Resistive processing unit cell having multiple weight update and read circuits for parallel processing of data using shared weight value

A device comprising:

• an array of resistive processing unit cells;
• a plurality of weight update control lines coupled to each resistive processing unit cell in the array;
• a plurality of weight read control lines coupled to each resistive processing unit cell in the array;
• and
• control circuitry configured to control the array of resistive processing unit cells, wherein the control circuitry is configured to (i) generate read control signals that are applied to the weight read control lines to read weights of the resistive processing unit cells, and (ii) generate weight update control signals that are applied to the weight update control lines to update weights of the resistive processing unit cells;
• wherein each resistive processing unit cell in the array of resistive processing unit cells comprises:
• a plurality of circuit blocks, wherein each circuit block comprises a weight update circuit coupled to dedicated update control lines, and a weight read circuit coupled to dedicated read control lines;
• and
• a weight storage device configured to store a weight voltage which represents a weight value of the resistive processing unit cell;
• wherein the resistive processing unit cell is configured to enable shared access to the weight storage device by the plurality of circuit blocks of the resistive processing unit cell;
• wherein the plurality of circuit blocks are configured to operate in parallel to perform separate weight read operations in which each read circuit of the resistive processing unit cell is configured to generate a read current based on the stored weight voltage of the resistive processing unit cell, and output the read current on the dedicated read control lines of the read circuit;
• wherein the plurality of circuit blocks are configured to operate in parallel to perform separate weight update operations in which each weight update circuit of the resistive processing unit cell is configured to receive respective update control signals on the dedicated update control lines, generate an update current based on the respective update control signals, and apply the update current to the weight storage device to adjust the stored weight voltage based on the update current;
• and
• wherein at least two of the plurality of circuit blocks of the resistive processing unit cell are configured to concurrently operate on a different data set using the same weight value of the resistive processing unit cell.

A method comprising:

• mapping kernel values of a convolutional kernel to an array of resistive processing unit cells which represents a convolutional layer of a neural network, wherein each resistive processing unit cell comprises (i) a weight storage device configured to store a weight voltage which represents a kernel value of the convolutional kernel, and (ii) a plurality of circuit blocks, wherein each circuit block comprises a weight update circuit coupled to dedicated update control lines, and a weight read circuit coupled to dedicated read control lines;
• inputting a plurality of different input volumes to the array of resistive processing unit cells to perform a training process, wherein a number of the different input volumes equals a number of the circuit blocks of each resistive processing unit cell;
• and
• performing a training process to train the kernel values of the convolutional kernel, wherein the training process comprises concurrently applying the convolutional kernel to each of the different input volumes in parallel by (i) performing separate weight read operations in parallel in each resistive processing unit cell wherein each read circuit is configured to generate a read current based on the stored weight voltage of the resistive processing unit cell, and output the read current on the dedicated read control lines of the read circuit, and (ii) performing separate weight update operations in parallel in each resistive processing unit cell wherein each weight update circuit is configured to receive respective update control signals on the dedicated update control lines, generate an update current based on the respective update control signals, and apply the update current to the weight storage device to adjust the stored weight voltage based on the update current;
• wherein each corresponding circuit block in the resistive processing unit cells processes a corresponding one of the input volumes;
• and
• wherein at least two of the plurality of circuit blocks of at least a given one of the resistive processing unit cells are configured to concurrently operate on a different one of the input volumes set using the same weight value of the given resistive processing unit cell.