Vertical mapping and computing for deep neural networks in non-volatile memory

An apparatus, comprising:

• a stack of a plurality of bonded die pairs connected along a plurality of vias configured to transfer data between the bonded die pairs, each of the bonded die pairs comprising:
• a memory die having one or more arrays of non-volatile memory cells, each array configured to store one or more weights of a layer of a neural network;
• and
• a control die having one or more control circuits configured to access data stored on the memory die of the bonded die pair,
• where the arrays of the memory dies are organized into columns along the vias and the control circuits are further configured to:
• receive an input for the neural network at first of bonded die pairs;
• and
• propagate the input through the layers of the neural network by propagating the input though a corresponding sequence of the arrays along the columns, performing a multiplication operation in each of the bonded die pairs using the weights stored in the corresponding array in the sequence.

A method, comprising:

• receiving a first input for a neural network;
• performing an inferencing operation for the input of the neural network and a plurality of layer of the neural network by:
• receiving the first input at a control die of a first bonded die pair of anon-volatile memory structure, the non-volatile memory structure comprising a plurality bonded die pairs, including the first bonded die pair, each of the bonded die pairs including a control die and a non-volatile memory die, each of the non-volatile memory dies storing weights for one or more layers of the neural network;
• propagating the first input through a sequence of the bonded die pairs, the bonded die pairs forming a three dimensional stack in which the bonded die pairs are connected by through silicon vias, the propagating including:
• at the first bonded die pair, generating an output for a layer of the neural network stored on the memory die of the first bonded die pair from the first input;
• and
• at each of one or more subsequent bonded die pairs in the sequence:
• receiving along the through silicon vias an output from the preceding bonded die pair in the sequence;
• and
• generating an output for a layer of the neural network stored on the memory die of the subsequent bonded die pair using the output from the preceding bonded die pair as input;
• and
• providing an output of the inferencing operation from the control die of one of the subsequent bonded die pairs.