Capacitive Processing Unit
Capacitive processing units integrated within memory devices address inefficiencies in ANNs by enabling in-memory computation, reducing latency and enhancing neural network performance through localized processing and storage.
Patent Information
- Application Number
- JP2022533316
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-12-09
- Filing Date
- 2020-11-20
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2040-11-20
AI Technical Summary
Existing artificial neural networks (ANNs) face inefficiencies in processing and storage capabilities, particularly in applications requiring complex computations and large data sets, due to the need for separate processing and memory components, leading to increased latency and resource utilization.
The integration of capacitive processing units (CPUs) within memory devices, utilizing metal oxide semiconductor (MOS) devices to facilitate both storage and computation, enabling in-memory processing through capacitive elements that store electrical energy and perform calculations locally, thereby reducing latency and enhancing efficiency.
The capacitive processing units enable efficient in-memory computation, reducing latency and improving the performance of neural networks by allowing weight updates and data processing directly within the memory cells, enhancing the speed and efficiency of neural network operations.
Smart Images

Figure 0007595669000001 
Figure 0007595669000002 
Figure 0007595669000003
Abstract
Description
[Technical field]
[0001] The present invention relates generally to memory devices for neuromorphic hardware and systems that can be used to implement artificial neural networks. [Background technology]
[0002] An artificial neural network (ANN) is a computational model inspired by the biological neural networks of the human or animal brain. ANNs learn tasks progressively and autonomously using examples. Applications of ANNs may include, but are not limited to, speech recognition, text processing, and image classification. ANNs contain a set of connected nodes (or neurons) and edges (connections between the nodes). Along the edges between the nodes, signals can be transmitted. Nodes that receive signals can process the received signals and communicate the processed signals (outputs) to the connected nodes. Weights can be associated with the connections and nodes, and the weights can be adjusted as learning progresses. Summary of the Invention
[0003] In some examples, a memory device structure is generally described. The structure can include an array of memory cells. The memory cell can include at least one metal oxide semiconductor (MOS) element, where a source terminal of the at least one MOS element is connected to a drain terminal of the MOS element. With the source terminal connected to the drain terminal, the at least one MOS element can exhibit capacitive behavior for storing electrical energy. A first transistor can be connected to the at least one MOS element, where activation of the first transistor can facilitate a write operation to the memory cell. A second transistor can be connected to the at least one MOS element, where activation of the second transistor can facilitate a read operation from the memory cell.
[0004] In some examples, a system including a memory, a processor, and a structure is generally described. The processor can be configured to communicate with the memory. The structure can be configured to communicate with the memory and the processor. The structure can include an array of memory cells. The memory cell can include at least one metal oxide semiconductor (MOS) element, where a source terminal of the at least one MOS element is connected to a drain terminal of the MOS element. With the source terminal connected to the drain terminal, the at least one MOS element can exhibit a capacitive behavior for storing electrical energy. A first transistor can be connected to the at least one MOS element, where activation of the first transistor can facilitate a write operation to the memory cell. A second transistor can be connected to the at least one MOS element, where activation of the second transistor can facilitate a read operation from the memory cell. The processor can be configured to generate one or more control signals for controlling write and read operations of the array of memory cells.
[0005] In some examples, a memory device is generally described. The memory device may include a non-volatile memory configured to store data and a structure coupled to the non-volatile memory. The structure may include an array of memory cells. The memory cell may include at least one metal oxide semiconductor (MOS) element, where a source terminal of the at least one MOS element is connected to a drain terminal of the MOS element. With the source terminal connected to the drain terminal, the at least one MOS element may exhibit a capacitive behavior for storing electrical energy. A first transistor may be connected to the at least one MOS element, where activation of the first transistor facilitates a write operation to the memory cell. A second transistor may be connected to the at least one MOS element, where activation of the second transistor facilitates a read operation from the memory cell. The write operation may change a capacitance of the at least one MOS element, where the changed capacitance may update data stored in the non-volatile memory.
[0006] Further features as well as the structure and operation of various embodiments are described in detail below with reference to the accompanying drawings, where like reference numbers indicate identical or functionally similar elements. [Brief description of the drawings]
[0007] [Figure 1] FIG. 1 illustrates an example of a system including one or more capacitive processing units in one embodiment. [Diagram 2] FIG. 2 illustrates additional details of a structure including a capacitive processing unit in one embodiment. [Diagram 3] FIG. 2 illustrates details of a capacitive processing unit in one embodiment. [Figure 4] FIG. 2 illustrates details of a circuit that may be implemented by the capacitive processing unit in one embodiment. [Diagram 5] FIG. 2 illustrates an exemplary implementation of a capacitive processing unit in one embodiment. [Figure 6] 1 is a table illustrating an example implementation of a capacitive processing unit in one embodiment. [Figure 7] FIG. 1 illustrates an example of a memory device including an integrated capacitive processing unit in one embodiment. [Figure 8] FIG. 2 is a schematic diagram of an example computer or processing system in which a capacitive processing unit can be implemented in accordance with an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0008] The following description of exemplary embodiments of various types of neural networks is presented by way of example. It should be noted that the embodiments of neural networks described herein are not limited to the examples described herein.
[0009] A deep neural network (DNN) is an ANN with multiple layers between the input and output layers. A DNN can learn relationships between input and output data and can use the learned relationships to transform future inputs into outputs. A DNN moves through layers that calculate the probability of each output, and each functional operation is considered a layer. In an example where a DNN is applied to image recognition, an example of a DNN's goal is to learn image features and classify new input images by comparing them to the learned features, without the need for human input. DNNs can also be used to model relatively complex nonlinear relationships and are typically feedforward networks.
[0010] In a feedforward neural network (such as a multi-layer perceptron), multiple nodes can be arranged in layers. Nodes in adjacent layers have connections between them, and the connections are associated with weights. A feedforward neural network can include three types of nodes arranged in each layer: input nodes, hidden nodes, and output nodes. Input nodes are nodes in the input layer and are responsible for the task of providing information to the network from the outside. In some examples, the input nodes may pass information to hidden nodes in the hidden layer and are not configured to perform calculations. Hidden nodes are arranged in one or more hidden layers and are hidden from the outside of the network. The hidden nodes are configured to perform calculations and forward information from the input layer to the output nodes. The output nodes in the output layer can perform calculations and forward information from the network to the outside. In one example, information propagates in one direction (forward direction) through hidden nodes in a feedforward network. The forward direction is from the input nodes to the output nodes. A feedforward network may be free of cycles or loops, which is different from a recurrent neural network (RNN) with circular connections between nodes. A recurrent neural network (RNN) is a network in which data can flow in any direction. RNNs can be used in applications such as language modeling.
[0011] A convolutional neural network (CNN) is a type of deep neural network. CNNs, for example, use a variety of multi-layer perceptrons designed to require minimal preprocessing. CNNs are also called shift-invariant or position-invariant artificial neural networks (SIANNs) based on their weight-sharing structure and translation-invariant properties. In one example, CNNs can learn filters in various applications, such as image processing applications, and the learning is independent of prior knowledge and human input. Applications of CNNs can include, but are not limited to, image and video recognition, computer vision, recommender systems, image classification, medical image analysis, acoustic modeling for automatic speech recognition (ASR), and natural language processing.
[0012] FIG. 1 illustrates an example of a system including one or more capacitive processing units in one embodiment. In the example illustrated in FIG. 1, the system 100 can include a device 101 configured to implement a neural network, such as a DNN 102. The device 101 can be a computing device configured to receive an input 130, which can include input data, such as image data including, but not limited to, pixel values. The device 101 can implement a neural network (e.g., the DNN 102) to generate an output 132, which can be, for example, a classification of the input 130. The device 102 can include a processor 110, a memory device (or memory) 120 configured to communicate with the processor 110, and a structure 150 configured to communicate with the processor 110 and the memory 120. In an exemplary embodiment, the structure 150 and the memory 120 can be integrated within the same memory device. In another exemplary embodiment, the structure 150 can be a memory accelerator.
[0013] In one example, the processor 110 can receive the input 130 and transmit the input 130 to the structure 150. The structure 150 can include multiple capacitive processing units or devices 160 ("cells 160"). In the example shown in FIG. 1, the structure 150 can include N×M capacitive processing units 160 (N rows with rows i=1, ..., N, and M columns with columns j=1, ..., M). Some of the capacitive processing units 160 in the structure 150 can be activated based on a desired implementation of the DNN 102, attributes of the DNN 102 (e.g., size, number of layers), attributes of the inputs 130 and outputs 132 (e.g., size), or other factors, or a combination thereof. For example, the structure 150 can include X memory cells, and N×M cells can be activated to train the DNN 102. The structure 150 can further include multiple conductive wires connecting the capacitive processing units 160. For example, the structure 150 may include N write word line (WWL) wires, N read word line (RWL) wires, and M true bit line (BLT) wires. The processor 110 may be configured to generate control signals to control the operation of the structure 150. For example, the processor 110 may write a logic "1" signal to all the RWL wires to perform a read operation, and may accumulate electrical energy that may be stored in the capacitive processing unit 160, such that the accumulated electrical energy may be read out as a digital value.
[0014] The structure 150 may further include a circuit 151, a circuit 152, and a plurality of circuits 170. The circuit 151 and the circuit 152 may be configured to receive control signals from the processor 110 and facilitate writing the control signals to the WWL and RWL wires. The circuit 151 may include a decoder (row decoder) configured to facilitate selection or activation of one or more rows (rows 1-N) of the cells 160. The circuit 152 may include a decoder (column decoder) configured to facilitate selection or activation of one or more columns (columns 1-M) of the cells 160. For example, the processor 110 may send a control signal having a logic "1" in bit positions 1 and 2 corresponding to the WWL wires of rows i=1 and i=2 to the circuit 151 and send a control signal having a logic "1" in bit position 2 corresponding to the BLT wire of column j=2 to the circuit 152 to facilitate a write operation to the cells (1,2) and (2,2). In some examples, the circuit 151 can include additional circuitry such as a digital-to-analog converter (DAC) that can be used to convert digital inputs into analog signals (or currents) that can be input to the cells 160. In some examples, the circuit 152 can include additional circuitry such as a sense amplifier that can be used to facilitate reading out accumulated data from the electrical energy stored in the cells 160. The control signals generated by the processor 110 can be used to control the structure 150 to perform operations such as matrix operations (e.g., addition, multiplication) and vector matrix operations (e.g., multiplication-accumulation, etc.) to train and implement the DNN 102. These operations performed by the structure 150 can facilitate training of the DNN 102, including updating the weights of the DNN 102 and implementing the DNN 102 to classify input data received by the device 101. The multiple circuits 170 can be analog circuits and can be configured to store the electrical energy output and accumulated from the cells 160. In an exemplary embodiment, each column of the cells 160 is connected to one circuit 170. In the example shown in FIG. 1, there are M circuits 170 corresponding to M columns of cells 160.
[0015] The cell 160 described in the embodiment according to the invention includes both processing and storage capabilities. The cell 160 may include one or more metal oxide semiconductor (MOS) elements, where the one or more MOS elements may be configured as one or more capacitive elements operable to store electrical energy. Furthermore, the capacitive processing device may be configured to perform calculations by switching the one or more MOS elements at different times using one or more clock signals. Thus, the cell 160 may provide in-memory computation in the implementation of a neural network. The in-memory computation may obtain weights from electrical energy stored in the capacitive elements of the capacitive processing device, and may also use the capacitive processing device to perform the calculations locally (in-memory computation), thereby increasing efficiency in the implementation of a neural network, for example by reducing latency.
[0016] Figure 2 is a diagram illustrating additional details of a structure including a capacitive processing unit in one embodiment. Figure 2 may include components with the same names as components in Figure 1, but for clarity, these will not be described again. The description of Figure 2 may refer to at least some of the components in Figure 1.
[0017] In the example shown in FIG. 2, the structure 150 may include additional components and wires to facilitate different operations of the cell 160. Focusing on one cell 160, the cell 160 may be connected to multiple components, such as pass gates or transmission gates, that are controllable using different signals provided by the processor 110 to cause the cell 160 to perform different operations. The multiple control signals that may be received by the structure 150 are, for example, 1) WRITE, 2) WEIGHT, 3) a write word line (WWL) signal, 4) a read word line (RWL) signal, 5) a precharge (PCH) signal, 6) a bias voltage labeled DATA, and 7) a bias voltage labeled DATA_T. It is noted that additional input signals may be received by the structure 150. The PCH signal may be one bit and may be a precharge signal to initialize an output line, such as a BLT wire. The WRITE signal may be one bit, and when WRITE="1", DATA_T may be passed to cell 160 through pass gate 204 as signal BIAS_T for biasing a first set of MOS elements in cell 160. Similarly, when WRITE="1", DATA may be passed to cell 160 through pass gate 206 as signal BIAS_B for biasing a second set of MOS elements in cell 160. DATA and DATA_T may each be M bits long corresponding to the number of columns of cells 160 in structure 150. The WWL and RWL signals may be N bits long corresponding to the number of rows of cells 160 in structure 150. The WWL and RWL signals may be generated by a local clock generator, which may be part of structure 150 or device 101.
[0018] The WEIGHT signal may be N bits long and may be a weight value being written to or updated in the memory 120. If the WWL signal is logic "1", the pass gate 202 may enable the WEIGHT signal to be sent to the cell 160. In an exemplary embodiment, the weight of the DNN 102 may be stored in the cell 160 and may be read by accumulating the electrical energy stored in the cell 160. The weight of the DNN 102 may be updated by the WEIGHT signal, such as by charging or discharging a capacitive element in the cell 160 selected by a write operation. Thus, the weight update may be performed locally in the cell 160. The updated amount of electrical energy between the cells 160 may be output via a WEIGHT_OUT pin that may be connected to an internal node labeled Bit Line Complement (BLC). The internal node BLC may be connected to a BLT wire, and the voltage difference between the BLT wire and the internal BLC node may represent the amount of electrical energy stored in the cell 160. The updated amount of electrical energy may be accumulated by the circuits 170, 152 and output to the memory 120 as a digital value.
[0019] FIG. 3 is a diagram illustrating details of a capacitive processing unit in one embodiment. Such a configuration may include a row and column arrangement of circuit components. FIG. 3 may include components with the same names as components in FIG. 1 and FIG. 2, but for clarity, these will not be described again. The description of FIG. 3 may refer to at least some of the components in FIG. 1 and FIG. 2.
[0020] The cell 160 may include multiple MOS elements, e.g., N-type MOS (NMOS) and P-type MOS (PMOS) transistors. In the example shown in FIG. 3, an NMOS transistor 302 may be controlled by a signal received from the WWL wire. Another NMOS transistor 304 may be controlled by a signal received from the RWL wire. Note that if the signal received from the WWL is logic "1", a current may flow from the BLT wire into the cell 160 to facilitate a write operation. Similarly, if the signal received from the RWL is logic "1", a current may flow from the cell 160 towards the BLC wire to facilitate a read operation.
[0021] The cell 160 may further include a PMOS transistor 310 and a PMOS transistor 312. The source terminal of the PMOS transistor 310 may be connected to the drain terminal of the PMOS transistor 310 or to the same voltage source. Similarly, the source terminal of the PMOS transistor 312 may be connected to the drain terminal of the PMOS transistor 312 or to the same voltage source. In the example shown in FIG. 3, the source terminals of the PMOS transistors 310, 312 are connected to their own drain terminals, their own gates, and the same voltage source. The PMOS transistors 310, 312 may be biased by the BIAS_T signal passed by the pass gate 204 in response to the WRITE signal being a logic "1". The BIAS_T signal may be the DATA_T signal.
[0022] The cell 160 may further include an NMOS transistor 314 and an NMOS transistor 316. The source terminal of the NMOS transistor 314 may be connected to the drain terminal of the NMOS transistor 314 or to the same voltage source. Similarly, the source terminal of the NMOS transistor 316 may be connected to the drain terminal of the NMOS transistor 316 or to the same voltage source. In the example shown in FIG. 3, the source terminals of the NMOS transistors 314, 316 are connected to their own drain terminals, their own gates, and the same voltage source. The NMOS transistors 314, 316 may be biased by the BIAS_B signal passed by the pass gate 206 in response to the WRITE signal being a logic "1". The BIAS_B signal may be a DATA signal.
[0023] When the source and drain terminals of a MOS are connected together, the MOS can exhibit different levels of capacitance in response to different voltages (gate voltages) being applied to the gate of the MOS. Thus, the MOS can behave similarly to a capacitor, which can be referred to as a MOS capacitor. Comparing capacitance to gate voltage, the capacitance change of a MOS capacitor can be linear in the gate voltage region 0 to 1 volt (V), or 0 to -1V. The capacitance of a MOS capacitor changes linearly with respect to the gate voltage in these regions. This linearity of the capacitance-voltage characteristic of a MOS capacitor can be utilized to use it as a capacitive element in a memory cell. By using a MOS capacitor as a capacitive element in a memory cell, the memory cell can be controlled to store electrical energy and to control the MOS element using various control signals for in-memory processing.
[0024] In an exemplary embodiment, the voltage difference 330 between the BLT and BLC wires can represent the amount of electrical energy stored in the cells 160. A column of cells 160 can output a respective voltage difference 330 at a respective output wire BLT or BLC. The voltage differences output from cells in the same column can be accumulated to generate an output for the column. Using an example in which the structure 150 has N×M cells 160, the structure 150 can output M accumulated voltage differences in parallel representing M values. The electrical energy transferred from the jth column of cells 160 to the jth BLT wire or the jth BLC wire can be differentially amplified and latched by the jth sense amplifier in the circuit 152, thereby outputting the accumulated electrical energy from the jth column of cells 160.
[0025] Figure 4 is a diagram illustrating details of a circuit that may be implemented by a capacitive processing unit in one embodiment. Figure 4 may include components with the same names as components in Figures 1-3, but for clarity, these will not be described again. The description of Figure 4 may refer to at least some of the components in Figures 1-3.
[0026] FIG. 4 shows the details of the circuit 170 in one embodiment. An analog_in (ana_in) pin 402 can be connected to the BLT wire of the corresponding column. For example, the pin 402 of the circuit 170_1 in column M=1 can be connected to the BLT wire of column M=1. The circuit 170 can receive a select signal at a pin 404, which can activate or deactivate the circuit 170. For example, when the SEL pin receives a logic "1" signal, the input impedance is high and the output impedance is low, so the circuit 170 acts as an analog buffer and allows current to flow through the circuit 170. When the SEL pin receives a logic "0" signal, the circuit 170 is turned off, both the input impedance and the output impedance are high, and no current flows through the circuit 170. The jth circuit 170 can collect the electrical energy output from the jth string of cells 160, and during a read operation, the jth circuit 170 can output the stored electrical energy accumulated from the jth string of cells 160 via the t_out pin 406. The t_out pin 406 can be connected to a read pin of the structure 150, and the stored energy from the string of cells 160 can be output via the read pin to one or more devices external to the structure 150.
[0027] FIG. 5 illustrates an exemplary implementation of a capacitive processing unit in one embodiment. FIG. 5 may include components with the same names as components in FIG. 1-4, but for clarity, these will not be described again. The description of FIG. 5 may refer to at least some of the components in FIG. 1-4.
[0028] The illustrative diagram of FIG. 5 can be used to represent some electrical characteristics of the cell 160. Comparing the diagram of FIG. 5 with FIG. 3, switch S1 can represent PMOS 312, switch S3 can represent PMOS 310, switch S2 can represent NMOS 316, and switch S4 can represent NMOS 314. The NMOS transistors 310, 312 and the PMOS transistors 314, 316 can be implemented as switches S1, S2, S3, S4 (e.g., these switches can be turned on or off by applying a clock signal to their respective gate terminals). The voltages V1, V3 can represent BIAS_T voltages, and the voltages V2, V4 can represent BIAS_B voltages. Capacitor C can represent a capacitance corresponding to a voltage difference 330 between the BLT and BLC wires. In one example, switches S1, S3 are the same (e.g., both NMOS) and switches S2, S4 are the same (e.g., both PMOS), so that the capacitance C is based on the first difference V1-V3 and the second difference V2-V4, allowing cell 160 to operate similarly to a differential in, differential out amplifier.
[0029] In another example, if the switches S1, S4 represent a first type of MOS and the switches S2, S3 represent a second type of MOS, an equivalent resistance can be formed between the voltage differences (V1-V4) and (V2-V3). S1-S4 can also be made from the same type of transistors, and for correct functioning, both S1 and S4 can be turned on in the first phase, and both S2 and S3 can be turned on in the second phase, but the first and second phases do not overlap. Alternatively, S1 and S2 can be opposite polarities of transistors operating in the same phase, to ensure that S1 and S4 are turned on at the same time, and S2 and S3 are turned on at the same time, and S1 and S2 can be essentially complementary. In another example, if switches S1, S4 represent a first type of MOS, switches S2, S3 represent a second type of MOS, and voltage V1=V3 and voltage V2=V4, then a resistor having a value of T / C can be formed between V1 (or V3) and V2 (or V4), where T indicates the period during which a clock signal is applied to the gate terminals of the transistors implemented as switches S1, S2, S3, S4.
[0030] Figure 6 is a table illustrating an exemplary implementation of a capacitive processing unit in one embodiment. Figure 6 may include components with the same names as components in Figures 1-5, but for clarity, these will not be described again. The description of Figure 6 may refer to at least some of the components in Figures 1-5.
[0031] In one example, table 600 illustrates an exemplary implementation of cell 160 described herein. In table 600, rows T1, T2, and T3 are three operations that write WEIGHT signals "00" to a first cell (CELL_0) and a second cell (CELL_1) and read the written weights as voltage quantities. At T1, the WWL wire (WWL_0) of CELL_0 receives logic "1" and the WWL wire (WWL_1) of CELL_1 does not receive logic "1", thus facilitating a write operation to CELL_0 but not facilitating a write operation to CELL_1. Note that at T1, RWL_0="1" and RWL_1="0", and no read operation is performed because all RWL wires must be logic "1" to perform a read operation. Thus, at T1, a "0" is written to CELL_0 (the MOS element of CELL_0 is not charged). At T2, the WWL wire (WWL_1) of CELL_1 receives a logic "1", facilitating a write operation to CELL_1. Note that at T2, RWL_0="0" and RWL_1="1", and no read operation is performed since all RWL wires must be at logic "1" to perform a read operation. Thus, at T2, a "0" is written to CELL_1. At T3, both RWL_0 and RWL_1 are at logic "1", and the electrical energy stored in cells CELL_0 and CELL_1 is read as a voltage (0.38mV), which in this case represents a logic "0".
[0032] Rows T4, T5, and T6 are three operations that write WEIGHT signals "01" to a first cell (CELL_0) and a second cell (CELL_1) and read the written weights as voltage quantities. At T4, the WWL wire (WWL_0) of CELL_0 receives logic "1" and the WWL wire (WWL_1) of CELL_1 does not receive logic "1", thus facilitating a write operation to CELL_0 but not facilitating a write operation to CELL_1. Note that at T4, RWL_0="1" and RWL_1="0", and no read operation is performed since all RWL wires must be at logic "1" to perform a read operation. Thus, at T4, a "0" is written to CELL_0. At T5, the WWL wire (WWL_1) of CELL_1 receives logic "1", thus facilitating a write operation to CELL_1. Note that at T5, RWL_0='0' and RWL_1='1', and no read operation is performed since all RWL wires must be at logic '1' to perform a read operation. Thus, at T5, a '1' is written to CELL_1. At T6, both RWL_0 and RWL_1 are at logic '1', and the electrical energy stored in cells CELL_0 and CELL_1 is read as a voltage (0.550mV), which in this case represents a logic '1'. Comparing T3 with T6, note that T6 outputs a higher voltage because CELL_1 first stored a '0' at T2 and was then updated (or charged) to store a '1' at T5.
[0033] Rows T7, T8, and T9 are three operations that write WEIGHT signals "11" to a first cell (CELL_0) and a second cell (CELL_1) and read the written weights as voltage quantities. At T7, the WWL wire (WWL_0) of CELL_0 receives logic "1" and the WWL wire (WWL_1) of CELL_1 does not receive logic "1", thus facilitating a write operation to CELL_0 but not facilitating a write operation to CELL_1. Note that at T7, RWL_0="1" and RWL_1="0", and a read operation is not performed because all RWL wires must be at logic "1" to perform a read operation. Thus, at T7, a "1" is written to CELL_0. At T8, the WWL wire (WWL_1) of CELL_1 receives logic "1", thus facilitating a write operation to CELL_1. Note that at T8, RWL_0='0' and RWL_1='1', and no read operation is performed since all RWL wires must be at logic '1' to perform a read operation. Thus, at T8, a '1' is written to CELL_1. At T9, both RWL_0 and RWL_1 are at logic '1', and the electrical energy stored in cells CELL_0 and CELL_1 is read as a voltage (0.799mV), which in this case represents a logic '1'. Comparing T6 to T9, note that T9 outputs a higher voltage because CELL_0 first stored a '0' at T5 and was then updated (or charged) to store a '1' at T8.
[0034] FIG. 7 illustrates an example of a memory device including an integrated capacitive processing unit according to one embodiment. FIG. 7 may include components with the same names as components in FIGs. 1-6, but for clarity, these will not be described again. The description of FIG. 7 may refer to at least some of the components in FIGs. 1-6.
[0035] Structure 150 coupled to memory 120 may be packaged as a memory device 700 configured to perform in-memory processing. In the example shown in FIG. 7, memory 120 may be a non-volatile memory (NVM), such as a magnetoresistive random access memory (MRAM), or other types of non-volatile memory (e.g., ReRAM, RRAM, FLASH, etc.). Based on structure 150 being coupled to NVM 120, as capacitive elements in cells 160 are charged or discharged, weights that may be stored in memory 120 may be updated accordingly. By coupling structure 150 to a non-volatile memory (e.g., memory 120), updates to weights stored in the non-volatile memory may be made locally in cells 160 without the need to have additional components to facilitate the transmission of data between structure 150 and memory 120. In some exemplary embodiments, cells 160 may be implemented as volatile memory cells. Memory device 700 may be implemented as a memory device that provides both volatile and non-volatile storage, and updates made to volatile storage elements may be transferred directly to the non-volatile storage elements.
[0036] In the example shown in FIG. 7, capacitance C can represent one or more of the capacitive elements in the cell 160 (e.g., PMOS 310, 312 and / or NMOS 314, 316). In some examples, capacitance C can be the capacitance of a capacitive element such as a transistor-based capacitor, a trench-based capacitor, a dielectric-based capacitor, a wire capacitor such as metal-insulator-metal (MIM), or other forms of capacitors, or combinations thereof. In some examples, the type of capacitor implemented in the cell 160 can be based on the desired implementation of the system 100. Transistor M1 can represent NMOS 304, transistor M2 can represent NMOS 302, clock signal Clk1 can represent an RWL signal, and clock signal Clk2 can represent a WWL signal. When Clk1 is high (logic "1") and Clk2 is low (logic "0"), transistor M1 can be activated, transistor M2 is not activated, and the MRAM 701 is not updated. Additionally, transistor M1 may be activated and transistor M2 may be deactivated to facilitate a read operation to read the stored energy from cell 160 as a voltage V. When Clk1 is low and Clk2 is high, transistor M1 may be deactivated and transistor M2 may be activated, and MRAM bit 701 may be updated according to capacitance C. Note that in an exemplary embodiment, MRAM bit 701 (or other non-volatile memory bit) may be disposed or connected between the BLT wire and the corresponding cell 160. For example, referring to the example of FIG. 3, MRAM bit 701 may be connected between the BLT wire and transistor 302.
[0037] FIG. 8 is a schematic diagram of an example computer or processing system capable of implementing a capacitive processing unit in an embodiment of the present invention. The computer system is merely one example of a suitable processing system and is not intended to suggest any limitation as to the scope of use or functionality of the embodiments of the methods described herein. The illustrated processing system is operable with numerous other general purpose or special purpose computing system environments or configurations. Examples of well-known computing systems, environments, or configurations, or combinations thereof, that may be suitable for use with the processing system illustrated in FIG. 8, may include, but are not limited to, personal computer systems, server computer systems, thin clients, thick clients, handheld or laptop devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputer systems, mainframe computer systems, supercomputers, and distributed cloud computing environments that include any of the above systems or devices.
[0038] A computer system may be described in the general context of computer system executable instructions, such as program modules, executed by a computer system. Generally, program modules may include routines, programs, objects, components, logic, data structures, etc. that perform particular tasks or implement particular abstract data types. The computer system may be practiced in a distributed cloud computing environment where tasks are performed by remote processing devices that are linked through a communications network. In a distributed cloud computing environment, program modules may be located in both local and remote computer system storage media, including memory storage devices.
[0039] Components of a computer system may include, but are not limited to, one or more processors or processing units 12, a system memory 16, and a bus 14 that couples various system components including the system memory 16 to the processor 12. The processor 12 may include modules 30 (e.g., machine learning module 30) that perform methods described herein. The modules 30 may be programmed into integrated circuits of the processor 12 or may be loaded from the memory 16, storage device 18, or network 24, or a combination thereof.
[0040] Bus 14 may represent any one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, and a processor or local bus using any of a variety of bus architectures, including, by way of example and not limitation, an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MCA) bus, an Enhanced ISA (EISA) bus, a Video Electronics Standards Association (VESA) local bus, and a Peripheral Component Interconnect (PCI) bus.
[0041] The computer system may include a variety of computer system readable media, which may be any available media that can be accessed by the computer system and may include both volatile and nonvolatile media, removable and non-removable media.
[0042] The system memory 16 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The computer system may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, a storage system 18 may be provided for reading from and writing to a non-removable, non-volatile magnetic medium (e.g., a "hard drive"). Although not shown, a magnetic disk drive may be provided for reading from and writing to a removable, non-volatile magnetic disk (e.g., a "floppy disk"), and an optical disk drive may be provided for reading from and writing to a removable, non-volatile optical disk, such as a CD-ROM, DVD-ROM, or other optical medium. In such instances, each may be connected to the bus 14 by one or more data media interfaces. In some examples, the system memory 16 may include structures including one or more capacitive processing units as described herein.
[0043] The computer system may also communicate with one or more external devices 26, such as a keyboard, a pointing device, a display 28, one or more devices that allow a user to interact with the computer system, or any device that allows the computer system to communicate with one or more other computing devices (e.g., network cards, modems, etc.), or a combination thereof. Such communication may occur through an input / output (I / O) interface 20.
[0044] Additionally, the computer system may communicate with one or more networks 24, such as a local area network (LAN), a general purpose wide area network (WAN), and / or a public network (e.g., the Internet) via a network adapter 22. As shown, the network adapter 22 communicates with other components of the computer system via a bus 14. Although not shown, it should be understood that other hardware and / or software components may be used in conjunction with the computer system. Examples include, but are not limited to, microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archival storage systems.
[0045] The present invention may be a system, method, or computer program product, or combination thereof, at any possible level of technical detail of integration. The computer program product may include one or more computer-readable storage media having computer-readable program instructions for causing a processor to perform aspects of the present invention.
[0046] A computer readable storage medium may be a tangible device capable of holding and storing instructions for use by an instruction execution device. A computer readable storage medium may be, for example, but not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. A non-exhaustive list of more specific examples of computer readable storage media includes the following: portable computer diskettes, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), static random access memories (SRAM), portable compact disk read-only memories (CD-ROM), digital versatile disks (DVD), memory sticks, floppy disks, mechanically encoded devices such as punch cards or ridge structures in grooves in which instructions are recorded, and any suitable combination thereof. As used herein, a computer-readable storage medium should not be construed as a transitory signal per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission medium (e.g., light pulses passing through a fiber optic cable), or electrical signals transmitted over electrical wires.
[0047] The computer readable program instructions described herein may be downloaded from a computer readable storage medium into each computing / processing device or may be downloaded to an external computer or storage device over a network, such as the Internet, a local area network, a wide area network, or a wireless network, or a combination thereof. The network may include copper transmission cables, optical transmission fiber, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing / processing device.
[0048] The computer readable program instructions for carrying out the operations of the present invention may be source or object code written in any combination of one or more programming languages, including assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state setting data, configuration data for an integrated circuit, or procedural programming languages, such as object oriented programming languages, such as Smalltalk®, C++, and the like, and the "C" programming language or similar programming languages. The computer readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA) may execute computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry to carry out aspects of the invention.
[0049] Aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer readable program instructions.
[0050] These computer readable program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, executed by the processor of the computer or other programmable data processing apparatus, produce means for performing the functions / operations specified in one or more blocks of the flowcharts and / or block diagrams. These computer readable program instructions may be stored on a computer readable storage medium, such that the computer readable storage medium on which the instructions are stored comprises an article of manufacture containing instructions for performing aspects of the functions / operations specified in one or more blocks of the flowcharts and / or block diagrams, and may direct a computer, programmable data processing apparatus, or other device, or combination thereof, to function in a particular manner.
[0051] The computer readable program instructions may be loaded into a computer, other programmable data processing apparatus, or other device to cause the computer, other programmable apparatus, or other device to perform a series of operational steps to produce a computer-implemented process such that the instructions, which execute on the computer, other programmable apparatus, or other device, perform the functions / operations specified in one or more blocks of the flowcharts and / or block diagrams.
[0052] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowcharts or block diagrams may represent a module, segment, or portion of instructions, including one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions shown in the blocks may be performed out of the order shown. For example, two blocks shown in succession may in fact be executed substantially simultaneously, or the blocks may sometimes be executed in the reverse order, depending on the functionality involved. It should also be noted that each block in the block diagrams and / or flowchart illustrations, and combinations of blocks in the block diagrams and / or flowchart illustrations, may be implemented by a dedicated hardware-based system that executes the specified functions or operations, or executes a combination of dedicated hardware instructions and computer instructions.
[0053] The terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the invention. As used herein, the singular forms "a", "an" and "the" are intended to include the plural unless the context clearly indicates otherwise. It will be further understood that the terms "comprise" and / or "comprising" as used herein specify the presence of stated features, integers, steps, operations, elements, or components, or combinations thereof, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or groups thereof, or combinations thereof.
[0054] Corresponding structures, materials, acts, and equivalents of all means or step-plus-function elements within the scope of the following claims are intended to include any structure, material, or act for performing a function in combination with other claimed elements, when present, as specifically claimed. The description of the present invention has been presented for purposes of illustration and description, but is not intended to be exhaustive or to limit the invention to the disclosed form. Many modifications and variations will become apparent to those skilled in the art without departing from the scope of the invention. The embodiments have been selected and described in order to best explain the principles and practical applications of the invention, and to enable those skilled in the art to understand the invention in various embodiments including various modifications suitable for the particular use contemplated.
Claims
1. A structure including an array of memory cells, the memory cells comprising: at least one metal oxide semiconductor (MOS) device, a source terminal of the at least one MOS device being connected to a drain terminal of the at least one MOS device, the source terminal being connected to the drain terminal such that the at least one MOS device exhibits a capacitive behavior for storing electrical energy; a first transistor coupled to the at least one MOS device, activation of the first transistor facilitating a write operation to the memory cell; and a second transistor coupled to the at least one MOS device, activation of the second transistor facilitating a read operation from the memory cell; and Including, the at least one MOS device includes a first p-channel metal-oxide-semiconductor (PMOS) transistor, a second PMOS transistor, a first n-channel metal-oxide-semiconductor (NMOS) transistor, and a second NMOS transistor; the first PMOS transistor and the second PMOS transistor are connected in parallel, and the first NMOS transistor and the second NMOS transistor are connected in parallel; structure.
2. 2. The structure of claim 1, wherein the electrical energy stored in the columns of memory cells is accumulated to represent weights of an artificial neural network.
3. the first transistor is a first NMOS transistor; the second transistor is a second NMOS transistor; a drain terminal of the first NMOS transistor of the first transistor is connected to a true bit line (BLT) wire; a source terminal of the first NMOS transistor of the first transistor is connected to the at least one MOS element; a drain terminal of the second NMOS transistor of the second transistor is connected to the at least one MOS element; 2. The structure of claim 1, wherein a source terminal of the second NMOS transistor of the second transistor is connected to a bitline complementary (BLC) wire.
4. A gate of the first NMOS transistor of the first transistor is controlled by a first clock signal; a gate of the second NMOS transistor of the second transistor is controlled by a second clock signal; an input from the BLT wire is sent to the at least one MOS element in response to the first clock signal activating the first NMOS transistor of the first transistor; 4. The structure of claim 3, wherein in response to the second clock signal activating the second NMOS transistor of the second transistor, a voltage difference between the BLT wire and the BLC wire is output.
5. 10. The structure of claim 1, further comprising at least one circuit configured to bias a capacitance of the at least one MOS device.
6. 2. The structure of claim 1, wherein said array of memory cells is integrated with a non-volatile memory.
7. 10. The structure of claim 1, further comprising a plurality of analog circuits configured to accumulate and store the electrical energy output from the corresponding columns of memory cells.
8. Memory, a processor configured to communicate with the memory; A structure according to any one of claims 1 to 7, arranged to communicate with said memory and said processor; Including, the system.
9. a non-volatile memory configured to store data; A structure according to any one of claims 1 to 7 coupled to said non-volatile memory; 2. A memory device comprising:
Citation Information
Patent Citations
Capacitive structure for weighted addition used in neural nets
JP1993507168A
Active matrix display device
JP1995333653A
Semiconductor device
JP2008152876A
Semiconductor device, and electronic equipment
JP2019033234A
Apparatus and methods for segmented variable capacitor arrays
US9362882B1
