Charge pump-based current-mode neurons for machine learning

By implementing efficient storage and computation of filter weights in an in-memory computing architecture using charge-pump-based current-mode neurons, the data flow bottleneck in the von Neumann architecture is solved, improving the processing speed and computational density of machine learning applications.

CN115699184BActive Publication Date: 2026-11-10QUALCOMM INC
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202180041872.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2021-06-21
Filing Date
2021-06-22
Publication Date
2026-11-10
Estimated Expiration
2041-06-22

AI Technical Summary

Technical Problem

In machine learning applications, the data flow to and from memory in the von Neumann architecture becomes a bottleneck for processing speed, limiting computational performance.

Method used

We employ a charge pump-based current-mode neuron, which distributes data processing hardware in an in-memory computing architecture and utilizes the series coupling of charge pump capacitors and filter weight capacitors to achieve efficient storage and computation of filter weights. We combine a voltage digital-to-analog converter and a switching matrix for neuron training and computation.

Benefits of technology

It increases the density and efficiency of in-memory computation, reduces reliance on data streams, and improves the processing speed and computational density of machine learning applications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115699184B_ABST
    Figure CN115699184B_ABST
Patent Text Reader

Abstract

An in-memory computing array is provided, wherein each neuron includes a capacitor and an output transistor. During the evaluation phase, filter weight voltages and the binary state of the input bits control whether the output transistor is turned on or off, thus affecting the voltage on the read bit line connected to the output transistor.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Cross-references to related applications

[0002] This application claims priority to U.S. Nonprovisional Patent Application No. 17 / 353,564, filed June 21, 2021, which in turn claims the benefit of U.S. Provisional Application No. 63 / 042,537, filed June 22, 2020, the contents of which are incorporated herein by reference in their entirety. Technical Field

[0003] This application relates to in-memory computation, and more specifically to charge-pump-based current-mode in-memory computation bit cells (neurons) for machine learning applications. Background Technology

[0004] Computer processing of data typically uses the von Neumann architecture, where data is retrieved from memory for processing in arithmetic and logic units. In compute-intensive applications such as machine learning, the flow of data to and from memory becomes a bottleneck for processing speed. To address this data movement bottleneck, in-memory computing architectures have been developed, where data processing hardware is distributed across bit units. Summary of the Invention

[0005] According to one aspect of this disclosure, a memory for machine learning is provided, comprising: a bias circuit, a bit line, and a neuron, the neuron including: an output transistor having a drain connected to the bit line; a filter weight capacitor; a digital-to-analog converter; a first switch coupled between the bias circuit and the filter weight capacitor; a second switch coupled between the digital-to-analog converter and the filter weight capacitor; and a third switch coupled between the filter weight capacitor and the gate of the output transistor.

[0006] According to another aspect of this disclosure, a memory for machine learning is provided, comprising: a read bit line; an output transistor having a drain connected to the read bit line; a charge pump capacitor; a filter weight capacitor; a voltage source; and a set of switches having a first switching configuration in which the filter weight capacitor is charged by the voltage source, and a second switching configuration in which the filter weight capacitor is coupled to the charge pump capacitor to provide a boost voltage to the gate of the output transistor.

[0007] According to another aspect of this disclosure, a memory for machine learning is provided, comprising: a plurality of neurons, each neuron including: a bit line, a bit line switch, and an output transistor having a terminal connected to the bit line via the bit line switch; and a bias circuit including: a charge source and a switching matrix, the switching matrix being configured to couple the charge source to a selected neuron of the neurons to charge the gate capacitance in the output transistor of the selected neuron of the neurons with a boost voltage.

[0008] Finally, according to another aspect of this disclosure, a method for operating a neuron is provided, comprising: driving current through a diode-connected transistor to charge a charge pump capacitor to a charge pump voltage; adjusting a voltage-to-analog converter (VDAC) in response to a training dataset to charge a filter weight capacitor with the VDAC voltage; series coupling the charge pump capacitor and the filter weight capacitor in response to a first binary value of an input bit to charge the gate of an output transistor to the sum of the VDAC voltage and the charge pump voltage; and conducting a mirror version of the current through the output transistor in response to the charging of the gate of the output transistor.

[0009] These and other advantageous features can be better understood through the following detailed description. Attached Figure Description

[0010] Figure 1A The diagram illustrates a charge pump-based current-mode neuron and a bias circuit according to aspects of this disclosure.

[0011] Figure 1B The illustration shows the charging stage during the training phase and the refresh stage during the computation phase, according to aspects of this disclosure. Figure 1A The switching configuration of neurons.

[0012] Figure 1C The illustration depicts an evaluation stage during the training or computation phase in response to a first binary value of an input bit, according to aspects of this disclosure. Figure 1A The switching configuration of neurons.

[0013] Figure 1D The illustration depicts an evaluation stage during the training or computation phase in response to a second binary value of an input bit, according to aspects of this disclosure. Figure 1A The switching configuration of neurons.

[0014] Figure 2 The illustration depicts aspects of this disclosure, during the training phase or during the computation phase. Figure 1A Some operational waveforms of neurons.

[0015] Figure 3An array of charge pump-based current-mode neurons biased by a common bias circuit is illustrated according to aspects of this disclosure.

[0016] Figure 4 The illustration shows multiple charge pump-based current-mode neurons arranged to form a filter according to aspects of this disclosure.

[0017] Figure 5 The illustration depicts a charge pump-based current-mode neuron according to aspects of this disclosure, including a variable capacitor for storing filter weight voltages.

[0018] Figure 6 The illustrations depict aspects according to this disclosure. Figure 5 Modifications to the neurons to include voltage-to-digital-to-analog converters.

[0019] Figure 7 The illustration shows a charge pump-based current-mode neuron with a filter weight capacitor according to an aspect of this disclosure, the filter weight capacitor being charged by a current digital-to-analog converter with a filter weight voltage.

[0020] Figure 8 The diagram illustrates a charge pump-based current-mode neuron and bias circuit according to aspects of this disclosure, wherein the output transistor is a PMOS transistor.

[0021] Figure 9 The diagram illustrates an array of charge-pump-based current-mode neurons and a common bias circuit according to aspects of this disclosure, wherein the gate capacitance of the output transistor in each neuron is configured as a filter weight capacitor and wherein the bias circuit includes a VDAC.

[0022] Figure 10 The diagram illustrates an array of charge-pump-based current-mode neurons and a common bias circuit according to aspects of this disclosure, wherein the gate capacitance of the output transistor in each neuron is configured to serve as a filter weight capacitor, and wherein the bias circuit includes an IDAC.

[0023] Figure 11 The diagram illustrates an array of current-mode neurons and a common bias circuit according to aspects of this disclosure, wherein the gate capacitance of the output transistor in each neuron is configured to serve as a filter weight capacitor, and wherein the bias circuit comprises only a VDAC and a switching matrix.

[0024] Figure 12 This is a flowchart of an example method for operating a charge pump-based current-mode circuit according to aspects of this disclosure.

[0025] The embodiments and advantages of this disclosure will be best understood by referring to the following detailed description. It should be understood that the same reference numerals are used to identify one or more of the same elements illustrated in the figures. Detailed Implementation

[0026] Using the traditional von Neumann architecture for machine learning applications is cumbersome because the data flow to and from memory becomes a bottleneck for increasing processing speed. Therefore, in-memory computation architectures have been developed, where data processing hardware is distributed across bit units or neurons. For example, each layer of a machine learning application typically implements one or more filters. Filters are multiplication and accumulation circuits that multiply multiple filter weights by corresponding multiple input bits and sum (accumulate) the results of the multiplications. Each filter weight is stored by an in-memory computation neuron, which is also used to multiply the stored filter weight by the input bits from the input vector. A particularly advantageous in-memory computation neuron is disclosed, which includes filter weight capacitors for storing the filter weights. In some implementations, during the training phase, the first plate of the filter weight capacitor is charged to an analog filter weight voltage, while the second plate of the filter weight capacitor is discharged to ground. Based on training, the analog filter weights are adjusted to appropriate values. For example, a voltage source such as a voltage digital-to-analog converter (VDAC) can generate the analog filter weight voltage. The validation phase can then follow the training phase, in which the analog filter weight voltage is not further altered.

[0027] As neurons are trained and validated, the filter weight capacitors of each neuron are charged to the appropriate filter weight voltage, and the neural network can then process the data in a phase typically represented as computation or prediction. To provide bias to each neuron in the layer, a charge pump capacitor can be coupled to the second plate of the filter weight capacitor, while the upper plate of the filter weight capacitor is coupled to the gate of the output capacitor, which has a drain or source terminal coupled to the read bit line. The charge pump capacitor can be part of a bias circuit that can also bias other neurons. Therefore, the bias circuit can also be represented herein as a common bias circuit, as it is shared across multiple neurons. In the charging stage of the training phase, a current source in the common bias block drives current to the drain of a diode-connected transistor. The gate of the diode-connected transistor is coupled to the charge pump capacitor. Thus, the charge pump capacitor will be charged to the charge pump voltage (VchargeP) during the charging stage. A similar charging stage can appear in the computation phase. With the charging stage complete, the training phase can proceed to the evaluation stage, where the charge pump capacitor is then coupled to the second plate of the filter weight capacitor, resulting in a series stack or combination of the filter weight capacitor and the charge pump capacitor. The filter weight capacitor is thus pumped in a charge pump manner to be charged during the evaluation stage to the sum of the filter weight voltage and the charge pump voltage. A similar evaluation stage follows the charging stage during the computation phase. The difference between the two evaluation stages is that the training phase responds to training data, while the computation phase processes new data. In both evaluation stages, the neurons respond to the binary value of the input bits. Based on this binary value, the first plate of the filter weight capacitor is either coupled to the gate of the output transistor or isolated from the gate of the output transistor, which has a drain coupled to the bit line.

[0028] If the input bit has a binary value that couples the first plate of the filter weight capacitor to the gate of the output transistor, the output transistor and the diode-connected transistor form a current mirror. Therefore, the output transistor will conduct a mirrored current through the bit line, which will vary according to the VDAC voltage. This current variation exists relative to the bias current established by the current source current. For example, suppose VDAC has a two-bit resolution. Then, the mirrored current to the bit line will have four possible values, depending on the resulting VDAC voltage established during the training phase. Given this current mirror and the use of the charge pump capacitor, the neuron disclosed herein can be represented as a charge pump-based current-mode neuron.

[0029] The neuron just described includes a filter weight capacitor and multiple switches for charging the filter weight capacitor. Each switch can be implemented by a corresponding transistor switch. Therefore, the neuron can include several transistors for switching (e.g., four or five switching transistors), as well as filter weight capacitors and output transistors. This set of transistors is similar in number to the in-memory computation bit cells of a conventional static random access memory (SRAM-based) neuron. However, note that the number of filter weight bits that need to be stored by a similar SRAM-based neuron depends on the number of bits per filter weight. If each filter weight is four bits, four SRAM bit cells are needed to store the filter weights. However, the filter weights of the charge pump-based current-mode neuron disclosed herein can be implemented by a VDAC, resulting in a significant increase in density (the semiconductor area required for the implementation of a given neuron).

[0030] The bit line coupled to the drain of the output transistor is also referred to herein as the read bit line. The pre-charging of the read bit line depends on the polarity of the output transistor. If the output transistor is an n-type metal-oxide-semiconductor (NMOS) transistor, the read bit line can be pre-charged by being charged to the supply voltage. If the output transistor is a p-type metal-oxide-semiconductor (PMOS) transistor, the read bit line can be pre-charged by being discharged. In other embodiments, the read bit line can be pre-charged to an intermediate voltage between ground and the supply voltage.

[0031] We will now discuss in more detail some example implementations of in-memory computation formed by charge-pump-based current-mode neurons. Figure 1A Example neuron 100 and associated bias circuitry 101 are shown. Neuron 100 includes a filter weight capacitor (Cweight), an NMOS output transistor M1, and a set of five transistor switches S1, S2, S3, S4, and S5. Switch S1 is coupled between the second plate of the filter weight capacitor and the VchargeP output node of bias circuitry 101. Switch S2 is coupled between the output of voltage digital-to-analog converter (VDAC) 110 and the first plate of the filter weight capacitor. Switch S3 is coupled between the first plate of the filter weight capacitor and the gate of output transistor M1. Switch S4 is coupled between the second plate of the filter weight capacitor and ground. Finally, switch S5 is coupled between the gate of output transistor M1 and ground. The source of output transistor M1 is connected to ground, while the drain of output transistor M1 is connected to the read bit line (RBL).

[0032] The state of a transistor switch depends on the operating mode. As described above, neurons in a neural network are trained during the training phase to establish filter weight values ​​for the trained neurons. The training is then evaluated during the validation phase. Although the filter weights are not changed during the validation phase, the neural network can be tuned, for example, by removing hidden layers during the validation phase. With the neural network trained and validated, it can be used to process input data during the computation phase (also referred to as the prediction phase in this paper). During the charging stage of the training phase, control signals... The control signals control the states of switches S1 and S4. complement Control switches S1, S3, and S5. During the training and evaluation phases of the computation phase, switches S1 and S3 are controlled by data control signals. Control, this data control signal, is activated when the input bit has a binary true state. It is activated. During the evaluation stage, switch S5 is controlled by the data control signal. The two's complement control is activated when the input bit has a binary false state. These switch states (which can also be referred to as switch configurations) will now be discussed in more detail.

[0033] The switching states of switches S1 to S5 depend on whether neuron 100 is in the training, validation, or computation phase. During the charging stage of the training (or computation) phase, the switching states of neuron 100 vary. Figure 1B The following is shown. Switches S2 and S4 respond to the control signal. The switch is closed upon activation. Switches S1, S3, and S5 respond to the complement control signal. The cancellation takes effect and the switch is turned on. This switch configuration for the switch set can be referred to herein as the first switch configuration. With switches S2 and S4 closed, VDAC 110 charges the filter weight capacitor to the filter weight voltage Vdac, which depends on neuron training. In the bias circuit 101, current source I drives bias current I to the drain of diode-connected NMOS transistor M2, which has a source connected to ground. The charge pump capacitor (CchargeP) has a second plate connected to ground and a first plate connected to the gate of diode-connected transistor M2. In response to current I, the gate of diode-connected transistor M2 and the charge pump capacitor are charged to the charge pump voltage VchargeP.

[0034] During the evaluation phase of training, the switching configuration for computing neurons 100 in memory depends on the input bits of the training dataset. Is it binary true or false? During the evaluation level of the computation phase, the same switch configuration behavior exists for its input bits. The following discussion will assume the input bits are active high, but the active low convention can also be used. During the evaluation level of the training (or computation) phase, the set of switches in neuron 100 is configured in response to the binary high state of the input training bits. Figure 1C The switch configuration is shown in the diagram. This switch configuration can also be represented herein as a second switch configuration. Responding to the input bit... In the high state, switches S1 and S3 are closed. Control signal The effect is canceled during the evaluation stage so that switches S2 and S4 are turned on. During the evaluation stage, switch S5 responds to the complement of the input bit. Because the input bits are Figure 1C When the input bit is set to high, switch S5 is opened. As previously described, the charge pump capacitor is charged to the charge pump voltage VchargeP during the charging stage. With switch S1 closed during the subsequent evaluation stage, the first plate of the filter weight capacitor is thus boosted to a boost voltage (the sum of the filter weight voltages Vdac and VchargeP) in a charge pump manner due to the series coupling of the filter weight capacitor and the charge pump capacitor. This boost voltage drives the gate of the output transistor M1 through the closed switch S3. Therefore, the charge pump capacitor and the filter weight capacitor are stacked in series in response to the input bit being true. Depending on the values ​​of the Vdac and VchargeP voltages, the output transistor M1 conducts a current mirror-adjusted version of the current I during the evaluation section to correspondingly reduce the read bit line voltage. For example, if the Vdac voltage is zero, the gate of the output transistor M1 is charged to the VChargeP voltage. Before the training phase, the read bit line voltage is pre-charged to the supply voltage or a portion of the supply voltage. Depending on the value of the Vdac voltage (which represents the filter weight), the bit line is discharged accordingly to equal a certain analog value.

[0035] During the training or computation phase, the input bits can also have a binary false (binary low) value during the evaluation stage. The switching configuration for neuron 100 in response to this binary low state of the input bits during the evaluation stage is... Figure 1D The switch configuration is shown in the diagram. This switch configuration can also be represented herein as a third switch configuration. Since the input bits have binary low values, the two's complement input bits... A binary high signal causes switch S5 to close, grounding the gate of output transistor M1, and switches S1, S2, S3, and S4 to open. Therefore, if the input bit is low, the read bit line floats. The read bit line then remains at its pre-charge voltage.

[0036] As training or input data flows through neurons 100, the sequential charging and evaluation parts, such as... Figure 2 The waveform appears as shown. Specifically, the first input bit from the training data (or input data during the computation phase) is evaluated during the first cycle (cycle 1), where the first cycle (cycle 1) is at time t1 with a control signal. The pulse begins during the charging stage of cycle 1. Then, neuron 100 will be in the first switching configuration, as per [previous information]. Figure 1B The discussion focuses on the control signal following the charging stage. The effect was canceled to begin the evaluation level of cycle 1. In the evaluation level of cycle 1, the input bits... A binary high value will cause neuron 100 to be in the second switch configuration, as shown in the following... Figure 1C The discussion continues. Similarly, the second training bit is evaluated during the second training cycle (cycle 2), which begins at time t2. Cycle 2 also begins at the charging stage, where the control signal... This is enabled. In the evaluation section of cycle 2, the input training bit is low, causing neuron 100 to be in the third switch configuration, as per [the previous sentence]. Figure 1D The subject of discussion.

[0037] During the computation phase, the machine learning application, including neuron 100, processes the input data. As data passes through the various layers and reaches the layer containing neuron 100, the input bit is received by neuron 100 and processed using the Vdac voltage stored on the filter weight capacitor. If the filter weight capacitor has no leakage, a charging phase is not needed to refresh the Vdac voltage on the filter weight capacitor before it processes the input bit. However, to address any such leakage, the computation phase can be initiated from the charging stage (which can also be represented as...). Figure 2 The refresh level shown in the diagram begins. Control signals. and complement control signal During this charging refresh period, such as regarding Figure 1B The operation is as discussed so that the filter weight capacitor voltage Vdac and the charging capacitor voltage VchargeP can be refreshed. If the input bit is in a logic high state after refresh charging, neuron 100 will be in a second switching configuration, such that the gate of the output transistor M1 is charged to the sum of the Vdac and VChargeP voltages. Conversely, if the input bit is low after refresh, neuron 100 will be in a third switching configuration, such that the read bit line floats (remains in its pre-charged state).

[0038] To increase density, the bias circuit 101 can be configured as follows: Figure 3The multiple neurons 100 in the in-memory computing array 300 shown are shared. The VchargeP voltage can be shared by multiple neurons 100, or it can be varied for each neuron 100. The Vdac voltage is set for each neuron 100 according to training. Each neuron can cause its read bit line to discharge or float depending on its input bit. The size of each filter weight (width in bits) depends on the precision of VDAC. In typical applications, filter weights are typically 4-bit or 8-bit wide, although this value can be reduced or increased in alternative implementations. Note that such filter weight SRAM implementations require four or eight bit units respectively. However, in each neuron 100, the functionality of these four or eight bit units can be performed using only five transistor switches S1 to S5, the output transistor, and the filter weight capacitor. The multi-bit filter weight implementation of the neuron 100 thus provides a significant increase in density compared to the SRAM implementation. Even if the filter weight is only one bit, the density of the neuron 100 is still comparable to that of a conventional SRAM bit unit. Furthermore, the neuron 100 does not require separate matching in addition to conventional analog circuitry matching.

[0039] Multiple neurons 100 can be organized to form a filter. As mentioned earlier, a filter is a multiplication and accumulation circuit that multiplies multiple filter weights with their corresponding multiple input bits and sums (accumulates) the results of the multiplication. Figure 4 An example filter 400 is shown, which includes multiple neurons 100, all of which are coupled to a common readout bit line RBL. For clarity, the common bias circuit is not shown. Figure 4 It is shown in the diagram, but can be arranged as discussed with respect to bias circuit 101.

[0040] As disclosed herein, there are several alternative implementation methods for charge pump-based current-mode neurons. For example, Figure 5 The diagram shows 500 neurons. Figure 5 In this circuit, the filter weight capacitors are implemented using a capacitive digital-to-analog converter (CDAC) that functions as a variable capacitor. The CDAC is trained similarly to that discussed for VDAC 110 so that neuron 500 does not require a VDAC. Bias circuit 101 functions as discussed for neuron 100. Switch S2 is coupled between the first plate of charge pump capacitor CchargeP and the first plate of CDAC. Switch S4 is coupled between the second plate of CDAC and ground. Control signal Control switches S2 and S4 are closed during the charging stage, causing the CDAC to be charged to VchargeP. Switch S3 is coupled between the first plate of the CDAC and the gate of the output transistor M1, similar to that discussed for neuron 100. During the computation phase, switch S1 can be selectively coupled to ground or to a node of the supply voltage Vdd, depending on the input bit. The value of Vdd is used. For example, if the input bit is binary high, switch S1 selects Vdd so that the second plate of the CDAC is charged accordingly. The first plate of the CDAC is then pumped onto CchargeP by a voltage that depends on the CDAC settings established during the training phase. Switch S5 functions as discussed for neuron 100.

[0041] For greater flexibility, the CDAC can be combined with the VDAC 110, such as... Figure 6 The bias circuit 101 is arranged as discussed for neuron 100. Switches S1 to S5 are arranged as discussed for neuron 500, except that switch S1 is coupled to VDAC 110.

[0042] In another alternative embodiment of neuron 100, the function of VDAC 100 can be determined by, for example... Figure 7 The current DAC (IDAC) 710 shown is implemented. The common bias circuit 705 includes a current source 105, a diode-connected transistor M2, and a charge pump capacitor as discussed with respect to bias circuit 101. The charge pump capacitor is thus charged to VchargeP by these circuit elements. Furthermore, the bias circuit 705 includes an IDAC 710, which is coupled to the first plate of the charge pump capacitor via a switch S1. Switch S1 is closed during a single or unique charging phase for each neuron. If there are n neurons, switch S1 therefore responds to n corresponding clock phases. During training, the charge pump capacitor is therefore charged to the equivalent value of VchargeP + Vdac according to the IDAC 710 settings. The filter weight capacitor Cweight has a second plate connected to ground and a first plate coupled to the first plate of the charge pump capacitor via switch S2. The first plate of the filter weight capacitor is also coupled to the gate of the output transistor M1. Switch S3 is coupled between the gate of the output transistor and ground. Control signal Control switch S2. Similarly, the two's complement input bits. Control switch S3.

[0043] In neurons 100, 500, 600, and 700, the output transistor M1 and the diode-connected transistor M2 are NMOS transistors. These transistors are implemented using PMOS transistors. Figure 8 The neuron 800 and bias circuit 805 are shown. The bias circuit includes a PMOS diode-connected transistor P1, which has a drain coupled to ground via a current source 810. The source of the diode-connected transistor P1 is coupled to a node for the power supply voltage Vdd. A charge pump capacitor CchargeP has a first plate coupled to the gate of the diode-connected transistor P1 and a second plate coupled to the power supply node. The first plate of the charge pump capacitor is coupled to the second plate of the filter weight capacitor CweightP via switch S1. During the charging stage, the first plate of the filter weight capacitor is coupled to ground via switch S2. During the charging stage, VDAC 110 is coupled to the second plate of the filter weight capacitor via switch S3. In the evaluation stage, the first plate of the filter weight capacitor is coupled to the gate of the PMOS output transistor P2 via switch S4, depending on the binary value of the input bit. During the evaluation stage, the gate of the output transistor P2 is also coupled to a node for the power supply voltage via switch S5, which is controlled by the complement of the input bit. During the charging stage, when the control signal... When activated to charge the filter weight capacitor to -Vdac, switches S2 and S3 are closed. After the charging stage, switches S2 and S3 are opened. If the input bit... If the input bit is binary high during the subsequent computation phase, switches S1 and S4 are closed so that the gate of the output transistor is charged to the boost voltage VchargeP-Vdac. Therefore, neuron 800 stacks the filter weight capacitor and charge pump capacitor in series, similar to the configuration discussed for neuron 100. The output transistor P2 then conducts a mirror version of the current conducted by current source 810. The ratio of this mirror current to the current conducted by current source 810 depends on the boost voltage VchargeP-Vdac. The read bit line is coupled to the drain of output transistor P2 and is discharged (or partially discharged) before the computation phase. Due to the conduction of the mirror current, the read bit line voltage will therefore rise from ground or that portion of the voltage level. If the input bit is binary 0 during the evaluation phase, switch S5 is closed to turn off transistor P2 and allow the read bit line to float.

[0044] The separate filter weight capacitors discussed for neurons 100, 500, 600, 700, and 800 can be eliminated in an alternative implementation in which the gate capacitance of the output transistor is used as the filter weight capacitor. Figure 9Example neuron 900 is shown. Common bias circuitry 905 includes a charge source such as VDAC 110, switches S1 through S4, a filter weight capacitor, a current source 105, a diode-connected transistor M2, and a charge pump capacitor, all arranged as discussed for neuron 100. However, neuron 900 only includes output transistor M1 and a bit line switch S5 coupled between the drain of output transistor M1 and the read bit line. During the charging stage, switches S2 and S4 are closed to allow the filter weight capacitor to be charged to Vdac. During the evaluation stage, switches S1 and S3 are closed and switches S2 and S4 are open to allow the filter weight capacitor to be charged to the boost voltage VchargeP + Vdac.

[0045] Although the bias circuit 905 is shared by multiple neurons, VDAC 110 is set to any appropriate value determined by the training phase of a particular neuron. To apply the resulting boosted voltage VchargeP+Vdac to the gate of the output transistor M1 in the selected neuron, the first plate of the filter weight capacitor is switched by switch S3 (in control signal). (The pulse is closed after the signal is received) and the switching matrix 910 is coupled to the gate of the output transistor M2 of the selected neuron. Figure 9 For clarity, additional neurons other than neuron 900 are not shown. However, the gate of the output transistor of each neuron can be coupled to the filter weight capacitor via switch matrix 910. Therefore, during the charging stage, the gate capacitance of the selected output transistor M1 is charged to the boost voltage VchargeP + Vdac. In subsequent computational phases of neuron 900, switch matrix 910 is configured to isolate the gate of the output transistor from the filter weight capacitor. Switch S5 is turned on or off depending on the input bit value. If switch S5 is closed, output transistor M1 will conduct a mirrored current through the read bit line, as discussed for neuron 100.

[0046] Since neuron 900 uses its gate capacitance as a filter weight capacitor, neuron 900 can also be represented as a gate capacitor cell. The bias circuit for such a gate capacitor cell can be implemented in various alternative embodiments. For example, bias circuit 1110 can be implemented as follows: Figure 10The switch matrix 910 shown biases the array 1000 of gate capacitor cells. The IDAC 1005 in the bias circuit 1110 is trained during the training phase, similar to that discussed with respect to VDAC 110. The IDAC 1005 drives the drain of the diode-connected transistor M2. The first plate of the charge pump capacitor is coupled to the gate of the diode-connected transistor and, via the switch matrix 910, to the gate of the selected output transistor. During the charging state, the filter weight capacitor is then charged to a voltage VchargeP depending on the training voltage of the IDAC 1005. The array 1000 includes n gate capacitor cells ranging from the first gate capacitor cell 1115 to the nth gate capacitor cell 1120. Each neuron / gate capacitor cell is arranged as discussed with respect to neuron 900.

[0047] During the charging stage, the switch matrix 910 is configured such that the gate of the output transistor of the selected neuron is coupled to a charged filter weight capacitor. The gate capacitance of the selected output transistor is then charged accordingly so that, during the evaluation stage, a mirror current is drawn from the corresponding read bit line based on the value of the input bit of the selected neuron. For example, the bit line switch S5 in neuron 1115 is controlled by the input bit... Control. Similarly, input bits. Control bit line switch S5 in neuron 1120.

[0048] In yet another implementation of the gate capacitor cell, the bias circuit 1105 may consist only of a charge source such as VDAC 110 and a switching matrix 910, as... Figure 11 As shown in the diagram, Vdac 110 is trained during the training phase to provide an appropriate Vdac voltage. During the evaluation phase, the resulting Vdac voltage is coupled to selected neurons from the gate capacitor cell array 1100 via a switch matrix 910. Each neuron is implemented as discussed for array 1000.

[0049] Now about Figure 12The flowchart discusses a method for operating a charge pump-based current-mode neuron. The method includes action 1200, which drives current through a diode-connected transistor to charge a charge pump capacitor to a charge pump voltage. Charging of the charge pump capacitors in bias circuits 101, 305, and 805 is an example of action 1200. The method also includes action 1205, which responds to a training dataset and includes adjusting a voltage digital-to-analog converter (VDAC) to charge a filter weight capacitor with a VDAC voltage. Operation of VDAC 110 during the evaluation stage of the training phase in neurons 100, 800, and bias circuit 1110 is an example of action 1205. Furthermore, the method includes action 1210, which responds to a first binary value of an input bit and includes series coupling of the charge pump capacitor and the filter weight capacitor to charge the gate of an output transistor to the sum of the VDAC voltage and the charge pump voltage. A second switching configuration of neuron 100 and an equivalent switching configuration of neuron 800 are examples of action 1210. Finally, the method includes action 1215, which, in response to charging of the gate of the output transistor, conducts a mirror version of the current through the output transistor. The conduction by the output transistor M1 in neuron 100 or the output transistor P2 in neuron 800 is an example of action 1215.

[0050] This disclosure will now be summarized by the following example clauses:

[0051] Clause 1. A memory for machine learning, comprising:

[0052] Bias circuit;

[0053] Bit lines; and

[0054] Neurons, including:

[0055] An output transistor having a drain connected to the bit line;

[0056] Filter weight capacitor;

[0057] Digital-to-analog converter;

[0058] A first switch is coupled between the bias circuit and the filter weight capacitor;

[0059] A second switch is coupled between the digital-to-analog converter and the filter weighting capacitor; and

[0060] A third switch is coupled between the filter weight capacitor and the gate of the output transistor.

[0061] Clause 2. The memory according to Clause 1, wherein the neuron further comprises:

[0062] The fourth switch is coupled between ground and the filter weight capacitor.

[0063] Clause 3. The memory according to Clause 2, wherein the neuron further comprises:

[0064] The fifth switch is coupled between ground and the gate of the output transistor.

[0065] Clause 4. The memory according to any one of Clauses 1-3, wherein the output transistor is an n-type metal-oxide-semiconductor (NMOS) transistor.

[0066] Clause 5. The memory according to Clause 1, wherein the bias circuitry comprises:

[0067] Diode-connected transistor;

[0068] A charge pump capacitor, coupled to the gate of the diode-connected transistor; and

[0069] A current source is coupled to the drain of the diode-connected transistor.

[0070] Clause 6. The memory according to Clause 5, wherein the diode-connected transistor includes an NMOS diode-connected transistor.

[0071] Clause 7. The memory according to any one of Clauses 1-6, wherein the filter weight capacitor includes a first plate and a second plate, the first switch is coupled between the second plate and the bias circuit, the second switch is coupled between the first plate and the digital-to-analog converter, and the third switch is coupled between the first plate and the gate of the output transistor.

[0072] Clause 8. The memory pursuant to Clause 1, wherein the filter weighting capacitor comprises a capacitive digital-to-analog converter.

[0073] Clause 9. The memory according to Clause 8, wherein the capacitive digital-to-analog converter includes a first plate and a second plate, wherein a first switch is coupled between the first plate and the bias circuit, a second switch is coupled between the second plate and the digital-to-analog converter, and a third switch is coupled between the first plate and the gate of the output transistor.

[0074] Clause 10. The memory according to Clause 1, wherein the output transistor comprises a p-type metal-oxide-semiconductor (PMOS) output transistor.

[0075] Clause 11. The memory according to Clause 10, wherein the filter weight capacitor includes a first plate and a second plate, the first switch is coupled between the second plate and the bias circuit, the second switch is coupled between the second plate and the digital-to-analog converter, and the third switch is coupled between the first plate and the gate of the output transistor, the neuron further comprising:

[0076] A fourth switch is coupled between the first electrode and ground; and

[0077] The fifth switch is coupled between the gate of the output transistor and the node for the power supply voltage.

[0078] Clause 12. A memory for machine learning applications, comprising:

[0079] Read the bit lines;

[0080] An output transistor having a drain connected to the read bit line;

[0081] Charge pump capacitor;

[0082] Filter weight capacitor;

[0083] Voltage source; and

[0084] The set of switches has a first switch configuration in which the filter weight capacitor is charged by the voltage source, and the set of switches has a second switch configuration in which the filter weight capacitor is coupled to the charge pump capacitor to provide a boost voltage to the gate of the output transistor.

[0085] Clause 13. The memory pursuant to Clause 12 further includes:

[0086] A bias circuit includes the charge pump capacitor and a current source, wherein the current source is configured to charge the charge pump capacitor to a bias voltage, and wherein the voltage source includes a voltage-to-analog converter.

[0087] Clause 14. The memory according to Clause 13, wherein the set of switches further includes a third switch configuration in which the output transistor is turned off.

[0088] Clause 15. The memory according to Clause 14, wherein the second switch is configured in response to an input bit being true, and the third switch is configured in response to an input bit being false.

[0089] Clause 16. The memory according to any one of Clauses 13-15, wherein the bias circuitry further comprises a diode-connected transistor having a drain connected to the current source and a gate connected to the charge pump capacitor.

[0090] Clause 17. The memory according to Clause 16, wherein the diode-connected transistor is a PMOS transistor having a source connected to a node for a power supply voltage.

[0091] Clause 18. The memory according to Clause 16, wherein the diode-connected transistor is an NMOS transistor having a source connected to ground.

[0092] Clause 19. A memory for machine learning applications, comprising:

[0093] Multiple neurons, each neuron comprising:

[0094] Bit line;

[0095] Position line switch; and

[0096] An output transistor having a terminal connected to the bit line via the bit line switch; and

[0097] Bias circuit, including:

[0098] Charge source; and

[0099] A switching matrix configured to couple the charge source to a selected neuron in the neuron to charge the gate capacitance in the output transistor of the selected neuron with a boost voltage.

[0100] Clause 20. The memory according to Clause 19, wherein each bit line switch is configured to close in response to the binary state of the corresponding input bit.

[0101] Clause 21. The memory according to Clause 19, wherein the bias circuitry further includes a current digital-to-analog converter configured to charge the charge pump capacitor with the boost voltage.

[0102] Clause 22. The memory according to Clause 21, wherein the biasing circuitry further comprises:

[0103] A diode-connected transistor having a gate and a drain coupled to the current-to-analog converter.

[0104] Clause 23. The memory according to Clause 19, wherein the charge source comprises:

[0105] Filter weight capacitor;

[0106] Digital-to-analog converters; and

[0107] The set of switches has a first switch configuration in which the filter weight capacitor is charged by the digital-to-analog converter, and the set of switches has a second switch configuration in which the filter weight capacitor is coupled to a charge pump capacitor to provide the boost voltage to the switch matrix.

[0108] Clause 24. The memory according to Clause 23, wherein the bias circuitry further comprises:

[0109] A diode-connected transistor having a gate coupled to the charge pump capacitor.

[0110] Clause 25. The memory according to Clause 24, wherein the biasing circuitry further comprises:

[0111] A current source is configured to drive current into the drain of the diode-connected transistor.

[0112] Clause 26. A method for manipulating neurons, comprising:

[0113] The drive current passes through the diode-connected transistor to charge the charge pump capacitor to the charge pump voltage;

[0114] In response to the training dataset, the voltage of the digital-to-analog converter (VDAC) is adjusted to charge the filter weight capacitors with the VDAC voltage;

[0115] In response to the first binary value of the input bit, the charge pump capacitor and the filter weight capacitor are coupled in series to charge the gate of the output transistor to the sum of the VDAC voltage and the charge pump voltage; and

[0116] A mirror version of the current is conducted through the output transistor in response to the charging of the gate of the output transistor.

[0117] Clause 27. The method according to Clause 26, wherein charging the filter weight capacitor with the VDAC voltage further comprises: turning on a first switch to couple a first plate of the filter weight capacitor to the voltage-to-analog converter, and turning on a second switch to ground a second plate of the filter weight capacitor.

[0118] Clause 28. The method described pursuant to Clause 26 further includes:

[0119] In response to the second binary value of the input bit, the output transistor is turned off.

[0120] Clause 29. The method according to Clause 28, wherein turning off the output transistor includes grounding the gate of the output transistor.

[0121] Clause 30. The method according to Clause 28, wherein turning off the output transistor includes charging the gate of the output transistor to a power supply voltage.

[0122] It should be understood that many modifications, substitutions, and variations can be made to the materials, apparatus, configurations, and methods of using the equipment disclosed herein without departing from the scope of this disclosure. Therefore, the scope of this disclosure should not be limited to the specific embodiments shown and described herein, as these are merely examples, but should be fully proportionate to the scope of the appended claims and their functional equivalents.

Claims

1. A memory for machine learning, comprising: Bias circuit, wherein the bias circuit includes a charge pump capacitor; Bit line; as well as Neurons, including: An output transistor having a drain connected to the bit line; The filter weighting capacitor includes a first plate and a second plate; Digital-to-analog converter; A first switch is coupled between the output node of the bias circuit and the second plate of the filter weight capacitor; The second switch is coupled between the output terminal of the digital-to-analog converter and the first plate of the filter weight capacitor; A third switch is coupled between the first plate of the filter weight capacitor and the gate of the output transistor; A fourth switch is coupled between ground and the second plate of the filter weight capacitor; and The fifth switch is coupled between ground and the gate of the output transistor.

2. The memory of claim 1, wherein the output transistor comprises an n-type metal-oxide-semiconductor (NMOS) transistor.

3. The memory according to claim 1, wherein the bias circuit comprises: Diode-connected transistor; A current source is coupled to the drain of the diode-connected transistor. The charge pump capacitor is coupled to the gate of the diode-connected transistor.

4. The memory according to claim 3, wherein the diode-connected transistor comprises an NMOS diode-connected transistor.

5. The memory of claim 1, wherein the filter weight capacitor comprises a capacitive digital-to-analog converter.

6. The memory of claim 1, wherein the output transistor comprises a p-type metal-oxide-semiconductor (PMOS) output transistor.

7. A method for operating on neurons included in a memory for machine learning according to claim 1, comprising: The drive current passes through the diode-connected transistor to charge the charge pump capacitor to the charge pump voltage; In response to the training dataset, the voltage digital-to-analog converter (VDAC) is adjusted to charge the filter weight capacitors with the VDAC voltage; In response to the first binary value of the input bit, the charge pump capacitor and the filter weight capacitor are coupled in series to charge the gate of the output transistor to the sum of the VDAC voltage and the charge pump voltage; as well as A mirror version of the current is conducted through the output transistor in response to the charging of the gate of the output transistor.

8. The method of claim 7, wherein charging the filter weight capacitor with the VDAC voltage further comprises: The first switch is turned on to couple the first plate of the filter weight capacitor to the voltage-to-analog converter, and the second switch is turned on to ground the second plate of the filter weight capacitor.

9. The method according to claim 7, further comprising: In response to the second binary value of the input bit, the output transistor is turned off.

10. The method of claim 9, wherein turning off the output transistor comprises grounding the gate of the output transistor.

11. The method of claim 9, wherein turning off the output transistor comprises charging the gate of the output transistor to a power supply voltage.

Citation Information

Patent Citations

  • Shift register circuit

    CN101147202A

  • A shift register circuit having threshold voltage compensation

    CN101765876A