Method and apparatus for performing a neural network operation
A method for performing a neural network operation includes receiving weight and bias values of a deep neural network (DNN). An array of feature values, a bias value and a set of weight values for a single layer of the DNN are coupled to a neural network engine. Multiply-and-accumulate operations are performed on the single layer at one or more multiply and accumulate circuit (MAC) to obtain a sum corresponding to each neuron in the single layer. A layer output value corresponding to each neuron in the single layer is coupled to a corresponding input of the MAC. The coupling a bias value and a set of weight values, the performing multiply-and-accumulate operations and the coupling a layer output value are repeated to generate an output-layer-sum corresponding to each output-layer neuron and an activation function is performed on each output-layer-sum to generate DNN output values.
1 . A method for performing a neural network operation at a flash controller integrated circuit comprising:
receiving one or more configuration files at the flash controller integrated circuit, the received one or more configuration files indicating a number of input neurons, a number of output neurons, a number of layers of hidden neurons, a number of hidden neurons in each layer of hidden neurons, and a type of activation function module;
programming a neural network engine of the flash controller integrated circuit in accordance with the information indicated in the one or more configuration files, the neural network engine programmed to include one or more multiply and accumulate circuits (MAC$) and to include one or more activation function modules of the indicated type, an input of a respective one of the one or more activation function modules coupled to an output of a respective one of the one or more MACs, the neural network engine programmed to include MAC and activation function modules to define a classification neural network (CNN) and a regression deep neural network (DNN);
receiving weight and bias values for the CNN at the flash controller integrated circuit;
monitoring the operation of a flash memory device comprising a flash memory die coupled to the flash controller integrated circuit at a status module of flash controller integrated circuit to identify feature values relating to the flash memory device, the feature values including a number program and erase (P/E) values for the flash memory device;
coupling a first array of feature values to the neural network engine of the flash controller integrated circuit, the feature values in the first array of feature values relating to a predicted reliability state of the flash memory device and including at least some of the identified feature values corresponding to the flash memory device;
coupling a bias value for a single layer of the CNN and a set of weight values for the single layer of the CNN to the neural network engine of the flash controller integrated circuit;
performing multiply-and-accumulate operations on the single layer of the CNN at the one or more MACs to obtain a sum corresponding to each neuron in the single layer of the CNN;
coupling a layer output value corresponding to each neuron in the single layer of the CNN to a corresponding input of the one or more MACs, each layer output value comprising one of the sums in the single layer or a function of one of the sums in the single layer;
repeating the coupling the bias value and the set of weight values for the single layer of the CNN to the neural network engine, the performing multiply-and-accumulate operations on the single layer of the CNN and the coupling the layer output value until multiply-and-accumulate operations have been performed on all layers of the CNN to generate an output-layer-sum corresponding to each output-layer neuron;
performing an activation function on each output-layer-sum using a respective activation function module of the plurality of activation function modules to generate one or more CNN output values representing an output layer of the CNN, the one or more CNN output values indicating a predicted reliability state of the flash memory device;
coupling a second array of feature values and weight and bias values of the regression DNN to the neural network engine of the flash controller integrated circuit, the second array of feature values including at least some of the identified feature values corresponding to the flash memory device;
performing neural network operations of the regression DNN to generate regression neural network (RNN) output values predicting the shape of respective threshold-voltage-shift read-error (TVS-RE) curves;
coupling the one or more RNN output values to one or more functional circuits of the flash controller integrated circuit that are external to the neural network engine of the flash controller integrated circuit, the one or more functional circuits including a read module, the one or more functional circuits to identify respective threshold voltage shift offset (TVSO) values proximate a minimum value of respective TVS-RE curves; and
sending one or more threshold-voltage-shift read instructions and respective TVSO values from the read module to the flash memory device, for reading the flash memory device.
2 . The method of claim 1 comprising:
receiving a file indicating an activation function shape; and
receiving coordinates indicating a portion of the activation function shape at the neural network engine,
wherein the performing an activation function comprises performing a linear activation function corresponding to the received coordinates on each of the sums of the particular layer to generate corresponding layer output values that represent a nonlinear function of the respective output-layer-sums.
3 . The method of claim 1 further comprising:
receiving a file indicating an activation function shape; and
receiving coordinates indicating a portion of the activation function shape at the neural network engine,
wherein the performing an activation function on each output-layer-sum further comprises performing a linear interpolation operation corresponding to the received coordinates on each output-layer-sum to generate the one or more CNN output values.
4 . The method of claim 1 wherein the activation function module type indicates a type selected from the group consisting of a Rectified Linear Unit (ReLU) activation function module, a Leaky ReLU activation function module, a Saturating Linear (SatLin) activation function module, a symmetric SatLin activation function module and a linear interpolation activation function module.
5 . The method of claim 1 comprising:
wherein coupling the first array of feature values to the neural network engine further comprises storing the array of feature values in a first cache memory,
wherein coupling the bias value and the set of weight values for the single layer of the CNN to the neural network engine comprises storing the bias value and the set of weight values for the single layer of the CNN in a second cache memory, and
wherein coupling the layer output value to the corresponding input of the one or more MACs comprises storing layer output values corresponding to a hidden layer of the CNN in a cache memory and moving the layer output values corresponding to the hidden layer to the corresponding input of the one or more MACs, the moving comprising storing a copy of the layer output values, corresponding to the hidden layer of the CNN stored in the neural output cache, into the feature and bias input cache, by overwriting previously stored values in the feature and bias input cache.
6 . A flash controller integrated circuit comprising:
an input to receive weight and bias values of a classification neural network (CNN) and a regression deep neural network (DNN);
a control module coupled to the input;
function circuits for performing flash memory control operations, the function circuits include a write module for programming codewords into a flash memory device that includes a flash memory die external to the flash controller integrated circuit, include a read module to perform reads of the flash memory device using Threshold Voltage Shift Offset (TVSO) values include a decode module to decode the results of the reads and include a status module to monitor the operation of the flash memory device to identify feature values relating to the flash memory device, the feature values including a number of program and erase (P/E) values for the flash memory device;
a neural network engine coupled to the control module, the control module to send aa first array of feature values to the neural network engine and to send a bias value and set of weight values for a single layer of the CNN to the neural network engine, wherein at least some of the feature values in the first array of feature values are the feature values identified by the status module, the neural network engine comprising:
one or more multiply and accumulate circuits (MACs) coupled to the control module, the one or more MACs to receive the array of feature values, the bias value and the set of weight values for a single layer of the CNN, the one or more MACs to perform multiply-and-accumulate operations on the single layer of the CNN to obtain a sum corresponding to each neuron in the single layer of the CNN;
a neural network engine controller coupled to the one or more MACs, the neural network engine controller to couple a layer output value corresponding to each neuron in the single layer of the CNN to a corresponding input of the one or more MACs, where each layer output value includes of one of the sums or a function of the one of the sums; and
one or more activation function module coupled to the one or more MACs,
wherein the control module is to repeat the coupling the bias value and the set of weight values, the one or more MACs are to repeat the performing multiply-and-accumulate operations and the neural network engine controller is to repeat the coupling the layer output value to the corresponding input of the one or more MAC until multiply-and-accumulate operations have been performed on all layers of the CNN to generate an output-layer-sum corresponding to each output-layer neuron,
wherein the activation function module is to perform an activation function on each output-layer-sum to generate one or more CNN output values representing an output layer of the CNN, the one or more CNN output values indicating a predicted reliability state of a flash memory device,
wherein the control module is to couple a second array of feature values and weight and bias values oh the regression DNN to the neural network engine of the flash controller integrated circuit, the second array of feature values including at least some of the identified feature values corresponding to the flash memory device, the neural network engine to perform neural network operations of the regression DNN to generate regression neural network (RNN) output values predicting the shape of respective threshold-voltage-shift read-error (TYS-RE) curves,
wherein the TVSO values used to perform the reads are TVSO values proximate a minimum value of respective TVS-RE curves.
7 . The integrated circuit of claim 6 wherein the control module is coupled to the MAC through one or more cache memory.
8 . The integrated circuit of claim 6 wherein the neural network engine comprises:
a feature and bias input cache coupled to the control module and the one or more MACs;
a weight input cache coupled to the control module and to the one or more MACs; and
a neural output cache coupled to the outputs of the one or more activation function modules and coupled to the neural network engine controller, the neural output cache to store layer output values corresponding to each hidden layer of the CNN and the one or more CNN output values,
wherein the coupling of each sum to the corresponding input of the one or more MACs comprises storing a copy of the contents of the neural output cache in the feature and bias input cache.
9 . The integrated circuit of claim 6 wherein the activation function module comprises a linear interpolation module coupled to the one or more MAC's, the linear interpolation module to receive coordinates indicating a portion of an activation function shape and perform a linear interpolation operation corresponding to the received coordinates on each of the sums of the particular layer of the DNN to generate corresponding layer output values that represent a nonlinear function of the respective sums.
10 . The integrated circuit of claim 6 wherein the activation function module performs an activation function selected from the group consisting of a Rectified Linear Unit (ReLU) activation function, a Leaky ReLU activation function, a Saturating Linear (SatLin) activation function, a symmetric SatLin activation module and a linear interpolation activation function module.