
wherein the controller is further configured to:
input at least two lower bits of the N-bit data to cell tiles included in a first tile group of the plurality of first tile groups, and
input remaining upper bits of the N-bit data to cell tiles included in a second tile group of the plurality of first tile groups.
17 . The neuromorphic device of claim 15 , wherein the input data input to the first layer is N-bit data, N being a natural number of 3 or more,
wherein a number of the plurality of first tile groups is N, and
wherein the controller is further configured to:
divide the input data by 1 bit, and
input the input data to the plurality of first tile groups by 1 bit.
18 . A neuromorphic device comprising:
a plurality of cell tiles, each of the plurality of cell tiles comprising:
a cell array comprising a plurality of memory cells storing weights of a neural network,
a row driver connected to the plurality of memory cells through a plurality of row lines, and
cell analog-to-digital converters ADCs connected to the plurality of memory cells through a plurality of column lines and configured to convert cell currents read through the plurality of column lines into digital cell data;
a layer buffer configured to store matching information between a plurality of layers included in the neural network and a plurality of valid cell tiles of the plurality of cell tiles storing the weights; and
a controller configured to:
select the plurality of valid cell tiles among the plurality of cell tiles based on the matching information, and
execute a neural network-based arithmetic operation based on the plurality of valid cell tiles.
19 . The neuromorphic device of claim 18 , wherein the neural network comprises a first layer and a second layer sequentially arranged between an input layer and an output layer,
wherein the plurality of valid cell tiles comprises a plurality of first valid cell tiles storing weights of the first layer, and a plurality of second valid cell tiles storing weights of the second layer, and
wherein the layer buffer is further configured to store, as the matching information, position information of the plurality of first valid cell tiles and position information of the plurality of second valid cell tiles.
20 . The neuromorphic device of claim 19 , wherein a number of the plurality of first valid cell tiles is different from a number of the plurality of second valid cell tiles.