In-memory computing crossbar and optimization method

The crossbar structure with zero-weight columns and adjustable analog-to-digital converters addresses the reliability and power consumption issues in analog IMC platforms, enhancing the accuracy and efficiency of MAC operations by mitigating noise and optimizing power usage.

WO2024245532A1PCT designated stage expired Publication Date: 2024-12-05HUAWEI TECH CO LTD +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
PCT/EP2023/064362
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-30
Publication Date
2024-12-05

AI Technical Summary

Technical Problem

Analog in-memory computing (IMC) platforms face reliability issues and high power consumption due to noise amplification during the digitization process of MAC operations, particularly in digital signal processing, which affects the accuracy of computations.

Method used

A crossbar structure with a plurality of input and output lines and cross-point devices, where at least one column has a zero weight to mitigate noise variations, and analog-to-digital converters with adjustable resolution to reduce power consumption, allowing for optimal trade-offs between reliability and complexity.

Benefits of technology

The solution enhances the reliability of analog IMC platforms by controlling noise levels and reducing power consumption during digitization, thereby improving the accuracy and efficiency of MAC operations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure EP2023064362_05122024_PF_FP_ABST
    Figure EP2023064362_05122024_PF_FP_ABST
Patent Text Reader

Abstract

The present disclosure relates to analog in-memory computing (BMC) platforms, especially to analog IMC crossbar structures based on static random access memory (SRAM) and resistive random access memory (RRAM) devices. A crossbar structure of this disclosure comprises a plurality of input lines and a plurality of output lines arranged in a crossed pattern; and a plurality of cross-point devices arranged in a plurality of rows corresponding to the plurality of input lines and in a plurality of columns corresponding to the plurality of output lines, wherein each cross-point device is arranged at a cross-point of one input line and one output line and is connected to the input line and the output line. In this regard, each input line is configured to receive a respective input signal. Furthermore, each cross-point device is configured to provide a weight for multiplication with the respective input signal received by the input line it is connected to. Moreover, at least one column of the plurality of columns comprises at least one cross-point device configured to provide a zero weight.
Need to check novelty before this filing date? Find Prior Art

Description

[0001]IN-MEMORY COMPUTING CROSSBAR AND OPTIMIZATION METHOD TECHNICAL FIELD The present disclosure relates to analog in-memory computing (IMC) platforms, for example, to analog IMC crossbar structures based on static random access memory (SRAM) and resistive random access memory (RRAM) devices. BACKGROUND Generally, digital multiplication and accumulation (MAC) operations may be used for equalization in communication systems to compensate linear and non-linear distortions. Normally, a large number of MAC operations may be needed in digital signal processing (DSP) due to their large number of adjustable parameters, which may contribute significantly to the total power consumption. In order to address such high power consumptions, analog IMC platforms, for example, based on devices like SRAM or emerging devices like RRAM, may be used to perform said MAC operations in the analog domain. For example, in the case of an analog IMC platform based on RRAM, the weights of an equalizer may be represented with the conductance of RRAM cells. As such, the partial products may be calculated in the analog domain, then may be converted to the digital domain, and digital shift-and-add operations may be applied accordingly. However, said digitization process may be affected by noise that may be further amplified by the shift operations, which may severely affect the reliability of the computation. SUMMARY In view of the above, this disclosure aims to provide an improved crossbar structure for IMC and an optimization method. An objective is to increase the reliability in analog IMC platforms. Another objective is to control the power consumption of the digitization process. These and other objectives are achieved by the solutions described in the independent claims. Advantageous implementations are further defined in the dependent claims. According to a first aspect of this disclosure, a crossbar structure for IMC is provided. The crossbar structure comprises a plurality of input lines and a plurality of output lines arranged in a crossed pattern; and a plurality of cross-point devices arranged in a plurality of rows corresponding to the plurality of input lines and in a plurality of columns corresponding to the plurality of output lines, wherein each cross-point device is arranged at a cross-point of one input line and one output line and is connected to the input line and the output line. In this regard, each input line is configured to receive a respective input signal. Furthermore, each cross-point device is configured to provide a weight for multiplication with the respective input signal received by the input line it is connected to. Moreover, at least one column of the plurality of columns comprises at least one cross-point device configured to provide a zero weight. Therefore, this disclosure presents a solution to adjust the noise levels according to the significance of the individual partial products, especially to mitigate the weight variations. For example, for binary expansion, the most and least significant portions may be referred to by the most significant bit (MSB) and the least significant bit (LSB), respectively. The solution of this disclosure foresees to provide IMC crossbar columns such that the columns together may have at least one non-zero or non-negative weight per row, and at least one zero weight in at least one column, for example, to reduce the amount of noise variation accumulated per column. Moreover, the number of columns may be defined according to the significance of the column, e.g., with higher significance corresponding to a greater number of columns, for instance, to achieve an optimal tradeoff between reliability and complexity. In an implementation form of the first aspect, each of the plurality of columns comprises an equal number of cross-point devices configured to provide a non-zero weight. Additionally, for example, the weights may be distributed over the columns without repetition. This may further reduce the amount of noise variation accumulated per column due to, for instance, the equal distribution of the weights over the columns. In an implementation form of the first aspect, each cross-point device is configured to provide a respective output signal at the output line it is connected to, wherein the respective output signal corresponds to a multiplication of the respective input signal received by the input line it is connected to and the weight provided by the cross-point device. For example, the input signal may be an electrical voltage signal and the output signal may be an electrical current signal. The weights may correspond to electrical conductance. The respective voltage signals may be received by the respective cross-point device in the horizontal direction along the crossbar structure, whereby providing the respective current signals by means of the conductance according to the Ohm’s law, which may be added or accumulated in the vertical direction along the crossbar structure according to the Kirchhoff’s current law. In an implementation form of the first aspect, the crossbar structure further comprises a plurality of analog-to-digital converters, wherein a respective analog-to-digital converter of the analog- to-digital converters is arranged in each column of the plurality of columns, and the respective analog-to-digital converter is connected to the output line corresponding to the column. In this regard, the respective analog-to-digital converter is configured to receive an accumulated signal from the output line, wherein the accumulated signal corresponds to an accumulation of the respective output signals from the cross-point devices arranged in the column, and to quantize the accumulated signal to generate a digital signal. Additionally, the crossbar structure may comprise a plurality of current-to-voltage converters, wherein a respective current-to-voltage converter of the current-to-voltage converters may be arranged in each column of the plurality of columns, especially between the output line corresponding to the column and the respective analog-to-digital converter. In this regard, the respective current-to-voltage converter may receive the accumulated signal from the output line, may convert the accumulated signal to generate a voltage signal, and to feed the voltage signal to the respective analog-to-digital converter for quantization. In an implementation form of the first aspect, the respective analog-to-digital converter has a resolution determined by the number of cross-point devices arranged in the column and configured to provide a non-zero weight. This may allow for quantization with low-resolution analog-to-digital converters, thereby controlling or reducing the power consumption of the deployed analog-to-digital converters. In an implementation form of the first aspect, the crossbar structure further comprises a digital computing circuitry connected to a set of analog-to-digital converters, each arranged in a respective column of a set of columns of the plurality of columns. In this regard, the digital computing circuitry is configured to receive the digital signals from the analog-to-digital converters of the set of analog-to-digital converters, and to perform a digital operation on the digital signals. The digital operation may comprise digital summation or addition operations, digital shift operations, or a combination thereof. For example, the digital computing circuitry may receive the converted signals, may shift one or more converted signals of the converted signals according to their significance, and may add the signals afterwards. In an implementation form of the first aspect, the input signal corresponds to a signed input signal or an unsigned input signal. Additionally or alternatively, the weights correspond to signed weights or unsigned weights. This may allow for an improved IMC for both signed and unsigned values, e.g., for both negative and positive values. In an implementation form of the first aspect, each of the plurality of cross-point devices comprises a RRAM device configured to generate a programmable conductance to realize the weight. This may allow for the realization of the weights in a simple and straightforward manner, whereby reducing the feature size of the implemented circuit. In an implementation form of the first aspect, each of the plurality of cross-point devices comprises a SRAM device, a switching arrangement, and a resistor bank connected to the switching arrangement. In this regard, the SRAM device is configured to control the switching arrangement to generate a programmable conductance via the resistor bank to realize the weight. For example, the SRAM device may store the switch controls as digital values and may control the switching arrangement accordingly to control the total or resultant conductance of the resistor bank. In an implementation form of the first aspect, each of the plurality of cross-point devices comprises a SRAM device, a switch, and a resistor connected to the switch. In this regard, the SRAM device is configured to control the switch to generate a programmable binary conductance via the resistor to realize the weight. For example, the SRAM device may store the switch controls as binary values and may control the switch accordingly to control the conductance of the resistor. According to a second aspect of this disclosure, a method is provided for optimizing a crossbar array structure for in-memory computing. The crossbar array structure comprises a plurality of cross-point devices arranged in a plurality of rows, corresponding to a plurality of input lines and each for receiving a respective input signal, and arranged in a column corresponding to an output line for providing an output signal, wherein each of the cross-point devices is connected to one input line and the output line and is configured to provide a weight to multiply with the respective input signal received by the input line it is connected to. The method comprises the steps of providing an additional column comprising a plurality of additional cross-point devices, and configuring the cross-point devices of the column and the additional column such that at least one of the column and the additional column comprises at least one cross-point device configured to provide a zero weight. In an implementation form of the second aspect, the method further comprises the step of configuring the cross-point devices of the column and the additional column such that each of the column and the additional column comprises an equal number of cross-point devices configured to provide a non-zero weight. In an implementation form of the second aspect, the method further comprises the steps of receiving an accumulated signal from the output line of each of the column and the additional column, wherein the accumulated signal corresponds to an accumulation of the output signals from the cross-point devices arranged in the column and the additional column, setting a resolution for an analog-to-digital conversion of the accumulated signal from each of the column and the additional column, wherein the resolution is determined by a total number of non-zero weights in the column and the additional column, and quantizing the accumulated signal from each of the column and the additional column to generate a respective digital signal. In an implementation form of the second aspect, the method further comprises the step of performing a digital operation on the digital signals from each of the column and the additional column. It is to be noted that the method according to the second aspect corresponds to the crossbar structure according to the first aspect and its implementation forms. Accordingly, the method of the second aspect may have corresponding implementation forms. Further, the method of the second aspect achieves the same advantages and effects as the crossbar structure of the first aspect and its respective implementation forms. BRIEF DESCRIPTION OF THE DRAWINGS The above described aspects and implementation forms will be explained in the following description of specific exemplary embodiments in relation to the enclosed drawings, in which: Fig.1 shows an exemplary crossbar array structure according to an exemplary design; Fig. 2 shows a first exemplary embodiment of the crossbar structure according to this disclosure; Fig. 3 shows a second exemplary embodiment of the crossbar structure according to this disclosure; Fig. 4 shows a first exemplary embodiment of the cross-point device according to this disclosure; Fig. 5 shows a second exemplary embodiment of the cross-point device according to this disclosure; Fig. 6 shows a third exemplary embodiment of the cross-point device according to this disclosure; Fig. 7 shows a fourth exemplary embodiment of the cross-point device according to this disclosure; Fig.8 shows an exemplary MAC operation according to this disclosure; and Fig.9 shows an exemplary flow diagram of the method according to this disclosure. DETAILED DESCRIPTION OF EMBODIMENTS Reference will now be made in detail to exemplary embodiments of the present disclosure, which are illustrated in the accompanying drawings. However, the following embodiments of the present disclosure may be variously modified and the range of the disclosure is not limited by the following embodiments. Reference signs for similar entities in different embodiments are partially omitted. In Fig.1, a crossbar array structure, especially one column 100 of the crossbar array structure, is illustrated according to an exemplary design. The column 100 may comprise four input lines 1011-1014, respectively connected to four cross-point devices 1021-1024, each producing a respective output signal I11-I14to an output line 103 to which the cross-point devices 1021-1024connected to. As such, the respective input signals may be received by the respective cross-point devices 1021-1024through the input lines 1011-1014in the horizontal direction, may be respectively multiplied by the respective weights W1-W4provided by the cross-point devices 1021-1024, and the resulting output signals I11-I14 may be added in the vertical direction as an accumulated signal IA1. Furthermore, an analog-to-digital converter (ADC) 104 may be provided in the column 100 to receive and to quantize the accumulated signal IA1, especially to generate a digital signal 105. The ADC output 105 may be affected by noise, which may be amplified by any further digital operations such as digital shift operations, e.g., due to the multiplication with an integer, thereby affecting the reliability of the computation severely. In Fig.2, a first exemplary embodiment of the crossbar structure 200 is illustrated according to this disclosure. The crossbar structure 200 may comprise four input lines 1011-1014, and two output lines 103, 203, where the input lines 1011-1014and the output lines 103, 203 may be arranged in a crossed pattern. The crossbar structure 200 may further comprise eight cross-point devices 1021-1024, 2021-2024arranged in four rows corresponding to the input lines 1011-1014and in two columns corresponding to the output line 103, 203, i.e., each column 201 may comprise four cross-point devices of the eight cross-point devices 1021-1024, 2021-2024. Accordingly, each cross-point device 1021-1024, 2021-2024 may be arranged at a cross-point of one input line and one output line and may be connected to said input line and said output line, and may receive a respective input signal via said input line. Furthermore, each cross-point device 1021-1024, 2021-2024 may provide a respective weight W1-W4 for multiplication with the respective input signal. For example, along the output line 103, the cross-point device 1021 may provide a weight W1 in terms of a conductance level or value, exemplarily a non-zero weight, may receive an input voltage signal via the input line 1011, and may further provide an output current signal I11 at the output line 103 by means of the multiplication of the input voltage signal and the conductance value. In this regard, the weight value may be represented by N-bits, and may be mapped to 2N conductance levels, e.g., between 0 and Gmax. Similarly, for example, along the output line 203, the cross-point device 2021 may provide a weight in terms of a conductance level or value, exemplarily a zero weight, may receive the input voltage signal via the input line 1011, and may further provide an output current signal I21at the output line 203 by means of the multiplication of the input voltage signal and the conductance value. Due to the zero weight, the output current signal I21 may correspond to a null or zero current. As such, each of the cross-point devices 1021-1024, 2021-2024 may be configured such that each of the cross-point devices 1021-1024, 2021-2024 may provide either a non-zero weight or a zero weight. Additionally or alternatively, each of the cross-point devices 1021-1024, 2021-2024may be configured such that each column may comprise an equal number of non-zero weights or zero weights or an unequal number of non-zero weights or zero weights. Furthermore, each column of the crossbar structure 200 may comprise an ADC 104, 204 connected to the output line 103, 203 corresponding to the column. The ADC 104, 204 may receive the accumulated signal IA1, IA2corresponding to an accumulation of the respective output signals, e.g., the accumulated signal IA1corresponding to an accumulation of the output signals I11-I14, and the accumulated signal IA2corresponding to an accumulation of the output signals I21-I24. The ADC 104, 204 may further quantize the accumulated signal IA1, IA2 to generate a digital signal 105, 205. In this regard, the respective ADC 104, 204 may have a resolution corresponding to the number of non-zero weights in the respective column. For example, the cross-point devices 1021-1024 along the output line 103 may provide two non-zero weights W1, W2, and thus the ADC 104 may have a resolution of 2 bits in case both weights and inputs represent binary values. Furthermore, the cross-point devices 2021-2024 along the output line 203 may provide two non- zero weights W3, W4, and thus the ADC 204 may also have a resolution of 2 bits in case both weights and inputs represent binary values. Optionally, each column of the crossbar structure 200 may comprise a current-to-voltage converter, e.g., a trans-impedance amplifier (TIA) 108, 208, arranged between the output line 103, 203 corresponding to the column and the respective ADC 104, 204. In this regard, the TIA 108, 208 may receive the accumulated signal IA1, IA2from the output line 103, 203, may convert the accumulated signal IA1, IA2to generate a voltage signal, and to feed the voltage signal to the ADC 104, 204 for quantization. Moreover, the crossbar structure 200 may comprise a digital computing circuitry (DCC) 206 connected to the ADCs 104, 204. The DCC 206 may receive the digital signals or the ADC outputs 105, 205 from the ADCs 104, 204, and may perform a digital operation, e.g., a summation operation, a shift operation, or a combination thereof, on the digital signals 105, 205 to generate an output signal 207. The output signal 207 may be used for further processing. As such, the crossbar structure 200 can be understood as an extension of the crossbar array structure of Fig.1, especially an extension of the column 100 by means of the additional column 201. Accordingly, the weights W1-W4 may be distributed or split over the extended crossbar structure 200, exemplarily evenly distributed or split over the columns, especially to reduce the amount of noise variation accumulated per column. Additionally, due to the splitting of the weights W1-W4, an ADC with a lower resolution per column can be used for quantization, thereby reducing the power consumption of the ADCs 104, 204. In Fig.3, a second exemplary embodiment of the crossbar structure 300 is illustrated according to this disclosure. The crossbar structure 300 may comprise four input lines 1011-1014, and three output lines 103, 203, 303, where the input lines 1011-1014 and the output lines 103, 203, 303 may be arranged in a crossed pattern. The crossbar structure 300 may further comprise twelve cross-point devices 1021-1024, 2021-2024, 3021-3024, arranged in four rows corresponding to the input lines 1011-1014 and in three columns corresponding to the output lines 103, 203, 303, i.e., each column 301 may comprise four cross-point devices of the twelve cross-point devices 1021-1024, 2021-2024, 3021-3024. Accordingly, each cross-point device 1021-1024, 2021-2024, 3021-3024 may be arranged at a cross-point of one input line and one output line and may be connected to said input line and said output line, and may receive a respective input signal via said input line. Furthermore, each cross-point device 1021-1024, 2021-2024, 3021-3024 may provide a respective weight W1-W4 for multiplication with the respective input signal. For example, along the output line 103, the cross-point device 1021 may provide a weight W1 in terms of a conductance level or value, exemplarily a non-zero weight, may receive an input voltage signal via the input line 1011, and may further provide an output current signal I11at the output line 103 by means of the multiplication of the input voltage signal and the conductance value. In this regard, the weight value may be represented by N-bits, and may be mapped to 2N conductance levels, e.g., between 0 and Gmax. Similarly, for example, along the output line 203, the cross-point device 2021 may provide a weight in terms of a conductance level or value, exemplarily a zero weight, may receive the input voltage signal via the input line 1011, and may further provide an output current signal I21at the output line 203 by means of the multiplication of the input voltage signal and the conductance value. Due to the zero weight, the output current signal I21 may correspond to a null or zero current. Similarly, for example, along the output line 303, the cross-point device 3021 may provide a weight in terms of a conductance level or value, exemplarily a zero weight, may receive the input voltage signal via the input line 1011, and may further provide an output current signal I31 at the output line 303 by means of the multiplication of the input voltage signal and the conductance value. Due to the zero weight, the output current signal I31may correspond to a null or zero current. As such, each of the cross-point devices 1021-1024, 2021-2024, 3021-3024may be configured such that each of the cross-point devices 1021-1024, 2021-2024, 3021-3024may provide either a non-zero weight or a zero weight. Additionally or alternatively, each of the cross-point devices 1021-1024, 2021-2024, 3021-3024 may be configured such that each column may comprise an equal number of non-zero weights or zero weights or an unequal number of non-zero weights or zero weights. Furthermore, each column of the crossbar structure 300 may comprise an ADC 104, 204, 304 connected to the output line 103, 203, 303 corresponding to the column. The ADC 104, 204, 304 may receive the accumulated signal IA1, IA2, IA3 corresponding to an accumulation of the respective output signals, e.g., the accumulated signal IA1 corresponding to an accumulation of the output signals I11-I14, the accumulated signal IA2corresponding to an accumulation of the output signals I21-I24, and the accumulated signal IA3 corresponding to an accumulation of the output signals I31-I34. The ADC 104, 204, 304 may further quantize the accumulated signal IA1, IA2, IA3or a further processed version thereof, e.g., by a TIA, to generate a digital signal 105, 205, 305. In this regard, the respective ADC 104, 204, 304 may have a resolution corresponding to the number of non-zero weights in the respective column. For example, the cross-point devices 1021-1024 along the output line 103 may provide two non-zero weights W1, W2, and thus the ADC 104 may have a resolution of 2 bits in case both weights and inputs represent binary values. Furthermore, the cross-point devices 2021-2024along the output line 203 may provide one non-zero weight W3, and thus the ADC 204 may have a resolution of 1 bit in case both weights and inputs represent binary values. Moreover, the cross-point devices 3021-3024 along the output line 303 may provide one non-zero weight W4, and thus the ADC 304 may also have a resolution of 1 bit in case both weights and inputs represent binary values. Although not shown in this example, each column of the crossbar structure 300 may comprise a current-to-voltage converter, e.g., a trans-impedance amplifier (TIA), between the output line 103, 203, 303 corresponding to the column and the respective ADC 104, 204, 304. In this regard, the TIA may receive the accumulated signal IA1, IA2, IA3from the output line 103, 203, 303, may convert the accumulated signal IA1, IA2, IA3to generate a voltage signal, and to feed the voltage signal to the ADC 104, 204, 304 for quantization. Moreover, the crossbar structure 300 may comprise a digital computing circuitry (DCC) 206 connected to the ADCs 104, 204, 304. The DCC 206 may receive the digital signals or the ADC outputs 105, 205, 305 from the ADCs 104, 204, and may perform a digital operation, e.g., a summation operation, a shift operation, or a combination thereof, on the digital signals 105, 205, 305 to generate an output signal 309. For example, the DCC 206 may comprise a first summing operator 306 that may add the digital signals 205, 305 from the ADCs 204, 304, respectively, and may generate a first signal 307. The DCC 206 may further comprise a second summing operator 308 that may add the digital signal 105 from the ADC 104 and the first signal 307, and may generate the output signal 309. The output signal 309 may be used for further processing. As such, the crossbar structure 300 can be understood as an extension of the crossbar structure 200, especially an extension of the column 201 by means of the additional column 301, e.g., due to a higher significance of the column 201. For example, the significance can be instantiated by applying a multiplication, e.g., a multiplication by 2, i.e., a shift to the left by 1 bit, to signal 307 before it may be added to the signal 105 by the adder 308. Accordingly, the weights W1- W4 may be distributed or split over the extended crossbar structure 300 to reduce the amount of noise variation accumulated per column. Additionally, due to the splitting of the weights W1- W4, an ADC with a lower resolution per column can be used for quantization, thereby reducing the power consumption of the ADCs 104, 204, 304. In Fig. 4, a first exemplary embodiment of the cross-point device 102, 202, 302 is illustrated according to this disclosure. Herein, the cross-point devices 1021-1024 are generally referred as the cross-point device 102, the cross-point devices 2021-2024 are generally referred as the cross- point device 202, and the cross-point devices 3021-3024are generally referred as the cross-point device 302. In this regard, the cross-point device 102, 202, 302 may comprise a programmable conductance 401, for example, realized by means of a RRAM cell or device, to realize the weight W1-W4. In other words, a bit state of the RRAM cell may be defined as a higher or lower resistance. Accordingly, the weight value W may be represented by N-bits, and may be mapped to 2N conductance levels, e.g., between 0 and Gmax. In Fig.5, a second exemplary embodiment of the cross-point device 102, 202, 302 is illustrated according to this disclosure. The cross-point device 102, 202, 302 may comprise a SRAM cell or device (not shown), and a switching arrangement 501 comprising a first switch 5011 and a second switch 5012operably coupled to the SRAM cell. The cross-point device 102, 202, 302 may further comprise a resistor bank 502 comprising a first resistor 5021and a second resistor 5022 operably coupled to the switching arrangement 501. For example, the first resistor 5021may provide a conductance G and may be coupled to the first switch 5011, and the second resistor 5022 may provide a conductance 2G and may be coupled to the second switch 5012. The SRAM cell may digitally store the switching information, e.g., 0 for an off state and 1 for an on state, and may control the switches to realize the total conductance as follows: if the first switch 5011 is off and the second switch 5012 is off, a total of conductance 0 can be realized, if the first switch 5011 is off and the second switch 5012 is on, a total of conductance G can be realized, if the first switch 5011is on and the second switch 5012is off, a total of conductance 2G can be realized, and if the first switch 5011 is on and the second switch 5012 is on, a total of conductance 3G can be realized. In Fig.6, a third exemplary embodiment of the cross-point device 102, 202, 302 is illustrated according to this disclosure. The cross-point device 102, 202, 302 may comprise a SRAM cell or device (not shown), and a switch 601 operably coupled to the SRAM cell. The cross-point device 102, 202, 302 may further comprise a resistor 602 that may provide a conductance G and may be coupled to the switch 601. The SRAM cell may digitally store the switching information, e.g., 0 for an off state and 1 for an on state, and may control the switch 601 to realize a binary conductance as follows: if the switch 601 is off, a conductance of 0 can be realized, and if the switch 601 is on, a conductance of G can be realized. Alternatively, the binary conductance can also be realized by the cross-point device 102, 202, 302 of Fig. 4, especially by mapping only two conductance levels in the programmable conductance 401 via the RRAM cell. In Fig.7, a fourth exemplary embodiment of the cross-point device 102, 202, 302 is illustrated according to this disclosure. In particular, the cross-point device 102, 202, 302 may be arranged with the input line in a differential manner, especially a positive line to receive an absolute value of the input ^^,^^^and a negative line to receive a negative of the absolute value of the input ^^,^^^. The cross-point device 102, 202, 302 may further comprise a first resistor 701 coupled between the positive line receiving the absolute value of the input ^^,^^^and a common node 703, and a second resistor 702 coupled between the negative line receiving the absolute value of the input ^^,^^^and the common node 703. Furthermore, the first resistor 701 may be coupled to the common node 703 via a first coupling switch ^^,^^^and the second resistor 702 may be coupled to the common node 703 via a second coupling switch ^^,^^^. The cross-point device 102, 202, 302 may further comprise a SRAM cell or device that may store the switching information, e.g., 0 for an off state and 1 for an on state, and may control the coupling switches based on the signs of the input ^^and the weight W. Each of the first resistor 701 and the second resistor 702 may provide a conductance of G, especially equal to the absolute value of the weight W. Accordingly, four possible realizations of the signs of input ^^and the weight W can be listed to realize the desired signed multiplications ^^. ^ as follows: ^^,^^^^^,^^^Resulting current ^ Xi>0 W>0 on off^^,^^^ ⋅ ^ = ^^ ⋅ ^Xi>0 W<0 off on ^^,^^^⋅ ^ = −|^^| ⋅ |^| = ^^⋅ ^ Xi<0 W>0 off on ^^,^^^⋅ ^ = −|^^| ⋅ |^| = ^^⋅ ^ Xi<0 W<0 on off ^^, ^^^ ⋅ ^ =|^^|⋅|^|= ^^⋅ ^ In Fig.8, an exemplary MAC operation is illustrated according to this disclosure. Suppose the MAC operation on input ^ = (^^, ^^) and weight ^ = (^^, ^^) should be carried out and suppose further the weight entries are two-bit values such as: ^^= ^^^+ ^^^⋅ 2, for ^ = 1,2 (1) Then, the MAC operation can be expanded as: ^^^= ^^^^+ ^^^^= ^^(^^^+ ^^^⋅ 2)+ ^^(^^^+ ^^^⋅ 2)(2) This can be performed by: ^ firstly calculating the partial products ^^^^^, ^^^^^, ^^^^^, ^^^^^, especially in the analog domain, ^ secondly by partially sum the partial products ^^^^^, ^^^^^, ^^^^^, ^^^^^according to their significance, i.e., ^^= ^^^^^+ ^^^^^and ^^= ^^^^^+ ^^^^^, especially in the analog domain, and ^ finally by shifting the partial sums ^^and ^^according to their significance and by adding afterwards, especially in the digital domain. In this regard, the exemplary crossbar structure can be realized by arranging two input lines for the inputs ^^, ^^and two output lines for the partial sums ^^, ^^in the crossed pattern, where four cross-point devices can be respectively arranged at the cross-points to provide the partial products It is to be noted that the partial sum ^^is more significant than the partial sum ^^as it is needed to be multiplied or shifted by 2, e.g., via a multiplier 801 of the DCC, in order to digitally shift the partial sum ^^. Afterwards, the DCC may add the shifted partial sum ^^and the partial sum ^^, e.g., via a summing operator 802 of the DCC, to complete the MAC operation. It is further to be noted, since the partial sum ^^is more significant than the partial sum ^^, the right column can be expanded by a further column, such as the crossbar structure 200 or 300, especially to distribute the noise variation of the significant portion and to further reduce the required ADC resolution for quantization. In Fig.9, an exemplary embodiment of the method 900 according to the second aspect of this disclosure is illustrated. In a first step 901, an additional column comprising a plurality of additional cross-point devices corresponding to a column of a conventional IMC crossbar array, such as the column 100, is provided. In a second step 902, the cross-point devices of the column and the additional column are configured such that each of the column and the additional column comprises at least one cross-point device configured to provide a zero weight. It is important to note that, in the description as well as in the claims, the word “comprising” does not exclude other elements or steps and the indefinite article “a” or “an” does not exclude a plurality. A single element or other unit may fulfill the functions of several entities or items recited in the claims. Furthermore, the word “coupled or connected” implies that the elements may be directly coupled together or may be coupled through one or more intervening elements. Moreover, the disclosure with regard to any of the aspects is also relevant with regard to the other aspects of the disclosure. Although the disclosure has been illustrated and described with respect to one or more implementations, equivalent alterations and modifications will occur to others skilled in the art upon the reading and understanding of this specification and the annexed drawings. In addition, while a particular feature of this disclosure may have been disclosed with respect to only one of several implementations, such feature may be combined with one or more other features of the other implementations as may be desired and advantageous for any given or particular application.

Claims

CLAIMS 1. A crossbar structure (200, 300) for in-memory computing comprising: a plurality of input lines (1011-1014) and a plurality of output lines (103, 203, 303) arranged in a crossed pattern; a plurality of cross-point devices (1021-1024, 2021-2024, 3031-3034) arranged in a plurality of rows corresponding to the plurality of input lines (1011-1014) and in a plurality of columns corresponding to the plurality of output lines (103, 203, 303), wherein each cross-point device (1021-1024, 2021-2024, 3031-3034) is arranged at a cross-point of one input line and one output line and is connected to the input line and the output line; wherein each input line (1011-1014) is configured to receive a respective input signal (Xi), wherein each cross-point device (1021-1024, 2021-2024, 3031-3034) is configured to provide a weight (W1-W4) for multiplication with the respective input signal (Xi) received by the input line it is connected to, and wherein at least one column of the plurality of columns comprises at least one cross- point device (1021-1024, 2021-2024, 3031-3034) configured to provide a zero weight.

2. The crossbar structure according to claim 1, wherein each of the plurality of columns comprises an equal number of cross-point devices (1021-1024, 2021-2024, 3031-3034) configured to provide a non-zero weight.

3. The crossbar structure according to claim 1 or 2, wherein each cross-point device (1021-1024, 2021-2024, 3031-3034) is configured to provide a respective output signal (I) at the output line (103, 203, 303) it is connected to, wherein the respective output signal (I) corresponds to a multiplication of the respective input signal (Xi) received by the input line (1011-1014) it is connected to and the weight (W1-W4) provided by the cross-point device.

4. The crossbar structure according to claim 3, further comprising: a plurality of analog-to-digital converters (104, 204, 304), wherein a respective analog-to-digital converter (104, 204, 304) of the analog-to-digital converters is arranged in each column of the plurality of columns, the respective analog-to- digital converter is connected to the output line corresponding to the column, and the respective analog-to-digital converter is configured to:receive an accumulated signal (IA1, IA2, IA3) from the output line, wherein the accumulated signal (IA1, IA2, IA3) corresponds to an accumulation of the respective output signals (I) from the cross-point devices (1021-1024, 2021-2024, 3031-3034) arranged in the column, and quantize the accumulated signal (IA1, IA2, IA3) to generate a digital signal (105, 205, 305).

5. The crossbar structure according to claim 4, wherein the respective analog-to-digital converter (104, 204, 304) has a resolution determined by the number of cross-point devices (1021-1024, 2021-2024, 3031-3034) arranged in the column and configured to provide a non-zero weight.

6. The crossbar structure according to claim 4 or 5, further comprising: a digital computing circuitry (206) connected to a set of analog-to-digital converters (104, 204, 304) each arranged in a respective column of a set of columns of the plurality of columns, wherein the digital computing circuitry (206) is configured to: receive the digital signals (105, 205, 305) from the analog-to-digital converters (104, 204, 304) of the set of analog-to-digital converters, and perform a digital operation on the digital signals (105, 205, 305).

7. The crossbar structure according to any of claims 1 to 6, wherein: the input signal (Xi) corresponds to a signed input signal or an unsigned input signal, and the weights (W1-W4) correspond to signed weights or unsigned weights.

8. The crossbar structure according to any of claims 1 to 7, wherein each of the plurality of cross-point devices (1021-1024, 2021-2024, 3031-3034) comprises a resistive random-access memory device configured to generate a programmable conductance (401) to realize the weight (W1-W4).

9. The crossbar structure according to any of claims 1 to 7, wherein each of the plurality of cross-point devices (1021-1024, 2021-2024, 3031-3034) comprises a static random-access memory device, a switching arrangement (501), and a resistor bank (502) connected to the switching arrangement (501), the static random-access memory device being configured tocontrol the switching arrangement (501) to generate a programmable conductance via the resistor bank (502) to realize the weight (W1-W4).

10. The crossbar structure according to any of claims 1 to 7, wherein each of the plurality of cross-point devices (1021-1024, 2021-2024, 3031-3034) comprises a static random-access memory device, a switch (601), and a resistor (602) connected to the switch (601), the static random-access memory device being configured to control the switch (601) to generate a programmable binary conductance via the resistor (602) to realize the weight (W1-W4).

11. A method (900) for optimizing a crossbar array structure for in-memory computing, the crossbar array structure comprising: a plurality of cross-point devices (1021-1024) arranged in a plurality of rows, corresponding to a plurality of input lines (1011-1014) and each for receiving a respective input signal (Xi), and arranged in a column (100) corresponding to an output line (103) for providing an output signal (I), wherein each of the cross-point devices (1021-1024) is connected to one input line and the output line and is configured to provide a weight (W1-W4) to multiply with the respective input signal (Xi) received by the input line it is connected to, and the method comprising: providing (901) an additional column (201) comprising a plurality of additional cross- point devices (2021-2024), and configuring (902) the cross-point devices (1021-1024, 2021-2024) of the column (100) and the additional column (201) such that at least one of the column (100) and the additional column (201) comprises at least one cross-point device (1021-1024, 2021-2024) configured to provide a zero weight.

12. The method according to claim 11, further comprising configuring the cross-point devices (1021-1024, 2021-2024) of the column (100) and the additional column (201) such that each of the column (100) and the additional column (201) comprises an equal number of cross-point devices (1021-1024, 2021-2024) configured to provide a non-zero weight.

13. The method according to claim 11 or 12, further comprising: receiving an accumulated signal (IA1, IA2) from the output line of each of the column (100) and the additional column (201), wherein the accumulated signal (IA1, IA2) correspondsto an accumulation of the output signals (I) from the cross-point devices (1021-1024, 2021-2024) arranged in the column (100) and the additional column (201), setting a resolution for an analog-to-digital conversion of the accumulated signal (IA1, IA2) from each of the column (100) and the additional column (201), wherein the resolution is determined by a total number of non-zero weights in the column (100) and the additional column (201), and quantizing the accumulated signal (IA1, IA2) from each of the column (100) and the additional column (201) to generate a respective digital signal (105, 205).

14. The method according to claim 13, further comprising performing a digital operation on the digital signals (105, 205) from each of the column (100) and the additional column (201).

Citation Information

Patent Citations

  • Electronic circuit for calculating weighted sums

    EP3992862A1

  • Gate based resistance control units

    US20220366229A1