In-memory computing device for performing signed MAC operations
By introducing multiple memory cell groups and row activation circuits into the memory computing device, the symbol and absolute value separation of the input value is solved, and the problem of long signed multiplication and accumulation calculation time in the prior art is improved. Efficiency and precision.
Patent Information
- Application Number
- CN202410720516.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-06-03
- Filing Date
- 2024-06-05
- Publication Date
- 2024-12-06
AI Technical Summary
In the prior art, in-memory computing devices that perform signed multiplication and accumulation operations require a long refinement time, resulting in a small number of operations that can be executed per second.
A in-memory computing device is designed to provide positive and negative activations by introducing multiple memory cell groups into a computing memory array, each cell group containing a positive and negative cell subset, and using a row activation circuit to provide positive and negative activations. Signals enable separation of symbols and absolute values of input values, thereby shortening the refinement time.
Improve the number of signed multiplication and accumulated operations that can be performed per second, simplifies the design of digitizers, and improves calculation accuracy and efficiency.
Smart Images

Figure CN119088341A_ABST
Abstract
Description
[0001] Priority declaration
[0002] This application claims the benefit of priority to Italian Patent Application No. 102023000011370, filed on June 5, 2023, and U.S. Patent Application No. 18 / 731,676, filed on June 3, 2024, the contents of which are incorporated herein by reference in their entirety to the maximum extent permitted by law. Technical Field
[0003] Embodiments herein relate to an in-memory computing (IMC) device for performing signed multiply and accumulate (MAC) operations. In addition, embodiments herein also relate to a corresponding method for controlling an IMC device. Background Art
[0004] As is known, in-memory computing (IMC) devices use a specific arrangement of memory cells of a memory array to perform analog elaboration of data. For example, IMC devices are used to perform multiply and accumulate (MAC) operations, which are used, for example, to implement automatic learning algorithms, such as neural networks.
[0005] The MAC operation provides the output vector Y=y1,...,y M , as the input vector X = x1, ..., x N The result of multiplying the weight vector or matrix G, for example:
[0006] Right now:
[0007]
[0008] The IMC device will calculate the weight g ij Stored in cells of memory, and perform multiplication and addition operations at the cell level.
[0009] In detail, for each value y of the output vector Y i , known IMC devices generate currents indicative of corresponding MAC operations, i.e., And includes a readout circuit having a corresponding analog-to-digital converter (ADC) for discretizing the current.
[0010] The IMC device allows avoiding the transfer of data back and forth between the memory and the elaboration (eg, processing) unit. Therefore, the performance of the IMC device is not limited by the bandwidth of data transfer between the memory and the elaboration unit and has low energy consumption.
[0011] It is known that a signed MAC operation needs to be performed (ie, where the input values x1, ..., x N and / or calculate the weight g ijTherefore, according to the specific combination of input values and calculation weights, the output values y1, ..., y M Can also be positive or negative.
[0012] It is recognized that known methods for performing signed MAC operations require long elaboration times. Therefore, the number of signed MAC operations that can be performed per second is small.
[0013] There is a need in the art to overcome the above-mentioned shortcomings. Summary of the invention
[0014] In an embodiment, an in-memory computing (IMC) device for performing a multiplication and accumulation (MAC) operation is configured to receive an input signal (X) indicating a plurality of input values, each input value having a corresponding sign and a corresponding absolute value, and to provide at least one output signal. The device includes: a computation memory array including a plurality of memory cell groups coupleable to a bit line, each memory cell group configured to store a corresponding computation weight having a sign and an absolute value, each memory cell group including a positive cell subset coupleable to a corresponding positive word line and a negative cell subset coupleable to a corresponding negative word line; and a row activation circuit configured to receive the input signal and provide a positive activation signal having a positive activation duration to the positive word line of the corresponding memory cell group and a negative activation signal having a negative activation duration to the negative word line of the corresponding memory cell group for each input value during a refinement interval having a refinement duration. Thus, for each input value, a linear combination of the corresponding positive activation duration and the corresponding negative activation duration is equal to the refinement duration, and a difference between the corresponding positive activation duration and the corresponding negative activation duration is a function of the sign and the absolute value of the corresponding input value.
[0015] A positive cell subset of a corresponding memory cell group is configured to be activated by a corresponding positive activation signal and is programmable to have a corresponding positive cell current flow through it when activated. A negative cell subset of a corresponding memory cell group is configured to be activated by a corresponding negative activation signal and is programmable to have a negative cell current flow through it when activated. For each memory cell group, a linear combination of the corresponding positive cell current and the corresponding negative cell current is equal to the total current, and a difference between the corresponding positive cell current and the corresponding negative cell current is a function of the sign and absolute value of the corresponding calculation weight. The bit line is configured to have a bit line current flow through it during a refinement interval, and the bit line current is a function of the sum of the positive cell current and the negative cell current of the memory cell group coupled to the bit line. The in-memory computing device also includes a column refinement circuit that can be coupled to the bit line and is configured to sample the bit line current and provide at least one output signal in response.
[0016] In an embodiment, an in-memory computing (IMC) device performs a multiplication and accumulation (MAC) operation. The IMC device includes a computation memory array, the computation memory array including a plurality of memory cell groups that can be coupled to a bit line, each memory cell group being configured to store a corresponding computation weight having a sign and an absolute value, each memory cell group including a positive cell subset that can be coupled to a corresponding positive word line and configured to be activated by a corresponding positive activation signal, and a negative cell subset that can be coupled to a corresponding negative word line and configured to be activated by a corresponding negative activation signal. A method for controlling the IMC device includes: for each memory cell group, programming the corresponding positive cell subset to cause a corresponding positive cell current to flow in response to activation of the positive cell subset, and programming the corresponding negative cell subset to cause a corresponding negative cell current to flow in response to activation of the negative cell subset, so that for each memory cell group, a linear combination of the corresponding positive cell current and the corresponding negative cell current is equal to the total current, and a difference between the corresponding positive cell current and the corresponding negative cell current is a function of the sign and the absolute value of the corresponding computation weight. The method also includes receiving an input signal from a row activation circuit indicative of a plurality of input values, each input value having a respective sign and a respective absolute value, and in response, during a refinement interval having a refinement duration, for each input value, providing a positive activation signal having a positive activation duration to a positive word line of a corresponding memory cell group and providing a negative activation signal having a negative activation duration to a negative word line of the corresponding memory cell group. Thus, for each input value, a linear combination of the respective positive activation duration and the respective negative activation duration is equal to the refinement duration, and a difference between the respective positive activation duration and the respective negative activation duration is a function of the sign and the absolute value of the respective input value. The method also includes coupling the memory cell group to the bit line such that during the refinement interval, a bit line current flows through the bit line, the bit line current being a function of a sum of a positive cell current and a negative cell current of the memory cell group coupled to the bit line. The method also includes processing by a column refinement circuit coupled to the bit line, the column refinement circuit sampling the bit line current and providing at least one output signal in response. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] For a better understanding of the present invention, embodiments of the present invention will now be described by way of non-limiting examples with reference to the accompanying drawings, in which:
[0018] Figure 1 A block diagram of an in-memory computing device is illustrated;
[0019] Figure 2 Shows Figure 1 A detailed circuit diagram of a memory cell group of the device;
[0020] Figure 3Shows Figure 1 A block diagram of a row activation circuit of a device;
[0021] Figure 4 Shows Figure 3 A detailed circuit diagram of a portion of the circuit;
[0022] Figure 5 Shown by Figure 4 An exemplary waveform of a row activation signal generated by the circuit of FIG. 1 when in use;
[0023] Figure 6 Shown when used with Figure 1 an exemplary table of possible charge values associated with each memory cell group of a device;
[0024] Figure 7 Shows Figure 1 Circuit diagram of the reference circuit of the equipment;
[0025] Figure 8 Shows Figure 1 Exemplary waveforms of the device in use;
[0026] Fig. 9 Shows Figure 1 A detailed circuit diagram of a memory cell group of the device;
[0027] Fig.10 Shows Figure 1 A circuit diagram of a bias circuit of the device;
[0028] Fig.11 and Fig.12 Shows Figure 1 Detailed circuit diagram of the digital detector;
[0029] Fig.13 Shows Fig.11 An exemplary waveform of a digital detector when in use;
[0030] Fig.14 and Fig.15 Shows Figure 3 A detailed circuit diagram of the timer of the circuit; and
[0031] Fig.16 Shown with Figure 1 An exemplary table of possible charge values associated with each memory cell group when the device is in use. DETAILED DESCRIPTION
[0032] Figure 1An in-memory computing (IMC) device 10 is shown, which includes a computing memory array (for simplicity, hereinafter referred to as just a "memory array" or a "computing array") 12, a row activation circuit 14, and an output refinement circuit, where the output refinement circuit includes a plurality of digital detectors 16 and a digital signal processor (DSP) 17.
[0033] The memory array 12 includes a plurality of memory cells 20 organized according to a matrix arrangement having M columns and K rows.
[0034] Next, a common memory cell in the plurality of memory cells 20 is denoted by reference numeral 20. i,k , M and k=1, . . . , K indicate columns and rows of universal memory cells in the memory array 12 , respectively.
[0035] In this embodiment, the memory cell 20 is of a non-volatile type.
[0036] The memory cells 20 are further organized to form a plurality of memory cell groups (in Figure 1 Each memory cell group is configured to store a corresponding calculated weight G having a sign and an absolute value. ij .
[0037] Calculate the weight G ij Can be used as weights to perform in-memory computations, particularly signed multiply and accumulate (MAC) operations.
[0038] In this embodiment, each memory cell group 22 is formed of a corresponding positive cell 23A and a corresponding negative cell 23B belonging to the plurality of memory cells 20 .
[0039] exist Figure 1 In the arrangement, the positive memory cells 23A and the negative memory cells 23B of each memory cell group 22 are arranged on the same column i and on two adjacent rows k, k+1 of the memory array 12.
[0040] In practice, in this embodiment, the memory cell groups 22 form an array having M columns and N rows, where N=K / 2.
[0041] In the following, a common cell group in the plurality of memory cell groups 22 is denoted by reference numeral 22. ij , where indexes i=1, ..., M and j=1, ..., N respectively indicate the general memory cell group 22 ij Columns and rows.
[0042] The memory cells 20 arranged in the same column of the memory array 12 are connected to each other via corresponding bit lines BLi , where i=1, ..., M. The memory cells 20 arranged in the same row k of the memory array 12 are connected together through corresponding word lines WL k Connect together, where k=1,…,K.
[0043] In detail, each memory cell group 22 i,j coupled to the positive word lines WL hereinafter respectively identified as j,+ and negative word line WL j,- Two adjacent word lines WL k , WL k+1 .
[0044] refer to Figure 1 The layout of the positive word line WL j,+ coupled to the memory cell group 22 1,j ,...,twenty two M,j The positive cell 23A, and the negative word line WL j,- coupled to the memory cell group 22 1,j ,...,twenty two M,j Negative cell 23B.
[0045] For example, the positive word line WL 1,+ coupled to all memory cells 20 arranged in row k=1 and columns i=1 to i=M in the memory array 12; negative word line WL 1,- Coupled to all memory cells 20 arranged in row k=2 and columns i=1 to i=M in the memory array 12 .
[0046] In practice, each memory cell group 22 i,j Coupled to the corresponding bit line BL i and two corresponding word lines WL j,+ , WL j,- .
[0047] The row activation unit 14 receives a plurality of input values x1, ..., x N The input vector X. The input values x1, ..., x N Each represents an input coefficient with sign and absolute value.
[0048] Input values x1, ..., x N Can be the input data of MAC operation.
[0049] The row activation unit 14 is for each input value x j (where j=1, ..., N) is the corresponding positive word line WL j,+ Provide positive activation signal S j,+ and the corresponding negative word line WL j,-Provide negative activation signal S j,- .
[0050] In practice, the row activation unit 14 sends a signal to the memory cell group 22. 1,j ,...,twenty two M,j The positive cell 23A provides a positive activation signal S j,+ , and to the memory cell group 22 1,j ,...,twenty two M,j The negative cell 23B provides a negative activation signal S j,- .
[0051] The row activation unit 14 is in the refinement interval T C Provide positive activation signal S during j,+ and negative activation signal S j,- .
[0052] Positive activation signal S j,+ In the refinement interval T C Activate the corresponding positive cell 23A for a positive activation duration T j,+ .
[0053] Negative activation signal S j,- In the refinement interval T C Activate the corresponding negative cell 23B for a negative activation duration T j,- .
[0054] Positive activation duration T j,+ and negative activation duration T j,- is the corresponding input value x j is a function of the sign and absolute value of , as described in detail below.
[0055] In detail, the positive activation signal S j,+ and negative activation signal S j,- It is a pulse, especially a rectangular pulse, the width (duration) of each pulse is equal to T j,+ and T j,- .
[0056] In this embodiment, the IMC device 10 further includes a bias circuit 18 that generates a bias voltage Vr and provides the bias voltage Vr to the bit lines BL1, . . . , BL M .
[0057] In the remaining description, for simplicity, it is considered that the bias circuit 18 is provided for all bit lines BL1, . . . , BL M The same bias voltage Vr is provided. However, depending on the specific application, the bias circuit 18 may start from the bias voltage Vr to provide the bit lines BL1, . . . , BL M Bias voltages different from each other are provided.
[0058] The digital detectors 16 are analog-to-digital converters (ADCs), each of which is coupled to a corresponding bit line BL i And each ADC provides a corresponding charge signal q i , the charge signal q i Indicates the refinement interval T C During this period, the corresponding bit line BL i The current I BL,i .
[0059] In detail, each digital detector 16 can provide a bit current I BL,i In the refinement interval T C The corresponding charge signal q within the start of the integration i ; That is, signal q i Indicates the refinement interval T C During the period, the bit line BL i The amount of charge flowing in the.
[0060] For example, the digital detector 16 may be formed by an integrating and sampling circuit known per se.
[0061] Reference below Figures 11 to 13 A specific embodiment of the digital detector 16 is described in detail.
[0062] The digital processor 17 processes the charge signals q1, ..., q M , and in response provides an output signal Y MAC , the output signal Y MAC Indicates the input vector X and the calculated weight G ij The result of the MAC operation between them.
[0063] Output signal Y MAC Indicates multiple output values y1, ..., y M , where each bit line BL i An output value.
[0064] As described in detail below, the digital processor 17 can convert the charge signal q i With the reference charge signal q ref compares and provides a corresponding output value y in response to the comparison i The output value y i The sign and absolute value of the corresponding charge signal q i With the reference charge signal q ref Comparison between.
[0065] Each memory cell 20 includes a storage element 25 and a selection element 26 .
[0066] 20 per memory cell i,k The storage element 25 is a variable resistance element which can be programmed to have a transconductance g ik , the transconductance g ik During programming of the memory array 12, the corresponding memory cell group 22 is selected i,j The calculation weight G ij function, as discussed in detail below.
[0067] In particular, the memory element 25 may be based on a phase change material (PCM), such as a chalcogenide.
[0068] Phase change materials have at least two phases, such as an amorphous phase and a crystalline phase, where each phase may have one or more different resistivity values.
[0069] Phase change materials can be transformed from one phase to another via heat transfer (eg, using a pulse of electric current).
[0070] The resistance of each storage element 25 associated with the corresponding phase state is used to distinguish between two or more logic states of the corresponding memory cell 20 .
[0071] For example, an amorphous phase may have a higher resistance (and therefore a lower transconductance) than a crystalline phase.
[0072] For example, a logic state “0” or a reset state may be associated with an amorphous phase of storage element 25 . A logic state “1” or a set state may be associated with a crystalline phase of storage element 25 .
[0073] In this embodiment, memory cells 20 are multi-level cells; that is, they can be programmed to represent a number of logical states greater than two.
[0074] In other words, each memory cell 20 i,k can be programmed so that the corresponding transconductance g i,k Can take a value selected from three or more values.
[0075] Purely as an example, in the following, consider each memory cell 20 i,k The transconductance g i,k can be selected to be one of three different transconductance values g1, g2 and g3, where g1 <g2<g3。
[0076] The storage element 25 has a corresponding bit line BL coupled to it. i A first terminal of a node 28 of the reference potential node 20 and a second terminal coupled to a reference potential node (here ground 29) via a selection element 26.
[0077] The selection element 26 is a switch, such as a BJT, a diode or a MOS transistor (here, an NMOS transistor), which is arranged in series with the corresponding storage element 25 and whose switching is controlled by the corresponding word line WL j,+ or WL j,- The activation signal S j,+ or S j,- control.
[0078] In this embodiment, the NMOS transistor forming the select element 26 has a source terminal coupled (here directly connected) to ground 29; a drain terminal coupled (here directly connected) to the second terminal of the storage element 25; and a drain terminal coupled (here directly connected) to the corresponding word line WL j,+ or WL j,- of the gate terminal.
[0079] In practice, the storage element 25 and the selection element 26 form a current path of the corresponding memory cell 20 .
[0080] refer to Figure 2 The general memory cell group 22 is shown in detail in i,j , the selection element 26 of the positive cell 23A responds to the corresponding positive activation signal S j,+ The corresponding current path is closed by receiving the positive cell current I + i,j can flow from the common node 28 to the ground 29 for a time interval that depends on the positive activation duration T j,+ , in particular, here is equal to the positive activation duration T j,+ .
[0081] The selection element 26 of the negative cell 23B responds to the corresponding negative activation signal S j,- The corresponding current path is closed by receiving the negative cell current I- i,j can flow from the common node 28 to the ground 29 for a time interval that depends on the negative activation duration T j,- , in particular, here is equal to the negative activation duration T j,- .
[0082] Universal memory cell group 22 i,j The positive cell 23A and the negative cell 23B are programmed so that the positive cell current I + i,j and negative cell current I - i,j The linear combination of the total current I tot .
[0083] In particular, in this embodiment, I + i,j +I- i,j =I tot .
[0084] The user may select the total current I according to the specific application of the IMC device 10 and / or the specific characteristics of the memory array 12. tot .
[0085] In this embodiment, for all memory cell groups 22, the total current I tot are all the same.
[0086] Universal memory cell group 22 i,j The positive cell 23A and the negative cell 23B are further programmed so that the positive cell current I + i,j With negative cell current I - i,j The difference between them is the corresponding calculation weight G i,j A function of the sign and absolute value of .
[0087] In detail, it can be based on the total current I tot The reference current I ref To select the positive cell current I + i,j and negative cell current I - i,j .
[0088] In this embodiment, the reference current I ref Equal to the total current I tot half of I ref =I tot / 2.
[0089] In more detail, the positive cell current I + i,j and negative cell current I - i,j, so that the positive cell current I + i,j With reference current I ref The difference between them is the corresponding calculation weight G i,j A function of the sign and absolute value of .
[0090] Positive cell current I + i,j With reference current I ref The sign of the difference between them depends on the calculated weight G i,j symbol.
[0091] Positive cell current I + i,j With reference current I refThe absolute value of the difference between them depends on the calculated weight G i,j The absolute value of .
[0092] In this example:
[0093] - If G i,j >0, then I + i,j >I ref ;
[0094] - If G i,j =0, then I + i,j =I ref ;as well as
[0095] - If G i,j <0, then I + i,j ref .
[0096] Positive cell current I + i,j With reference current I ref The absolute value of the difference between |I + i,j -I ref | is the calculation weight G i,j The absolute value of |G j,j |function.
[0097] Absolute value |I + i,j -I ref | can be used to calculate the weight G i,j The absolute value of |G j,j |, such as an increasing monotonic function.
[0098] Just by way of explanatory numerical example, assuming that the current I + i,j ,I - i,j can take values equal to 0μA, 1μA or 2μA respectively, that is, I tot =2μA and I ref =I tot / 2=1μA, then calculate the weight G i,j It can be expressed as:
[0099] - If G i,j = +1, then I + i,j =2μA and I - i,j =0μA;
[0100] - If Gi,j =0, then I + i,j =1μA and I - i,j = 1 μA; and
[0101] - If G i,j = -1, then I + i,j =0μA and I - i,j =2μA.
[0102] In practice, for each memory cell group 22 i,j , the corresponding positive cell current I + i,j With the corresponding negative cell current I - i,j The difference between + i,j -I - i,j The symbol can be the corresponding calculation weight G ij is a function of the sign of , not its absolute value. In other words, if G ij >0, then I + i,j -I - i,j >0, if G ij <0, then I + i,j -I - i,j < 0, and if G ij =0, then I + i,j -I - i,j =0.
[0103] The corresponding positive cell current I + i,j With the corresponding negative cell current I - i,j The difference between + i,j -I - i,j The absolute value of can be the corresponding calculation weight G ij is a function of the absolute value of , not its sign.
[0104] Specifically, in this embodiment, for each memory cell group 22 i,j , the corresponding positive cell current I + i,j With the corresponding negative cell current I- i,j The difference between + i,j -I - i,j is the corresponding calculation weight G ij A monotonic function of , especially an increasing monotonic function.
[0105] The IMC device 10 may also include an interface circuit 30 including row decoding and selection circuits, column decoding and selection circuits, and read / write circuits known per se for operation of the IMC device 10. For example, the read / write circuits may be used to program the value of the transconductance of the memory cell 20.
[0106] refer to Figure 3 , the row activation unit 14 includes a timer or global counter 45 and a plurality of input-to-time converters 46, each input value x1, ..., x N There is an input to the time converter.
[0107] The timer 45 provides a timing signal TM, which is defined at the start time t start and the end time t end The refinement interval T C (For example, Figure 5 and Figure 8 ).
[0108] Refinement interval T C With a refined duration (still T for simplicity) C Indication), such as about several hundred nanoseconds or less than approximately 100ns, which can be selected by the user according to the specific application.
[0109] The row activation unit 14 receives (eg, from an external user of the IMC device 10) a determined start time t start And thus determine the refinement interval T C For example, the calculation start signal START_CNT can be a step logic signal, and its rising edge determines the starting time t start ( Figure 8 ).
[0110] The timing signal TM may be an analog signal or a digital signal that increases with time. Here, it is an L-bit digital signal, which is also referred to as the timing signal TM in the following text and in the figure. <l:1>.
[0111] At the beginning of the new refinement operation, that is, at the starting time t start , the timer 45 can reset the timer signal TM to a starting value, for example, to zero.
[0112] The timer 45 increments a timing signal TM which increases with time, for example with a monotonic trend, in particular a strictly monotonic trend, which increases from the time t start Starting from the starting value at t, until time t end The ending value at .
[0113] Timer 45 updates at a frequency f u Incremental timing signal TM, the update frequency f u It depends on the specific implementation of the timer 45 and can be selected by the user of the IMC device 10 according to the specific application and the specific implementation of the timer 45.
[0114] In practice, the timer signal TM is a counter signal.
[0115] The row activation unit 14 further comprises a comparator 47 which compares the timing signal TM with a signal indicating a refinement interval T C Duration T C The maximum count signal MAX_CNT is compared and a count end (or calculation end) signal STOP_CNT, such as a step logic signal, is provided in response.
[0116] When the timing signal TM becomes equal to the maximum count signal MAX_CNT, the count end signal STOP_CNT switches, in particular switches to a high logic value, thereby determining the end time t end ( Figure 8 ).
[0117] The input to the time converter 46 each receives a corresponding input value x j And in the refinement interval T C The corresponding input value x is provided as j The positive activation signal S of the function j,+ and negative activation signal S j,- .
[0118] The positive activation signal S is input to the time converter 46. j,+ and negative activation signal S j,- , so that the positive activation duration T j,+ and negative activation duration T j,- The linear combination of is equal to the refinement duration T C .
[0119] In detail, in this embodiment, T j,+ +T j,- =T C .
[0120] In addition, the positive activation duration T is selected j,+ and negative activation duration T j,- , so that the positive activation duration T j,+ With negative activation duration T j,- The difference between the corresponding input values x j A function of the sign and absolute value of .
[0121] For example, for each input value x j , difference T j,+ -T j,- The sign of can be a function of the sign and / or absolute value of the corresponding input value.
[0122] Difference T j,+ -T j,- The absolute value of may be a function of the sign and / or absolute value of the corresponding input value.
[0123] Difference T j,+ -T j,- Can be the corresponding input value x j A monotonic function of .
[0124] In detail, it can be based on a refinement duration T less than C The reference duration T ref To select the positive activation duration T j,+ and negative activation duration T j,- .
[0125] In this embodiment, the reference duration T ref Equal to the refinement duration T C half of ; that is, T ref =T C / 2.
[0126] In detail, the positive activation signal S is input to the time converter 46. j,+ , so that the positive refinement duration T j,+ With reference duration T C The difference between / 2 is the input value x j A function of the sign and absolute value of .
[0127] Positive refinement duration T j,+ With reference duration T C The sign of the difference between x and x2 depends on the input value x. j The sign of the positive refinement duration T j,+ With reference duration T C The absolute value of the difference between / 2 depends on the input value x j The absolute value of .
[0128] In the following, reference Figure 4 The embodiment of the general input to time converter 46 is described in detail.
[0129] The input to the time converter 46 includes a register 50 which receives an input value x j and provide an indicative input value x j The sign signal SGN indicates the sign (+ or -) of the input value x j The absolute value of the absolute value of the signal MOD.
[0130] The logic block 51 receives the sign signal SGN and the absolute value signal MOD and in response provides a signal WL_DATA (here of digital type) configured to set the positive activation duration T j,+ and negative activation duration T j,- Set to input value x j A function of the sign and absolute value of .
[0131] The signal WL_DATA indicates that the timing signal TM is at time t start and t end One of the values between .
[0132] In detail, the switching signal WL_DATA indicates that the switch is arranged in the refinement interval T C Within (i.e., starting time t start and the end time t end The switching time t s,j .
[0133] For example, in the following, for simplicity and without loss of generality, consider t start = 0. Therefore, t end =T C And T ref =T C / 2.
[0134] If the input value x j is equal to zero, then the signal WL_DATA indicates the intermediate time t ref =T C / 2, that is, t s,j =T C / 2.
[0135] If the input value x j is greater than zero, then the signal WL_DATA indicates the intermediate time t in this embodiment. ref =T C / 2, that is, t s,j >T C / 2.
[0136] If the input value x j is less than zero, then the signal WL_DATA indicates the intermediate time t in this embodiment. ref =T C / 2 before, that is, t s,j <T C / 2.
[0137] Switching time t s,j With reference time t ref The absolute value of the difference between |t s,j -t ref | is the input value x j The absolute value of |x j |function.
[0138] Absolute value |t s,j -t ref | can be the input value x j The absolute value of |x j | is a monotonically increasing function of.
[0139] Absolute value |t s,j -t ref | can be combined with the input value x j The absolute value of |x j |Linear or nonlinear dependence.
[0140] In this embodiment, the absolute value |t s,j -t ref |With input value x j The absolute value of |x j | is linearly dependent, in particular, proportional; that is, |t s,j -t ref |=p·|x j |.
[0141] Furthermore, in this embodiment, when the input value x j Positive value or input value x j When it is a negative value, |t s,j -t ref |with|x j | are all the same. However, if x j >0, then |t s,j -t ref |with|x j The relationship between | can have a proportionality factor p + , and if x j <0, then |t s,j -t ref |with|x j The relationship between | can have a proportionality factor p - , where p + With p - different.
[0142] Comparator 52 compares signal WL_DATA with timing signal TM and provides switching signal SWITCH_WL in response. Comparator 52 may be a bit checking circuit based on an XOR logic gate, for example.
[0143] For example, the switching signal SWITCH_WL may have a high logic value as long as the timing signal TM is different from the signal WL_DATA. s,j ), the comparator 52 switches the switching signal SWITCH_WL.
[0144] The latch 53 of the set-reset type receives the switching signal SWITCH_WL at the set input terminal S and the count end signal STOP_CNT at the reset input terminal R, and provides the row selection signal SEL_WL at the Q output terminal.
[0145] The set-reset type latch 54 receives the count start signal START_CNT at the set input terminal S and the count end signal STOP_CNT at the reset input terminal R, and provides a row enable signal EN_WL at the Q output terminal.
[0146] In practice, the row enable signal EN_WL can be set at the refinement interval T C Period (i.e., from the start time t start To the end time t end ) has a high logic value.
[0147] The first AND logic gate 55A receives the selection signal SEL_WL and the enable signal EN_WL at the input terminal, and is coupled to the first row driver 57A (eg, a buffer) at the output terminal, which outputs the selection signal SEL_WL to the corresponding word line WL. j,+ Provide positive activation signal S j,+ .
[0148] Inverter 56 receives the select signal SEL_WL and provides a negated select signal SEL_WL_N.
[0149] The second AND logic gate 55B receives the negated selection signal SEL_WL_N and the enable signal EN_WL at the input terminal, and is coupled to the second row driver 57B (eg, a buffer) at the output terminal, which supplies the corresponding word line WL j,- Provide negative activation signal S j,- .
[0150] In practice, Figure 4 In the embodiment, the positive activation signal S j,+ It is asserted that, that is, here from the starting time t start To the switching time t s,j (ie, in response to the signal WL_DATA becoming equal to the timing signal TM) has a high logic value. The negative activation signal S j,- is asserted, that is, here from the switching time t s,j To the end time t end Has a high logic value.
[0151] Purely as an example, Figure 5 shows the positive activation signal S associated with the input value x1 in the example of x1>0, x2<0 and |x2|<|x1| 1,+ and negative activation signal S 1,- and the positive activation signal S associated with the input value x2 2,+ and negative activation signal S 2,- . Figure 5 The dotted line indicates the time t ref .
[0152] Since x1>0, the switching time t s,1 Greater than the reference time t ref Therefore, the positive activation duration T 1,+ Greater than the reference duration T C / 2(T 1,+ -T C / 2>0).
[0153] In addition, due to T 1,+ +T 1,- =T C , so the positive activation duration T 1,+ Greater than negative activation duration T 1,- In practice, T 1,+ -T 1,- >0.
[0154] Since x2<0, the switching time t s,2 Less than the reference time t ref Therefore, the positive activation duration T 2,+ Less than the reference duration T C / 2(T 2,+ -T C / 2<0).
[0155] In addition, due to T 2,+ +T 2,- =T C , so the positive activation duration T 2,+ Less than negative activation duration T 2,- In practice, T 2,+ -T 2,- <0.
[0156] exist Figure 5 In the example, since |x2|<|x1|, |t s,2 -t ref |<|t s,1 -t ref |, so |T 2,+ -T C / 2|<|T 1,+ -T C / 2|.
[0157] In other words, the negative activation duration T 2,- Less than the positive activation duration T 1,+ .
[0158] In practice, in this embodiment, if the input value x j is equal to zero, then the corresponding input to the time converter 46 generates a positive activation signal S j,+ , so that the positive activation duration T j,+ Equal to the reference duration, i.e., here equal to half of the refinement time T C / 2. Therefore, the negative activation duration T j,- Also equal to the reference duration T C / 2.
[0159] If the input value x j is greater than zero, then the corresponding input to the time converter 46 generates a positive activation signal S j,+ , so that the positive activation duration T j,+ Greater than the reference duration T C / 2. Therefore, the negative activation duration T j,- Less than the reference duration T C / 2.
[0160] If the input value x j is less than zero, then the corresponding input to the time converter 46 generates a positive activation signal S j,+ , so that the positive activation duration T j,+ Less than the reference duration T C / 2. Therefore, the negative activation duration T j,- Greater than the reference duration T C / 2.
[0161] In other words, the positive activation duration T j,+ With negative activation duration T j,- The sign of the difference between the input value x j is a function of the sign of , not of its absolute value.
[0162] Positive activation duration T j,+ With negative activation duration T j,- The absolute value of the difference between j is a function of the absolute value, not the sign, as mentioned above with reference to the switching time t s,j With reference time t ref The absolute value of the difference between |t s,j -t ref |Described.
[0163] In other words, in this embodiment, the positive activation duration T j,+ With reference duration T C The difference between / 2 is the input value x j A monotonic function of , especially a monotonic increasing function.
[0164] For example, the positive activation duration T j,+ With negative activation duration T j,- The absolute value of the difference between the input value x j The absolute value of is linearly or nonlinearly dependent.
[0165] In detail, in this embodiment, the positive activation duration T j,+ With negative activation duration T j,- The absolute value of the difference between the input value x j The absolute value of j is proportional to the absolute value of .
[0166] Purely as an example, consider the refinement duration T C =128ns (therefore T C / 2=64ns) and considering the input value x j Can be included between -64 and +64:
[0167] -If x j = -64, then T j,+ = 0ns and T j,- =128ns, that is, T j,+ -T C / 2 = -64;
[0168] -If x j = -32, then T j,+ =32ns and T j,- =96ns, that is, T j,+ -T C / 2 = -32;
[0169] -If x j =0, then T j,+ = 64ns and T j,- = 64ns, that is, T j,+ -T C / 2=0;
[0170] -If x j = +32, then T j,+ =96ns and T j,- =32ns, that is, T j,+ -T C / 2=32;
[0171] as well as
[0172] -If x j = +64, then T j,+ = 128ns and T j,- = 0ns, that is, T j,+ -T C / 2=64.
[0173] In use, the IMC device 10 may be used to provide an output vector Y MAC =y1,...,y M , which indicates that the input vector X = x1, ..., x N and the weight G is calculated by ij The multiplication and accumulation (MAC) operation between the matrices formed.
[0174] As reference Figure 2 As discussed, the general memory cell group 22 i,j The positive cell 23A is in the refinement interval T C During this period, a cell current is absorbed, here from the corresponding node 28 to ground 29, which depends on the transconductance g of the corresponding storage element 25. i,k and the activation duration of the corresponding selection element 26, i.e., depending on the positive activation signal S j,+ The positive activation duration T j,+ .
[0175] In detail, in the refinement interval T C During this period, the current flowing through the positive cell 23A has a current equal to the positive cell current I + i,j The strength of the i,k ) and depends on the positive activation duration T j,+ The duration of time.
[0176] Universal memory cell group 22 i,j The negative cell 23B is in the refinement interval T C During this period, a cell current is absorbed, here from the corresponding node 28 to ground 29, which depends on the transconductance g of the corresponding storage element 25. i,k+1 and the activation duration of the corresponding selection element 26, i.e., depending on the negative activation signal S j,- The negative activation duration T j,- .
[0177] In detail, during the refinement interval TC, the current flowing through the negative cell 23B has a current equal to the negative cell current I - i,j The strength of the i,k+1 ) and depends on the negative activation duration T j,- The duration of time.
[0178] In this embodiment, the bias voltage Vr is applied to all bit lines BL1, . . . , BL M are the same; therefore, the current intensity I + i,j and I - i,j It depends only on the transconductance of the positive memory cell 23A and the negative memory cell 23B, respectively.
[0179] In the refinement interval T C During this time, reference memory cell group 22 i,j , positive cell 23A has a charge displacement Q + i,j Contributes to the charge displacement Q + i,j is the current intensity I + i,j With positive activation duration T j,+ The function of the product of; negative cell 23B for charge displacement Q - i,j Contributes to the charge displacement Q - i,j is the current intensity I - i,j With negative activation duration T j,- The function of the product of .
[0180] In practice, in the refinement interval T C During this period, each memory cell group 22 i,j Both charge displacement Q i,j =I + i,j ·T j,+ +I - i,j ·T j,- Have contributed.
[0181] Since the positive activation duration T j,+ and negative activation duration T j,- Depends on the corresponding input value x j , and the current intensity I + i,j ,I - i,j Depends on the calculation weight G ij , so with the memory cell group 22 i,j The associated charge Q i,j is the product G ij ·x j function.
[0182] Purely as an example, Figure 6 The table below provides an example where T C =128ns, input value x j can take values between -64 and +64, and the current intensity I + i,j ,I - i,j It can be 0μA, 1μA or 2μA.
[0183] As from Figure 6 As can be seen from the table, if T j,+ =T j,- =T ref =T C / 2 and / or I + i,j =I - i,j =I ref =I tot / 2, then the charge Q i,j In the refinement interval T C At the end, it has a reference value Q ref In the example shown, the reference value Q ref Equal to 128.
[0184] In practice, when the corresponding positive activation duration T j,+ Equal to the reference duration (here T C / 2), and / or when the positive cell current I + i,j Equal to the reference current (here I tot / 2), memory cell group 22 i,j The corresponding bit line BL i The middle value is equal to the reference value Q ref contributes to the charge displacement.
[0185] In other words, if x j =0 and / or G ij =0, then the memory cell group 22 i,j Equal to Q ref The charge displacement contribution.
[0186] If x j >0 and G i,j >0, or x j <0 and G i,j <0, then Q i,j >Q ref .
[0187] If x j >0 and G i,j <0, or x j <0 and G i,j >0, then Q i,j ref .
[0188] In practice, by placing the charge Q i,j With reference charge Q ref By comparison, we can determine the operation G ij ·x j symbol.
[0189] In addition, the charge Q i,j With reference charge Q ref The absolute value of the difference between the two indicates the operation G ij ·x j The absolute value of .
[0190] In fact, consider, for example, the first row of the table, where G i,j =1, as x j The absolute value of the charge Q i,j Take a higher value.
[0191] In detail, as the absolute value |x j |Increase, absolute value of difference|Q i,j -Q ref |Increase, especially here it increases in a proportional manner.
[0192] In addition, the charge Q i,j The same value as G ij and x j These combinations are associated with the operation G ij ·x j The same result corresponds to (for example, G ij = -1 and x j = -32, or G ij = +1 and x j =+32).
[0193] For each bit line BL i , the corresponding bit line current I BL,i By coupling to the bit line BL i The sum of the currents of all cells 20 gives . Therefore, in the refinement interval T C At the end, it flows into the bit line BL i The total charge Q tot,i Equal to the coupling to the bit line BL i All memory cells of group 22 i,1 ,...,twenty two i,N The charge contribution Q i,1 , ..., Q i,N The sum of; that is,
[0194] In practice, the total charge Q tot,i Indicative operation G i1 x1+G i2 ·x2+...+G iN ·x N .
[0195] Therefore, it is equal to the reference value Q associated with each memory cell group 22 ref The sum of the total reference charge Q tot,i,ref and bit line BL i In other words, Q tot,i,ref =N·Q ref , where N is the bit line BL coupled to i The number of memory cell groups 22.
[0196] DSP 17 converts the charge signal q i With the reference charge signal q ref For comparison, the reference charge signal q ref Indicates the total reference charge Q tot,i,ref .
[0197] In response to q i With q ref By comparing the MAC operation G i1 x1+G i2 ·x2+...+G iN ·x N The sign and absolute value of the result of the above operation are determined, and a corresponding output signal y indicating the result is provided at the output terminal. i .
[0198] In detail, the DSP 17 may determine as q i With q ref The output signal y of the function of the difference between i The sign and absolute value of .
[0199] Output signal y i The sign of the difference q i -q ref Starting from the previous discussion, in this embodiment, the following conclusions are drawn:
[0200] -If q i =q ref , then y i =0;
[0201] -If q i >q ref , then y i >0; and
[0202] -If q i ref , then y i <0.
[0203] Based on the input values x1, ..., x N Mapped to the corresponding activation duration T 1,+ , ..., T N,+ and T 1,- , ..., T N,- and for calculating the weight G i1 , ..., G iN Mapped to the corresponding current value I + i,1 , ..., i + i,N and I - i,1 , ..., i - i,N The specific conversion function, output signal y i The absolute value of the difference q i -q ref The absolute value of .
[0204] For example, the DSP 17 may use a specific conversion table previously stored in the IMC device 10 to convert the difference value q i -q ref The absolute value of the output signal y i The specific value of .
[0205] For each input value x j , refinement duration T C is the positive activation duration T j,+ and negative activation duration T j,- The linear combination of and positive activation duration T j,+ With negative activation duration T j,- The difference between the input value x j The sign and absolute value of the function facts, plus 22 for each memory cell group i,j , total current I tot is the positive cell current I + i,j and negative cell current I - i,j The linear combination of the positive cell current I + i,j With negative cell current I - i,j The difference between them is used to calculate the weight G ij The fact that the sign and absolute value of the function makes the refinement interval T C Ends with memory cell group 22 i,j The associated charge Q i,j Necessary indication operation G ij ·x j The sign and absolute value of .
[0206] In addition, the refinement interval T C End with bit line BL i The associated charge Q i,tot Indicative operation G i1 x1+G i2 ·x2+...+G iN ·x N The sign and absolute value of .
[0207] In practice, the IMC device 10 allows for a single refinement interval T C Get the instruction MAC operation The sign and absolute value of the result, where the positive word line WL 1,+ , ..., W.L. N,+ and negative word line WL 1,- , ..., W.L. N,- All are activated.
[0208] In other words, the instruction operation The sign and absolute value of the output signal y i can be obtained in a single MAC refinement operation of the IMC device 10, that is, only the starting value (at time t start ) and the end value (at time t end ) only needs to scan the timing signal TM once.
[0209] Therefore, with two different refinement intervals T C Compared with the solution that performs positive refinement and negative refinement in , the IMC device 10 can perform more signed MAC operations per second. The number of operations per second (TOPS) can be doubled.
[0210] In particular, Figure 4 The embodiment of the input to the time converter 46 enables each bit line BL i The current is distributed to the refined interval T C Two different subintervals (T j,+ and T j,- ). This makes it possible to simplify the design of the digitizer 16 and improve its accuracy.
[0211] According to one embodiment, the IMC device 10 may include a reference circuit 70 ( Figure 7 ), the reference circuit 70 is configured to generate a reference charge signal q ref DSP 17 uses the reference charge signal q ref To determine the output signal y i The sign and absolute value of .
[0212] The reference circuit 70 includes a reference array 71 formed by one or more memory cells 72 and at least one digitizer (also indicated here at 16 ).
[0213] The memory cell 72 can be connected to the reference Figure 1 The memory cells 20 depicted are identical; ie, they are formed from a storage element 25 and a selection element 26 .
[0214] In the illustrated embodiment, the reference array 71 is composed of reference bit lines BL coupled to K memory cells 72. ref form.
[0215] The memory cells 72 are organized so as to form N reference cell groups 74, such as for Figure 1 The memory cell group 22 is discussed.
[0216] The reference cell groups 74 are also each denoted by reference numeral 74 ref,j Specify, where j=1,...,N.
[0217] Each reference cell group 74 ref,j It is formed by positive cell 72A and negative cell 72B, which are respectively coupled to positive word line WL refj,+ and negative word line WL refj,- , such as for Figure 1 The memory cell group 22 is discussed.
[0218] Word Line WL refj,+ , WL refj, - receiving a corresponding reference activation signal S in use refj,+ , S refj,- , these signals can be Figure 4 The same reference input to the time converter 77 is generated as the input to the time converter 46.
[0219] The reference circuit 70 is configured to, in use, C The reference bit line BL ref The charge displacement in the bit line BL indicates i The MAC refinement of has zero results (y i =0), for example, where all input values x1, ..., x N are all equal to zero and / or all calculated weights G i,1 , ..., G i,N All are equal to zero.
[0220] In this embodiment, all reference activation signals S refj,+ , S refj,- have a reference duration T ref The same activation duration, for example T ref =T C / 2.
[0221] Furthermore, all reference groups 74 are programmed so that the corresponding stored calculation weights G i1 , ..., G iN corresponds to a zero calculation weight, i.e., such that the current intensity flowing in the corresponding memory cell when used is I + ref,j =I - ref,j =I ref .
[0222] However, during the initialization phase, the reference activation signal S refj,+ , S refj,- Different combinations of activation duration and / or reference current I + ref,j ,I - ref,j Different combinations of C During the period, the reference bit line BL ref The charge shifted in corresponds to the MAC refinement result of array 12 being equal to zero.
[0223] Figure 8 Purely by way of example, the refinement interval T C internal timing signal TM, a reference charge signal q generated by a reference circuit 70 ref as well as Figure 1 The two common charge signals q shown in i1 ,q i2 .
[0224] In detail, the charge signal q i1 is coupled to the bit line BL i1 The digital detector 16 provides a signal, and the charge signal q i2 is coupled to the bit line BL i2 The digital detector 16 provides a signal.
[0225] exist Figure 8 For simplicity, the signals TM and q are shown. ref ,q i1 and q i2 The corresponding simulation diagram of C Its possible evolution.
[0226] Furthermore, purely as an example, Figure 8 signal q in i1 ,q i2 ,q ref It has a slope-like trend. In the refinement interval T C However, the effective trend depends on the specific input values and calculation weights.
[0227] In the example shown, the charge signal q i2 Greater than the reference charge signal q ref Therefore, DSP 17 obtains the charge signal q i2 starts to provide a corresponding output signal y indicating a value greater than zero i2 .
[0228] Charge signal q i1 Ground to the reference charge signal q ref Therefore, the corresponding output signal y provided by DSP 17 is i1 Indicates a value less than zero.
[0229] In addition, q i1 With q ref The difference between i1 The absolute value of is less than q i2 With q ref The difference between i2 Therefore, the output signal y i1 Indicated by the output signal y i2 In other words, the MAC operation associated with column i1 is has a MAC operation associated with column i2 lower than The absolute value of .
[0230] Reference charge signal q ref The fact that the reference cell 72 is used by the reference circuit 70 allows tracking the reference charge value Q ref possible modifications of , for example due to a drift of the transconductance value of the storage element 25 over time.
[0231] Therefore, the IMC device 10 can output accurate results even if the memory array 12 deviates from the behavior established during the initialization phase.
[0232] In addition, when calculating the weight G ij To the corresponding current value I + i,j and I - i,j The mapping and / or input value x j To the corresponding duration T j,+ and T j,- In the case where the mapping is not symmetric with respect to zero, the reference charge signal q ref The possibility of indicating the charge value can be useful.
[0233] Fig. 9 Shows Figure 1 The general cell group 80 (herein denoted by reference numeral 80) of the memory array 12 i,j denoting different embodiments of the present invention.
[0234] Cell Group 80 i,j It consists of four cells, of which two positive cells 82A', 82A" and two negative cells 82B', 82B". In practice, the positive cells 82A', 82A" form a positive cell subset, while the negative cells 82B', 82B" form a negative cell subset.
[0235] Cells 82A', 82A", 82B', 82B" and Figure 1 The cells 20 are identical, and each cell includes a corresponding storage element (here represented by 85) and a selection element 26.
[0236] The common node 28 of each cell 82A', 82A", 82B', 82B" is coupled (directly connected here) to the bit line BL i .
[0237] The selection elements 26 of the positive cells 82A′, 82A″ are coupled to the positive word line WL j,+ The selection elements 26 of the negative cells 82B', 82B" are coupled to the negative word line WL j,- .
[0238] The memory elements 85 are each variable resistive elements, in particular here also based on phase change material (PCM).
[0239] In this embodiment, each memory element 85 can be programmed to represent two different logic states, for example '0' and '1'; that is, they can be programmed to have two different transconductance values g1, g2, where g1 < g2, which correspond to two respective current values I1 and I2.
[0240] For example, for Vr = 100 mV, if the corresponding transconductance is equal to g1, then each cell 82A', 82A", 82B', 82B" can conduct a current I1 = 0 μA, or if the corresponding transconductance is equal to g2, then a current I2 = 1 μA flows.
[0241] The positive cells 82A', 82A" are arranged in parallel with each other electrically and coupled to the positive word line WL j,+ itself; thus, the positive cells 82A', 82A" act as a single positive cell and can be programmed to absorb a positive cell current I i having one of the following three values from the bit line BL + i,j : I1 + I1, I1 + I2 or I2 + I2.
[0242] Similarly, the negative cells 82B', 82B" are arranged in parallel with each other electrically and coupled to the same negative word line WL j,- ; thus, the negative cells 82B', 82B" act as a single negative cell and can be programmed to absorb a current I i having one of the following three values from the bit line BL - i,j : I1 + I1, I1 + I2 or I2 + I2.
[0243] Thus, electrically, the overall behavior of the cell group 80 i,j is the same as that described for the memory cell group 22 Figure 1 for i,j .
[0244] Thus, the cell group 80 can be used in the same way as the memory cell group 22 Figure 1 discussed for i,j to represent the corresponding computational weights G with sign and absolute value i,j . ij .
[0245] Fig.10 An embodiment of the bias circuit 18 of the IMC device 10 is shown.
[0246] The bias circuit 18 is biased by the bias current I b Start generating bias voltage Vr and provide bias voltage Vr to Figure 1 The bit lines BL1, . . . , BL M .
[0247] In this embodiment, bias circuit 18 biases all bit lines BL1, . . . , BL M However, the bias circuit 18 may bias the bit lines BL1, . . . , BL M Bias voltages different from each other starting from the bias voltage Vr are supplied.
[0248] The bias circuit 18 includes a circuit for generating a bias current I b A current source 90 having an input node 92 and a bias impedance Z b The reference network is 91.
[0249] In this embodiment, the current source 90 is a controllable current source that receives an external signal EXT (eg, from a user of the IMC device 10 ) indicating the bias current I b expected value.
[0250] The bias network 91 receives a bias current I at an input node 92. b The input node 92 is in the bias current I b and bias impedance Z b The voltage as a function of .
[0251] The bias circuit 18 further comprises a voltage distribution circuit, here formed by an operational amplifier 95 having an output 96 providing a bias voltage Vr.
[0252] The operational amplifier 95 has a non-inverting input terminal coupled to an input node 92 of the bias network 91. The operational amplifier 95 has an inverting input terminal coupled (particularly here directly coupled) to an output terminal 96 of the operational amplifier 95.
[0253] Depending on the specific application, the output terminal 96 of the amplifier 95 is coupled to the bit lines BL1, . . . , BL M .
[0254] In practice, in this embodiment, the voltage at the input node 92 of the bias network 91 constitutes the bias voltage Vr.
[0255] In detail, in this embodiment, the bias network 91 is formed by a bias memory array, and is therefore also denoted by reference numeral 91 hereinafter, the bias memory array having a total bias transconductance value g b And comprises one or more memory cells, here a plurality of memory cells 97 of non-volatile type.
[0256] Bias array 91 may be part of memory array 12 or may be a separate memory array.
[0257] The unit cell 97 of the bias array 91 has a Figure 1 The memory cells 20 of the memory array 12 have the same circuit configuration.
[0258] In detail, the cells 97 of the bias array 91 each comprise a storage element 98 and a selection element 99 , in particular identical to the storage element 25 and the selection element 26 of the memory cell 20 , respectively.
[0259] In practice, the storage element 98 is based on the same technology as used to obtain the storage element 25. For example, if the storage element 25 is based on a PCM material, then the storage element 98 is also based on a PCM material, in particular the same PCM material.
[0260] Depending on the specific application, cells 97 of bias array 91 are programmed to store respective transconductance values that may be equal to or different from one another.
[0261] The number of cells 97 of bias array 91 and the corresponding transconductance values may be selected at the design stage such that bias array 91 forms a statistically significant sample of memory array 12 .
[0262] In practice, from a statistical point of view, the total bias transconductance g of the bias array 91 is b represents the total transconductance of the memory array 12.
[0263] For example, the total transconductance of the memory array 12 may be equal to the total transconductance of all memory cells 20 during the refinement interval T C The memory array 12 will have a transconductance of when the periods are simultaneously activated.
[0264] For example, the number of memory cells 97 may be higher than one hundred.
[0265] For example, the memory cell 97 may be programmed so that the total bias transconductance of the bias array 91 is g b is equal to the average value of the total transconductance of the memory array 12.
[0266] For example, the average value may represent an average transconductance value of the memory array 12 in use, such as measured during a calibration or initialization phase of the IMC device 10 .
[0267] The storage element 98 has a first terminal coupled to the input node 92 of the bias network 91 and a second terminal coupled to a reference potential node (here, ground 29 ) via a selection element 99 .
[0268] In practice, all memory cells 97 share the same bias bit line BL bias .
[0269] The selection elements 99 are each formed by a respective switch, for example a BJT transistor, a diode or a MOSFET transistor, here an NMOS transistor, in particular identical to the selection element 26 of the memory cell 20 , which is arranged in series with the respective storage element 98 .
[0270] In this embodiment, the selection elements 99 of all cells 97 are coupled to the same word line WL_B so as to be controlled by the same bias activation signal, which may be generated by the word line activation unit 14 or other components of the IMC device 10 (not shown here), for example.
[0271] However, each selection element 99 may be controlled by respective activation signals that are different from one another, for example depending on which memory cells are to be activated during calculations performed by the IMC device 10 .
[0272] Bias impedance Z b is a function of the transconductance value stored in memory cell 97 and the duration of the bias activation signal provided to bias word line WL_B.
[0273] From the bias current I b The fact that the bias voltage Vr is generated makes it possible to change the bias impedance Z b and / or bias current I b To adjust the bias voltage Vr.
[0274] During the lifetime of the IMC device 10, the IMC device 10 is subject to temperature variations, which affect the value of the transconductance of the memory cell 20. Furthermore, the value of the transconductance of the memory cell 20 is subject to drift; for example, in case the memory cell 20 is a PCM memory cell, the storage element 25 is subject to aging phenomena (e.g., amorphization).
[0275] The above deviation of the transconductance value from the programmed value will cause the positive current I + i,j and negative current I - i,j relative to the programmed value, resulting in the calculation of the weight G ij The mapping of i An error occurred.
[0276] However, temperature changes or drifts that affect memory cell 20 also affect the value of the transconductance of memory cell 97 of bias array 91 because memory cell 97 represents a statistically significant sample of memory cells 20 .
[0277] Therefore, the transconductance g of the bias array 91 is b The overall value of will also be affected by such deviations or drifts.
[0278] Total transconductance g b The change in bias impedance Z b The change of Vr and thus means the change of the voltage at the input node 92 of the reference network 91. Therefore, the bias voltage Vr may also change according to the temperature and the drift phenomenon of the memory cell 20.
[0279] In detail, the variation of the bias voltage Vr makes it possible to compensate for the variation of the transconductance value of the memory cell 20 .
[0280] For example, if the memory cell 20 experiences a drift that causes the corresponding transconductance value to increase, then the current I + i,j ,I - i,j Therefore, the corresponding bit line current I BL,i However, at the same time, the bias transconductance g of the bias array 91 b The total value also increases.
[0281] Transconductance g b The increase in bias impedance Z b and therefore means a reduction in the bias voltage Vr.
[0282] The reduction of bias voltage Vr involves current I + i,j ,I - i,j The decrease in the bit line current I BL,i of reduction.
[0283] in other words, Fig.10 The bias circuit 18 enables compensation of the current I + i,j ,I - i,j possible changes that may originate from the drift of the transconductance value of the memory cell 20. In this way, the current I + i,j ,I - i,j Keep it equal to the corresponding design value, thereby preventing errors that may occur during the execution of MAC operations by the IMC device 10.
[0284] Fig.11 An embodiment of the digital detector 16 is shown in detail. Referring now to the coupled to the plurality of bit lines BL1, . . . , BL M Any bit line BL in i The digital detector 16 will be described in detail below.
[0285] The digital detector 16 includes an integrating stage 110 and a counter stage 111 .
[0286] In this embodiment, the integrating stage 110 includes a current mirror 115 which converts the bit line BL i The bit line current I BL,i Mirrored into the input node 116 of the corresponding integrating stage 111 .
[0287] The current mirror 115 has a current ratio of 1:k, so that the mirrored bit line current k·I BL,i flows through the input node 116 of the integrating stage 110 .
[0288] In detail, the current mirror 115 has a current mirror coupled to the bit line BL. i The first branch formed here by the corresponding PMOS transistor 117 and the second branch formed here by the corresponding PMOS transistor 118 coupled to the corresponding integrating stage 110
[0289] The sources of PMOS transistors 117 and 118 are coupled to a supply node 120, where the voltage is V DD , and the gates of the PMOS transistors 117, 118 are coupled to each other and to the drain of the PMOS transistor 117. The drain of the PMOS transistor 118 is coupled (specifically directly connected here) to the input node 116 of the integrating stage 110.
[0290] The integration stage 110 includes a first integration circuit 121 , a second integration circuit 122 , and a switch circuit 123 coupled between the first integration circuit 121 and the second integration circuit 122 .
[0291] The first integration circuit 121 and the second integration circuit 122 are coupled to the input node 116 to receive the mirrored bit line current k·I BL,i .
[0292] The first integrator circuit 121 includes a first inverter 124 having an output terminal 125, a capacitor C coupled to the output terminal 125 of the first inverter 124, and a capacitor C coupled to the output terminal 125 of the first inverter 124. A A capacitor 127 is provided, and a second inverter 128 whose input terminal is coupled to the output terminal 125 of the first inverter 124.
[0293] The first inverter 124 has a supply node coupled to the input node 116 of the integrating stage 110 and receives the first control signal IN at an input terminal. A .
[0294] In practice, the first inverter 124 is provided by mirroring the bit line current k·I BL,i Bias.
[0295] Capacitor 127 has a first terminal coupled to the output node 125 of first inverter 124 and a second terminal coupled to a reference node (here, ground).
[0296] The output node 125 of the first inverter 124 is at the first integrated voltage drop V across the capacitor 127. A .
[0297] The second inverter 128 has a first sampling threshold (hereinafter defined as “first threshold V th1 ”), receiving a first integrated voltage V at the input terminal A and provides a first threshold V at the output th1 and the first integrated voltage V A A first switching signal S1 is a function of .
[0298] In detail, the first switch signal S1 is when the first integrated voltage V A Below the first threshold V th1 has a high logic value and when the first integrated voltage V A Above the first threshold V th1 A logic signal with a low logic value when
[0299] The second integrator circuit 122 includes a first inverter 130 having an output terminal 131, a capacitor C coupled to the output terminal 131 of the first inverter 130, and a B A capacitor 132 , and a second inverter 133 whose input terminal is coupled to the output terminal 131 of the first inverter 130 .
[0300] The first inverter 130 has a supply node coupled to the input node 116 of the integrating stage 110 and receives the second control signal IN at an input terminal. B .
[0301] In practice, the first inverter 130 is provided by mirroring the bit line current k·I BL,i Bias.
[0302] Capacitor 132 has a first terminal coupled to the output node 131 of first inverter 130 and a second terminal coupled to a reference node (here, ground).
[0303] The output node 131 of the first inverter 130 is at the second integrated voltage drop V across the capacitor 131. B .
[0304] The second inverter 133 has a second sampling threshold V th2 (hereinafter defined as the second threshold V th2 ), receiving a second integrated voltage V at the input terminal B and provides a second threshold V at the output th2 and the second integrated voltage V B The second switching signal S2 is a function of .
[0305] In detail, the second switch signal S2 is when the second integrated voltage V B Below the second threshold V th2 has a high logic value and when the second integrated voltage V B Above the second threshold V th2 A logic signal with a low logic value when
[0306] In this embodiment, the first threshold V th1 Equal to the second threshold V th2 However, depending on the specific application, the first threshold V th2 can be compared with the second threshold V th2 different.
[0307] The switch circuit 123 is a latch formed by two inverters 135 , 136 arranged in a ring configuration, a first switch 137 controlled by a first switch signal S1 , and a second switch 138 controlled by a second switch signal S2 .
[0308] The switch circuit 123 has a first node 140 coupled to the input of the inverter 136 and the output of the inverter 135 , and a second node 141 coupled to the output of the inverter 136 and the input of the inverter 135 .
[0309] The first node 140 provides a first control signal IN A The second node 141 provides a second control signal IN B .
[0310] The first switch 137 is coupled to the first node 140 and the voltage V′ DD and the second switch 138 is coupled between the second node 141 and the voltage V' DD between nodes.
[0311] Voltage V' DD can be equal to the voltage V DD Or different from it. For example, if the voltage V' DD With voltage V DD different, especially less than the voltage V DD , the digital detector 16 may then include a voltage scaling circuit, such as a transistor (not shown here), whose source and drain terminals are coupled between the supply node 120 and the input node 116 of the integrating stage 110 .
[0312] In this embodiment, the switch circuit 123 also receives an enable signal EN, which controls the activation of the switch circuit 123. For example, the enable signal EN can be used to keep the switch circuit 123 off when it is not in use, thereby optimizing energy consumption. In addition, the enable signal EN can be used to set the switch circuit 123 to a defined state, such as when the IMC device 10 is turned on.
[0313] The charge counting stage 111 is coupled to first and second nodes 140 , 141 of the switching circuit 123 .
[0314] In detail, the charge counter stage 111 includes an inverter 144 whose input is coupled to the second node 141 and a counter, which includes an inverter 145 whose input is coupled to the first node 140 and a plurality of D-type flip-flops 147, the plurality of D-type flip-flops 147 including a first flip-flop 147.2, a second flip-flop 147.3 and a last flip-flop 147.F, where F is the charge signal q i The number of digits.
[0315] In practice, the counter of the charge counter stage 111 has F−1 flip-flops 147 .
[0316] The output of inverter 144 provides a charge signal q i The first q i (1), which is the least significant bit.
[0317] In other words, at the end of the calculation performed by the IMC device 10, the output of the inverter 144 can be used as the charge signal q i The least significant bit q i (1).
[0318] The flip-flops 147 are sequentially cascaded to one another from the first flip-flop 147 . 2 to the last flip-flop 147 .F.
[0319] Each flip-flop 147 has a clock input (CK input), a data input (D input), a Q output, and Output.
[0320] The CK input of the first flip-flop 147.2 is coupled to the output of the inverter 145. The output is sent back to the D input of the first flip-flop 147.2. The Q output of the first flip-flop 147.2 constitutes the charge signal q i The second q i (2).
[0321] The CK input of the second flip-flop 147.3 is coupled to the CK input of the first flip-flop 147.2. The second flip-flop 147.3 The output is sent back to the D input of the second flip-flop 147.3. The Q output of the second flip-flop 147.3 constitutes the charge signal q i The third q i (2).
[0322] The description for the second flip-flop 147 . 3 applies, mutatis mutandis, to all subsequent flip-flops (not shown here) up to the F−1st flip-flop (also not shown).
[0323] Finally, the CK input of the last flip-flop 147.F is coupled to the CK input of the F-1th flip-flop. Output. The last flip-flop 147.F The output is sent back to the D input of the last flip-flop 147.F, and the Q output of the last flip-flop 147.F constitutes the charge signal q i The most significant bit q i (F).
[0324] refer to Fig.12 The first inverter 124 of the first integration circuit 121 is a CMOS inverter formed by a series circuit of a PMOS transistor 150 and an NMOS transistor 151 coupled to each other at an output node 125. The PMOS transistor 150 and the NMOS transistor 151 receive a first control signal IN at respective gate terminals. A .
[0325] The source of the PMOS transistor 150 is coupled to the input node 116 of the integrating stage 110 .
[0326] The second inverter 128 of the first integrating circuit 121 is a CMOS inverter formed by a series circuit of a PMOS transistor 152 and an NMOS transistor 153 coupled to each other at a node 154 where the first switching signal S1 is provided.
[0327] The first threshold V of the second inverter 128 th1 is the switching threshold of the second inverter 128 and therefore depends on the characteristics, such as threshold or on-resistance, of the PMOS transistor 72 and the NMOS transistor 73. In practice, the switching threshold may be the input voltage at which the output of the inverter has a high logic value or the input voltage at which the output of the inverter has a low logic value.
[0328] For example, the switching threshold of the second inverter 128 can be defined as the corresponding input voltage (ie, the first integrated voltage V A ) is equal to the operating point of the corresponding output voltage (ie, the first switching signal S1).
[0329] The first inverter 130 of the second integration circuit 122 is a CMOS inverter formed by a series circuit of a PMOS transistor 155 and an NMOS transistor 156 coupled to each other at an output node 131. The PMOS transistor 155 and the NMOS transistor 156 receive a second control signal IN at their respective gate terminals. B .
[0330] The source of the PMOS transistor 155 is coupled to the input node 116 of the integrating stage 110 .
[0331] The second inverter 133 of the second integrating circuit 122 is a CMOS inverter formed by a series circuit of a PMOS transistor 157 and an NMOS transistor 158 coupled to each other at a node 159 where the second switching signal S2 is provided.
[0332] The second threshold V of the second inverter 133 th2 is the switching threshold of the second inverter 133; that is, depends on the characteristics of the PMOS transistor 157 and the NMOS transistor 158. For example, the switching threshold depends on the gate-source voltage that enables current to flow through the source-drain path of the PMOS transistor 157 and the NMOS transistor 158.
[0333] like Fig.12 As shown in the detailed embodiment of the present invention, the first inverter 135 and the second inverter 136 of the switch circuit 123 are cross-coupled CMOS inverters, each of which includes a supply node (here at a voltage V') DD ) and the ground and a corresponding PMOS transistor 160 and a corresponding NMOS transistor 161 are coupled in series with each other.
[0334] In addition, the first inverter 135 and the second inverter 136 of the switch circuit 123 each include a corresponding enable switch, which is a PMOS transistor 162, which is coupled to the voltage V'. DD between the power supply node of the inverter and the PMOS transistor 160 of the corresponding inverter.
[0335] The PMOS transistor 162 is controlled by an enable signal EN.
[0336] In use, the bit line current I i of the bit line BL BL,i is mirrored into the integrating stage 110 of the corresponding digital detector 16.
[0337] Fig.13 Shows an example of the Fig.11 time behavior of the first control signal IN A , the first integration voltage V A and the second integration voltage V B of the digital detector 16.
[0338] For t0 < t < t1, the first integration voltage V A is below the first threshold V th,1 . Therefore, the PMOS transistor 152 of the second inverter 128 is turned on, while the NMOS transistor 153 of the second inverter 128 is turned off. Thus, the first switch signal S1 (not shown here) has a high value and the first switch 137 is open. The first control signal IN A has a low value.
[0339] Therefore, referring to the first inverter 124 of the first integration circuit 121, for t0 < t < t1, the PMOS transistor 150 is turned on and the NMOS transistor 151 is turned off.
[0340] At the same time, for t0 < t < t1, the second control signal IN B has a high value. Therefore, referring to the first inverter 130 of the second integration circuit 122, for t0 < t < t1, the PMOS transistor 155 is turned off and the NMOS transistor 156 is turned on.
[0341] Therefore, the mirrored bit line current k·I BL,i flows out from the input node 116 and only flows through the first inverter 124 of the first integration circuit 121, but does not flow through the first inverter 130 of the second integration circuit 122.
[0342] Specifically, the mirrored bit line current k·I BL,i passes through the PMOS transistor 150 and charges the capacitor 127. Thus, the first integration voltage V A increases with time for t0 < t < t1.
[0343] Specifically, in the Fig.13 example, the first integration voltage V A increases linearly with time for t0 < t < t1; however, the evolution of the first integration voltage V A depends on the bit line current I within the time interval t0 < t < t1 BL,i Specific evolution
[0344] When the first integration voltage V A equals the first threshold voltage V th,1 the NMOS transistor 153 of the second inverter 128 is turned on and the PMOS transistor 152 is turned off
[0345] In this embodiment, the first control signal IN A presents a high value at time t2
[0346] The time delay between time t1 and t2 can correspond to, for example, the propagation delay of the second inverter 128 of the first integration circuit 121 and / or the switching time of the first switch 137
[0347] For t1 < t < t2, the mirrored bit-line current I BL,i continues to charge the capacitor 127; thus, the first integration voltage V A increases to the maximum value (time t2)
[0348] At time t2, when the first control signal IN A presents a high value, the second control signal IN B (not shown here) presents a low value (the inverter 136 of the switching circuit 123 receives the first control signal IN at the input terminal A )
[0349] When the first control signal IN A has a high value, the PMOS transistor 150 and the NMOS transistor 151 of the first inverter 124 of the first integration circuit 121 are respectively in the off and on states. At the same time, when the second control signal IN B has a low value, the PMOS transistor 155 and the NMOS transistor 156 of the first inverter 130 of the second integration circuit 122 are respectively in the on and off states
[0350] Therefore, for t > t2, the mirrored bit-line current k·I BL,i flows out from the input node 116 and only flows through the first inverter 130 of the second integration circuit 122, and does not flow through the first inverter 124 of the first integration circuit 121
[0351] Specifically, the mirrored bit-line current k·I BL,i flows through the PMOS transistor 155 and charges the capacitor 132 of the second integration circuit 122. Therefore, the second integration voltage V B increases with time starting from time t2
[0352] Specifically, in Fig.13 the example of, the second integration voltage V B increases linearly with time for t > t2; however, the evolution of the second integration voltage V B depends on the specific evolution of the bit line current I BL,i .
[0353] When the first control signal IN A has a high value, the capacitor 127 of the first integration circuit 121 discharges through the NMOS transistor 151 of the first inverter 124, and the first integration voltage V A thus decreases to zero.
[0354] When the second integration voltage V B becomes equal to the second threshold voltage V th,2 (at time t3), the NMOS transistor 158 of the second inverter 133 turns on, while the PMOS transistor 157 turns off.
[0355] Thus, at time t4, in a manner similar to that discussed above for the first control signal IN A at time t2, the second control signal IN B presents a high value.
[0356] Specifically, in response to the second integration voltage V B reaching the second threshold V th,2 , the second switch signal S2 switches to a low value and the second switch 138 closes, causing the second node 141 to be at voltage V' DD , and thus the second control signal IN B presents a high value.
[0357] The time delay between times t3 and t4 can correspond, for example, to the propagation delay of the second inverter 133 of the second integration circuit 122 and / or the switching time of the second switch 138.
[0358] For t3 < t < t4, the mirrored bit line current k·I BL,i continues to charge the capacitor 132 of the second integration circuit 122; thus, the second integration voltage V B increases to a maximum value (at time t4).
[0359] For t2 < t < t4, the switching circuit 123 keeps the first control signal IN A high, while the second control signal IN B is low.
[0360] At time t4, in response to the second control signal IN B presenting a high value, the first control signal IN A again presents a low value.
[0361] In response to the first control signal IN A Presents a low value and, in a manner similar to that discussed for t1 < t < t2, the mirrored bit-line current k·I BL,i Recharges the capacitor 127 of the first integration circuit 121 again until time t6.
[0362] Thus, from time t6 to time t7, in a manner similar to that discussed for t2 < t < t4, the mirrored bit-line current k·I BL,i Charges the capacitor 132 of the second integration circuit 122 until time t7.
[0363] Still referring to Fig.11 , the counter stage 111, specifically the flip-flop 147, counts the number of switching events of the first control signal IN A , and in particular, in this embodiment, counts the number of rising edges of the first control signal IN A .
[0364] In practice, Fig.11 and Fig.12 's digital detectors 16 can each integrate the bit-line current I of the corresponding bit-line BL by performing multiple successive sampling iterations. i of the bit-line current I BL,i . In each sampling iteration, for example, referring to the sampling iteration in which the mirrored bit-line current k·I BL,i flows through the first integration circuit 121, the integration stage 110 generates a first integration voltage V BL,i that is the time integral of the mirrored bit-line current k·I A , compares the first integration voltage V A with a first threshold V th,1 , and in response to the first integration voltage V A reaching the first threshold V th,1 , resets the first integration voltage V A , specifically here by switching the first control signal IN A . The counter stage 110 updates the corresponding charge signal q A in response to the first integration voltage V th,1 reaching the first threshold V i .
[0365] In this embodiment, the least significant bit of the charge signal q i is the value of the second control signal IN end at the end of the calculation performed by the IMC device 10 (time t B ).
[0366] In other words, the digital detectors 16 each integrate the bit-line current I BL,i Converts into multiple charge packets and counts the charge packets to sample the corresponding bit line current I BL,i , where each charge packet corresponds to a charge accumulated on capacitors 127 , 132 , which causes switching of second inverters 128 , 133 .
[0367] Therefore, the bit line current is integrated into the capacitance C tot Compared with the case of a single capacitor, the capacitance of the capacitors 127 and 132 can be reduced. In detail, the capacitance of the capacitors 127 and 132 can be smaller than the capacitance C tot Low 2 F times, where F is the charge signal q i The number of digits.
[0368] Therefore, the die area occupied by the digital detector 16 is small, and thus the manufacturing cost of the IMC device 10 is low.
[0369] In addition, when the bit line current I BL,i Flows through the corresponding bit line BL i When (i.e., in the refinement interval T C During this period), each digital detector 16 begins to discretize the corresponding bit line current I BL,i Therefore, the charge signal q i In the refinement interval T C Immediately after completion, be ready for subsequent refinement.
[0370] For example, each digital detector 16 may detect the corresponding bit line current I BL,i Sampling is performed until the count end signal STOP_CNT indicates the refinement interval T C The end of.
[0371] In addition, according to the embodiment shown, the switch circuit 123 responds to the first integrated signal V A Reaching the first threshold V th,1 The first integration circuit 121 is disabled and the second integration circuit 122 is enabled, and in response to the second integration signal V B Reaching the second threshold V th,2 The first integration circuit 121 is enabled and the second integration circuit 122 is disabled.
[0372] This allows the bit line current I BL,i The first integration circuit 121 and the second integration circuit 122 are alternately sampled, so that the bit line current I BL,i Capacitor 127 can be charged while capacitor 132 is discharging, and capacitor 132 can be charged while capacitor 127 is discharging. By doing so, no charge is lost during sampling, and digital detector 16 can achieve bit line current I BL,i High accuracy of integration.
[0373] Still reference Fig.12 The set stage 110 may further include a first stop switch and a second stop switch, here a first NMOS transistor 171 and a second NMOS transistor 172, indicated by dashed lines, which prevent the corresponding digital detector 16 from responding to the corresponding bit line current I BL,i Take samples.
[0374] In detail, the first NMOS transistor 171 and the second NMOS transistor 172 have drain terminals coupled to the output nodes 127, 131 of the first integration circuit 121 and the second integration circuit 122, respectively; and source terminals coupled to a reference potential line (here, ground). The first NMOS transistor 171 and the second NMOS transistor 172 receive the count end signal STOP_CNT at their respective gate terminals.
[0375] In the refinement interval T C When the counting is finished, the count end signal STOP_CNT switches to a high logic value, thereby turning on the first NMOS transistor 171 and the second NMOS transistor 172 and short-circuiting the output nodes 127 and 131 of the first integration circuit 121 and the second integration circuit 122 to the ground.
[0376] Therefore, the first integration circuit 121 and the second integration circuit 122 stop the bit line current BL i integral.
[0377] According to one embodiment, Figure 3 The timer 45 may be formed by a current-controlled oscillator, such as a ring oscillator formed by an odd number of inverters or a Fig.14 The oscillator shown in FIG. 1 has an oscillation frequency f o In this case, the timer 45 receives the supply current. The oscillation frequency f o Depends on the supply current.
[0378] The timer 45 updates the timing signal TM at an update frequency f u Depends on the oscillation frequency f o , and thus depends on the supply current.
[0379] Therefore, the refinement time T C , that is, the timing signal TM changes from an initial value (e.g., zero) to a value at time t start )) reaches the final value MAX_CNT (final moment t end ) depends on the supply current of timer 45.
[0380] In practice, the desired refinement duration T can be set by adjusting the supply current of the timer 45. C .
[0381] like Figure 3 As indicated by the dashed arrow in FIG. 1 , the supply current of the timer 45 can be Fig.10 The bias circuit 18 generates a bias current I b In this way, the supply current of the timer 45 can be controlled by the external signal EXT.
[0382] In practice, the refinement duration T of the IMC device 10 is C It can be adjusted by the user of the IMC device 10.
[0383] In addition, due to the bias current I b The update frequency f of the timer signal TM is therefore 1 / 200 MHz, which is generated by the current source 90 and is therefore not affected by the drift of the value transconductance of the bias array 91. u and thus the refinement time T C It is also independent of the drift of the memory cell 97.
[0384] Fig.14 and Fig.15 An embodiment of a current-controlled timer 45 is shown in detail.
[0385] In detail, the timer 45 includes a circuit that generates an oscillator current I from a supply current. OSC The current mirror 180, where the supply current is equal to Fig.10 The bias current I b , and the oscillator current I OSC A counting portion 181 is provided which provides a timer signal TM.
[0386] The current mirror 180 has a mirror ratio of 1:p, so the oscillator current I OSC p·I REF .
[0387] In detail, the current mirror 180 has a first branch (here formed by a corresponding PMOS transistor 183 ) coupled to the current source 32 and a second branch (here formed by a corresponding PMOS transistor 184 ) coupled to the counting portion 181 .
[0388] The sources of PMOS transistors 183, 184 are coupled to a supply node 185 (which is at a voltage V DD ), and the gates of the PMOS transistors 183 and 184 are coupled together and coupled to the drain of the PMOS transistor 183. The drain of the PMOS transistor 184 is coupled (specifically, directly connected here) to the input node 187 of the counting part 181.
[0389] The counting portion 181 of the timer 45 includes an integration stage 190, here formed by a first integration circuit 191, a second integration circuit 192 and a switch circuit 193 coupled between the first integration circuit 191 and the second integration circuit 192, and a counter stage 195 coupled to the integration stage 190 and providing a timer signal TM.
[0390] The first integrator circuit 191 and the second integrator circuit 192 are coupled to the input node 187 to receive the oscillator current I OSC .
[0391] The first integrator circuit 191 includes a first inverter 197 having an output terminal 198, a first integrator circuit 198 having a capacitor C′ coupled to the output terminal 198 of the first inverter 197, and a first integrator circuit 199 having a capacitor C′ coupled to the output terminal 198 of the first invertor 197. A A capacitor 199 is provided, and a second inverter 200 whose input terminal is coupled to the output terminal 198 of the first inverter 197.
[0392] The first inverter 197 has a supply node coupled to the input node 187 of the counting section 181 and receives the first oscillator control signal OS at an input terminal. A .
[0393] In practice, the first inverter 197 is driven by the oscillator current I OSC Bias.
[0394] Capacitor 199 has a first terminal coupled to the output node 198 of the first inverter 197 and a second terminal coupled to a reference potential node (here, ground).
[0395] The output node 198 of the first inverter 197 is at the first oscillator integrated voltage V′ across the capacitor 199. A .
[0396] The second inverter 200 has a first oscillator threshold V' th1 (hereinafter referred to as the "first threshold V' th1 ”), receiving the first oscillator integrated voltage V' at the input terminal A and provides a first threshold V' at the output th1 and the first oscillator integrated voltage V' A A function of the first oscillator switching signal S'1.
[0397] In detail, when the first oscillator integrates the voltage V' A Below the first threshold V' th1 When the first oscillator integrated voltage V' A Above the first threshold V' th1 When , the first oscillator switch signal S'1 has a low logic value.
[0398] The second integrator circuit 192 includes a first inverter 202 having an output terminal 203, a capacitor C′ coupled to the output terminal 203 of the first inverter 202, and a capacitor C′ coupled to the output terminal 203 of the first inverter 202. B A capacitor 204 is provided, and a second inverter 205 whose input terminal is coupled to the output terminal 203 of the first inverter 202.
[0399] The first inverter 202 has a supply node coupled to the input node 187 of the counting section 181 and receives the second oscillator control signal OS at an input terminal. B .
[0400] In practice, the first inverter 202 is driven by the oscillator current I OSC Bias.
[0401] The capacitor 204 has a first terminal coupled to the output node 203 of the first inverter 202 and a second terminal coupled to a reference potential node (here, ground).
[0402] The output node 203 of the first inverter 202 is at the second oscillator integrated voltage V′ dropped across the capacitor 204 B .
[0403] The second inverter 205 has a second oscillator threshold V' th2 (hereinafter referred to as the second threshold V' th2 ”), receiving the second oscillator integrated voltage V' at the input terminal B and provides a second threshold value V' at the output th2 and the second oscillator integrated voltage V B The second oscillator switching signal S'2 is a function of .
[0404] In detail, when the second oscillator integrates the voltage V B Below the second threshold V' th2 When the second oscillator integrated voltage V' B Above the second threshold V' th2 When , the second oscillator switch signal S'2 has a low logic value.
[0405] In this embodiment, the first threshold V′ of the second inverter 200 is th1 is equal to the first threshold V of the second inverter 128 of the digital detector 16 th1 ( Fig.11 and Fig.12 ). The second threshold value V′ of the second inverter 205 th2 Equal to digital detector 16 ( Fig.11 and Fig.12 ) of the second inverter 133 th2 .
[0406] Still reference Fig.14 The switch circuit 193 is a latch formed by two inverters 208 and 209 arranged in a ring configuration. The first switch 210 is controlled by the first oscillator switch signal S'1, and the second switch 211 is controlled by the second oscillator switch signal S'2.
[0407] The switch circuit 193 has a first node 213 coupled to the input of the inverter 209 and the output of the inverter 208 , and a second node 214 coupled to the output of the inverter 209 and the input of the inverter 208 .
[0408] The first node 213 provides a first oscillator control signal OS A The second node 214 provides a second oscillator control signal OS B .
[0409] The first switch 210 is coupled to a first node 213 and a voltage V′ DD The second switch 211 is coupled between the second node 214 and the voltage V' DD between nodes.
[0410] According to an embodiment, the voltage V′ of the counting portion 181 of the timer 45 is DD can be equal to the voltage V′ of the integrating stage 110 of the digital detector 16 DD ( Fig.11 ).
[0411] exist Fig.14 In an embodiment of the present invention, the switch circuit 193 also receives an oscillator enable signal EN', which controls the activation of the switch circuit 193. For example, the oscillator enable signal EN' can be used to keep the switch circuit 193 off when it is not in use, thereby optimizing energy consumption. In addition, the enable signal EN' can be used to set the switch circuit 193 to a clear state, such as when the IMC device 10 is turned on.
[0412] The counter stage 195 is coupled to a first node 213 and a second node 214 of the switch circuit 193 .
[0413] In detail, the charge counter stage 195 includes an inverter 216 having an input coupled to the second node 214, and a counter including an inverter 217 having an input coupled to the first node 213 and a plurality of D flip-flops 218 including a first flip-flop 218.2, a second flip-flop 218.3 and a last flip-flop 218.L, where L is the timer signal TM <l:1>The number of digits.
[0414] In practice, the counter of the charge counter stage 195 has L−1 flip-flops 218 .
[0415] The output of inverter 216 provides the first bit TM(1), ie, the least significant bit, of the timer signal TM.
[0416] The flip-flops 218 are sequentially cascaded to one another from the first flip-flop 218 . 2 to the last flip-flop 218 .L.
[0417] The flip-flops 218 each have a clock input (CK input), a data input (D input), a Q output, and a Output.
[0418] The CK input of the first flip-flop 218.2 is coupled to the output of the inverter 217. The output is sent back to the D input of the first flip-flop 218.2. The Q output of the first flip-flop 218.2 constitutes the timer signal TM <l:1>The second TM(2).
[0419] The CK input of the second flip-flop 218.3 is coupled to the CK input of the first flip-flop 218.2. The second flip-flop 218.3 The output is fed back to the D input of the second flip-flop 218.3. The Q output of the second flip-flop 218.3 constitutes the timer signal TM <l:1>The third TM(3).
[0420] The description for the second flip-flop 218 . 3 applies, mutatis mutandis, to all subsequent flip-flops (not shown here) up to the L−1th flip-flop (also not shown).
[0421] Finally, the CK input of the last flip-flop 218.L is coupled to the L-1th flip-flop Output. The last flip-flop 218.L The output is sent back to the D input of the last flip-flop 218.L. The Q output of the last flip-flop 218.L constitutes the timer signal TM <l:1>The most significant bit TM(L).
[0422] refer to Fig.15 Detailed implementation of the integration stage 190 shown in FIG. 1 , the first inverter 197 of the first integration circuit 191 is a CMOS inverter formed by a series circuit of a PMOS transistor 220 and an NMOS transistor 221 coupled to each other at an output node 198. The PMOS transistor 220 and the NMOS transistor 221 receive the first oscillator control signal OS at their respective gate terminals. A .
[0423] The source of PMOS transistor 220 is coupled to input node 187 of integrating stage 190 of timer 45 .
[0424] The second inverter 200 of the first integrating circuit 191 is a CMOS inverter formed by a series circuit of a PMOS transistor 222 and an NMOS transistor 223 coupled to each other at a node 224 providing the first oscillator switching signal S′ 1 .
[0425] The first threshold V′ of the second inverter 200 th1 is the switching threshold of the second inverter 200, and thus depends on the characteristics, such as threshold or on-resistance, of the PMOS transistor 222 and the NMOS transistor 223. In practice, the switching threshold may be an input voltage at which the output of the inverter has a high logic value or an input voltage at which the output of the inverter has a low logic value.
[0426] For example, the switching threshold of the second inverter 200 can be defined as the corresponding input voltage (ie, the first oscillator integrated voltage V') A ) is equal to the operating point of the corresponding output voltage (ie, the first oscillator switching signal S'1).
[0427] The first inverter 202 of the second integrating circuit 192 is a CMOS inverter formed by a series circuit of a PMOS transistor 225 and an NMOS transistor 226 coupled to each other at an output node 203. The PMOS transistor 225 and the NMOS transistor 226 receive the second oscillator control signal OS at their respective gate terminals. B .
[0428] The source of PMOS transistor 225 is coupled to input node 187 of count portion 190 of timer 45 .
[0429] The second inverter 205 of the second integrating circuit 192 is a CMOS inverter formed by a series circuit of a PMOS transistor 227 and an NMOS transistor 228 coupled to each other at a node 229 providing a second oscillator switching signal S'2.
[0430] The second threshold V′ of the second inverter 205 th2 is the switching threshold of the second inverter 205, and thus depends on the characteristics, such as threshold or on-resistance, of the PMOS transistor 227 and the NMOS transistor 228. In practice, the switching threshold may be the input voltage at which the output of the inverter has a high logic value or the input voltage at which the output of the inverter has a low logic value.
[0431] For example, the switching threshold of the second inverter 205 can be defined as the corresponding input voltage (ie, the second oscillator integrated voltage V') B ) is equal to the operating point of the corresponding output voltage (ie, the second oscillator switching signal S'2).
[0432] like Fig.15 As shown in the detailed embodiment of the switch circuit 193, the first inverter 208 and the second inverter 209 are cross-coupled CMOS inverters, each of which includes a supply node (here at a voltage V') DD ) and the ground are coupled in series with each other.
[0433] In addition, the first inverter 208 and the second inverter 209 of the switch circuit 193 each include a corresponding enable switch, which is a PMOS transistor 232, which is coupled to the voltage V'. DD between the power supply node of the inverter and the PMOS transistor 230 of the corresponding inverter.
[0434] The PMOS transistor 232 is controlled by the oscillator enable signal EN′.
[0435] In this embodiment, reference Fig.15 The timer 45 includes a first stop switch and a second stop switch, here a first NMOS transistor 234 and a second NMOS transistor 235 , configured to interrupt the updating of the timer signal TM by the timer 45 .
[0436] In detail, the first NMOS transistor 234 and the second NMOS transistor 235 have drain terminals coupled to the output nodes 198, 203 of the first integration circuit 191 and the second integration circuit 192, respectively; and source terminals coupled to a reference terminal (here, ground). The first NMOS transistor 234 and the second NMOS transistor 235 receive the count end signal STOP_CNT at their respective gate terminals.
[0437] When the comparator 47 ( Figure 3 ) switches the count end signal STOP_CNT to a high logic value, the first NMOS transistor 234 and the second NMOS transistor 235 are turned on, thereby short-circuiting the output nodes 198 and 203 of the first integration circuit 191 and the second integration circuit 192 to ground.
[0438] Therefore, the timer 45 stops updating the timer signal TM.
[0439] In practice, in this embodiment, the timer 45 (particularly the corresponding integration stage 181) has the same Fig.11 The circuit diagram of any one of the digital detectors 16 (particularly the circuit diagram of the corresponding integration stage 110) is the same circuit diagram.
[0440] Therefore, by comparing with any one of the digital detectors 16 according to the corresponding bit line current I BL,i Generate charge signal q i In the same way, the timer 45 can be set according to the oscillator current I OSC Generates timer signal TM <l:1>.
[0441] Thus, the timer 45 can generate the timer signal TM by executing a plurality of successive timing iterations. <l:1>In each timing iteration, for example, the oscillator current I OSC The integrator stage 190 generates a current I OSC The time integral of the first oscillator integrated voltage V' A , the first oscillator integrated voltage V' A With the first threshold V' th,1 is compared, and in response to the first oscillator integrated voltage V' A Reaching the first threshold V' th,1 The first oscillator integrated voltage V' is reset A , here in particular the first oscillator control signal OS is switched A The counter stage 195 is responsive to the first oscillator integrated voltage V' A Reaching the first threshold V' th,1 The update timer signal TM <l:1>.
[0442] In this embodiment, the least significant bit of the timer signal TM is the second oscillator control signal OS B The value of .
[0443] In other words, the timer 45 can be switched by setting the oscillator current I OSC The oscillator current I is converted into a plurality of charge packets and counted. OSC Sampling is performed, wherein each charge packet corresponds to a charge accumulated on the capacitor 199 , 204 , which charge causes the switching of the second inverter 200 , 205 .
[0444] Therefore, the timer signal TM <l:1>The update frequency f u The first oscillator control signal OS A The frequency of switching events is given by Fig.11 The first control signal IN discussed A Therefore, the update frequency f u Depends on the oscillator current I OSC The value of, that is, depends on the bias current I b and the mirror factor p of the current mirror 180, the capacitance C' A , C' B and the first threshold V′ of the second inverter 200, 205 th,1 and the second threshold V' th,2 .
[0445] In practice, the integrating stage 110 of the timer 45 behaves like a current controlled oscillator.
[0446] In use, the timer 45 is used to synchronize the current I from the corresponding bit line of each digital detector 16. BL,i Generate the corresponding charge signal q i In the same way, the oscillator current I OSC Generates timer signal TM <l:1>The fact that the corresponding integration circuits 110, 181 have the same circuit diagram allows obtaining the timer signal TM <l:1>With the charge signal q1, ..., q M The strong correlation between .
[0447] Therefore, global changes that may affect the IMC device 10 (such as the supply voltage V DD 、V' DD Drift and / or temperature variation) is compensated by the timer 45 and the digital detector 16 without affecting the accuracy of the MAC operation performed by the IMC device 10.
[0448] By changing the oscillator current I OSC , for example by changing the bias current I via an external signal EXT b , you can modify the timer signal TM <l:1>The update frequency f u , thereby modifying the refinement time T of the IMC device 10 C .
[0449] In fact, for example, the oscillator current I OSC The increase in the oscillator integrated voltage V' A 、V' B ( Fig.11 and Fig.12 ) increases faster; therefore, the first oscillator control signal OSA and the second oscillator control signal OSB switch faster, thereby also increasing the timer signal TM <l:1>The update frequency f u .
[0450] Finally, it is clear that modifications and variations may be made to what has been described and illustrated herein without departing from the scope of the present invention as defined in the appended claims.
[0451] For example, the positive activation duration T j,+ With negative activation duration T j,- The difference between the two values may be different from the corresponding input values x described and illustrated above. j A function of the sign and absolute value of , for example, a linear function that is different from direct or inverse proportionality, or a nonlinear function.
[0452] For example, the positive current I + i,j With negative current I - i,j The difference between them may be a different corresponding calculation weight G than described and illustrated above. ij A function of the sign and absolute value of ; for example, a linear function that is different from a direct or inverse proportional function, or a nonlinear function.
[0453] For example, the input to the time converter 46 may be configured such that the difference T j,+ -T ref Symbols and references Figure 4 and Figure 5 The signs of the descriptions are opposite; that is, if x j >0, then it can be T j,+ <T ref , and if x j <0, then it can be T j,+ >T ref In other words, if x j >0, then it can be T j,+ -T j,- <0, and if x j >0, then it can be T j,+ -T j,- >0.
[0454] For example, the cell group can be programmed so that the difference I + i,j -I ref The sign of is opposite to that described above; that is, if G ij >0, then it can be I + i,j ref , and if G ij <0, then it can be I + i,j >I ref , so if G ij >0, then it can be I + i,j -I - i,j <0, and if G ij <0, then it can be I + i,j -I - i,j >0.
[0455] For example, with reference Figure 4 and Figure 5 Different from what has been described, the input to the time converter 46 may be configured to be in the refinement interval T C During the period, the negative cell 23B is activated first, and then the positive cell 23A is activated. In other words, the input to the time converter 46 can be configured so that the negative activation duration T j,- From the starting time t start Extends to the corresponding switching time t s,j , and the positive activation duration T j,+ From the corresponding switching time t s,j Extend to the end time t end .
[0456] Alternatively, the input to the time converter 46 may be configured such that the positive activation duration T j,+ With negative activation duration T j,- In the refinement interval T C For example, the input to the time converter 46 may be configured such that the positive activation duration T j,+ and negative activation duration T j,- All in the refinement interval T C The starting time t start Start at.
[0457] For example, the memory cell 20 may be a multi-level memory cell, the number of levels of which is greater than three. Fig.16 The table shows that in the case where the cell 20 can be programmed in one of five different transconductance levels, the refinement interval T C Period and memory cell group 22 i,j The associated charge value Q ij .
[0458] exist Fig.16 In the example of FIG. 1 , the cell 20 can be programmed so that the corresponding current I + i,j ,I - i ,j Five values equal to 0 μA, 1 μA, 2 μA, 3 μA or 4 μA can be assumed respectively.
[0459] Furthermore, in this example, I tot =4μA and I ref =I tot / 2=2μA.
[0460] In addition, in this example, the input value x j can assume values between -64 and +64, T C = 128ns and T ref =T C / 2.
[0461] As from Fig.16 As can be seen in the example, for I + i,j =I - i,j =I - i,j =I ref and / or T j,+ =T j,- =T ref , charge Q i,j Presented with reference charge Q ref Same value.
[0462] Product G ij ·x j Combinations with positive results and Q i,j >Q ref Corresponding, and the product G ij ·x j Combinations with negative results and Q i,j ref correspond.
[0463] For example, depending on the specific implementation of the memory cell and the corresponding calculation weight G to be mapped ij , each memory cell group of the computing array 12 may include a positive cell subset and a negative cell subset, wherein the number of cells in the positive cell subset and the negative cell subset is greater than one ( Figure 1 and Figure 2 ) or two ( Fig. 9 ).
[0464] For example, the memory cells 20, 72, 97 may not be based on PCM material but on resistive memory cells based on a different technology; for example, they may be magnetoresistive memory cells (MRAM), resistive memory cells (RRAM) or static memory cells (SRAM).
[0465] The row activation unit 14 can also receive the address signal ADR ( Figure 3 The dashed arrow in the figure indicates that the C For example, the address signal ADR may be used to activate only some of the plurality of word lines during use, for example if the number of values of the input vector X is less than N.
[0466] For example, refer to Figure 1 , memory cell group 22 i,j The positive cell group 23A and the negative cell group 23B in the memory array 12 may be arranged in two rows that are not adjacent to each other in the memory array 12. i,j The positive cell 23A and the negative cell 23B in the memory array 12 may be arranged on two different columns of the memory array 12, and when in use, the two columns may be coupled together via specific column selection circuits known per se.
[0467] For example, the IMC device 10 may include a number of digital detectors 16 that is less than the number of columns M of the array 12. In this case, the corresponding current I BL,1 , ..., I BL,M The initial charge signal q1, ..., q M The generation of can be controlled by a dedicated multiplexer circuit known per se.
[0468] For example, the bias circuit 18 may be configured to bias each column of the array 12 or each memory cell group 22. i,j Different bias voltages are generated so that each memory cell group 22 i,j The positive cell current I + i,j and negative cell current I - i,j The regulation can be performed not only as a function of the corresponding programmed transconductance value, but also as a function of the corresponding bias voltage activated.
[0469] Finally, the different embodiments described above can be combined to provide further solutions.
Claims
1. An in-memory computing (IMC) device for performing a multiply and accumulate (MAC) operation, comprising: an input terminal configured to receive an input signal indicative of a plurality of input values, each input value having a corresponding sign and a corresponding absolute value; an output terminal configured to provide at least one output signal; a computation memory array comprising a plurality of memory cell groups coupled to the bit lines, wherein each memory cell group is configured to store a respective computation weight having a sign and an absolute value, each memory cell group comprising a positive memory cell coupled to a respective positive word line and a negative memory cell coupled to a respective negative word line; a row activation circuit configured to provide a positive activation signal having a positive activation duration to a positive word line of a corresponding memory cell group and to provide a negative activation signal having a negative activation duration to a negative word line of a corresponding memory cell group for each input value of the input signal during a refinement interval having a refinement duration, wherein the refinement duration comprises a combination of the positive activation duration and the negative activation duration, and wherein a difference between the positive activation duration and the negative activation duration is a function of a sign and an absolute value of the input value; wherein the positive memory cell of the corresponding memory cell group is configured to be activated by a positive activation signal and to have a corresponding positive cell current flow therethrough when activated; wherein a negative memory cell of a corresponding memory cell group is configured to be activated by a corresponding negative activation signal and to have a negative cell current flow therethrough when activated; wherein a combination of the corresponding positive cell current and the corresponding negative cell current equals the total current, and a difference between the corresponding positive cell current and the corresponding negative cell current is a function of the sign and absolute value of the corresponding calculation weight stored by the memory cell group; wherein a bit line current flowing through the bit line during the refinement interval is a function of a sum of a positive cell current and a negative cell current of the group of memory cells coupled to the bit line; as well as A column elaboration circuit is coupled to the bit line and configured to sample the bit line current and provide the at least one output signal in response thereto.
2. The IMC device of claim 1 , wherein the row activation circuit is configured to provide a corresponding positive activation signal and a corresponding negative activation signal for each input value such that a sum of the corresponding positive activation duration and the corresponding negative activation duration is equal to the refinement duration.
3. The IMC device according to claim 1, wherein: For each input value, the difference between the corresponding positive activation duration and the corresponding negative activation duration is a function of the sign and absolute value of the corresponding input value.
4. The IMC device according to claim 1, wherein: For each input value, the sign of the difference between the corresponding positive activation duration and the corresponding negative activation duration is a function of the sign of the corresponding input value and not a function of the absolute value of the corresponding input value.
5. The IMC device according to claim 1, wherein: For each input value, the absolute value of the difference between the corresponding positive activation duration and the corresponding negative activation duration is a function of the sign and / or absolute value of the corresponding input value.
6. The IMC device according to claim 1, wherein: For each input value, the absolute value of the difference between the corresponding positive activation duration and the corresponding negative activation duration is a function of the absolute value of the corresponding input value and not a function of the sign of the corresponding input value.
7. The IMC device according to claim 1, wherein: For each input value, the difference between the corresponding positive activation duration and the corresponding negative activation duration is a monotonic function of the corresponding input value.
8. The IMC device according to claim 1, wherein: For each input value, when the input value is equal to zero, the positive activation duration is equal to the negative activation duration.
9. The IMC device of claim 1 , wherein the row activation circuit comprises: a timer configured to provide a timing signal, the timing signal defining a refined interval between a start time and an end time; as well as A plurality of input-to-time converters, one for each input value, each input-to-time converter being configured to compare the timing signal with a signal indicative of a sign and an absolute value of a corresponding input value and to provide a corresponding positive activation signal and a corresponding negative activation signal in response thereto.
10. The IMC device according to claim 9, wherein: For each input value, the corresponding input-to-time converter is configured to provide one of the corresponding positive activation signal and the corresponding negative activation signal from the start time of the refinement interval to the corresponding switching time between the start time and the end time of the refinement interval, and to provide the other of the corresponding positive activation signal and the corresponding negative activation signal from the corresponding switching time to the end time of the refinement interval.
11. The IMC device of claim 1 , wherein the column elaboration circuit is configured to generate a charge signal indicative of an integral of a bit line current during a elaboration interval, compare the charge signal to a reference charge signal, and determine a sign and an absolute value of a result of a multiplication and accumulation operation based on the comparison; wherein the reference charge signal indicates a charge flowing in the bit line during the elaboration interval when all input values are zero and / or when all computation weights stored by a group of memory cells coupled to the bit line are zero.
12. The IMC device of claim 11 , further comprising a reference circuit configured to generate a reference charge signal, the reference circuit comprising: a reference memory array comprising at least one reference memory cell group coupled to a reference bit line and configured to receive a reference activation signal having a reference activation duration during a refinement interval, and wherein a reference current flows for the reference activation duration to generate a charge during the refinement interval corresponding to a charge that would flow in a bit line of the computational memory array if all input values were zero and / or if all computational weights stored by the memory cell group coupled to the bit line of the computational memory array were zero; as well as A refinement circuit is configured to sample the reference bit current during a refinement interval and provide a reference charge signal in response thereto.
13. The IMC device of claim 1, wherein the refinement duration is set by a user of the IMC device prior to the start of the refinement interval.
14. The IMC device of claim 1, wherein the positive memory cells and the negative memory cells of each memory cell group each comprise a corresponding multi-level memory cell configured to be programmed in one of three or more transconductance levels.
15. The IMC device of claim 1, wherein the positive memory cells and the negative memory cells of each memory cell group each include two corresponding two-level memory cells configured to be programmed in one of the two transconductance levels.
16. The IMC device of claim 1 , wherein the positive memory cells and the negative memory cells of each memory cell group comprise nonvolatile phase change memory cells; and wherein each nonvolatile phase change memory cell has a current path, the current path comprising a storage element and a selection element and extending between a common node and a reference potential node, the selection element of each positive memory cell being configured to selectively close the corresponding current path in response to receipt of a positive activation signal, and the selection element of each negative memory cell being configured to selectively close the corresponding current path in response to receipt of a negative activation signal.
17. A method for controlling an in-memory compute (IMC) device to perform a multiplication and accumulation (MAC) operation, the IMC device comprising a computation memory array, the computation memory array comprising a plurality of memory cell groups coupled to bit lines, the method comprising: storing the corresponding calculated weight having a sign and an absolute value in each memory cell group, wherein each memory cell group includes a positive memory cell coupled to a corresponding positive word line and a negative memory cell coupled to a corresponding negative word line; activating a positive memory cell by a corresponding positive activation signal on a corresponding positive word line; activating the negative memory cell by a corresponding negative activation signal on a corresponding negative word line; causing a corresponding positive cell current to flow through a corresponding positive memory cell of each memory cell group in response to a corresponding positive activation signal; causing a corresponding negative cell current to flow through a corresponding negative memory cell of each memory cell group in response to a corresponding negative activation signal; wherein, for each memory cell group, a combination of the corresponding positive cell current and the corresponding negative cell current equals the total current, and wherein a difference between the corresponding positive cell current and the corresponding negative cell current is a function of the sign and absolute value of the corresponding computational weight stored by the memory cell group; providing, during a refinement interval having a refinement duration, a positive activation signal having a positive activation duration to a positive word line of a corresponding group of memory cells and a negative activation signal having a negative activation duration to a negative word line of a corresponding group of memory cells for each input value in response to an input signal indicative of a plurality of input values, wherein each input value has a corresponding sign and a corresponding absolute value; wherein, for each input value, the refinement duration comprises a combination of a corresponding positive activation duration and a corresponding negative activation duration; wherein the difference between the corresponding positive activation duration and the corresponding negative activation duration is a function of the sign and absolute value of the corresponding input value; causing a bit line current to flow through a bit line coupled to the group of memory cells during a refinement interval, the bit line current being a function of a sum of a positive cell current and a negative cell current of the group of memory cells coupled to the bit line; and The bit line current is sampled and at least one output signal is provided in response thereto.
18. The method of claim 17, wherein each memory cell group is programmed so that the sum of the corresponding positive cell current and the corresponding negative cell current equals the total current.
19. The method according to claim 17, wherein: For each memory cell group, the sign of the difference between the corresponding positive cell current and the corresponding negative cell current is a function of the sign and / or absolute value of the corresponding calculation weight.
20. The method according to claim 17, wherein: For each memory cell group, the sign of the difference between the corresponding positive cell current and the corresponding negative cell current is a function of the sign of the corresponding calculation weight and not a function of the absolute value of the corresponding calculation weight.
21. The method according to claim 17, wherein: For each memory cell group, the absolute value of the difference between the corresponding positive cell current and the corresponding negative cell current is a function of the sign and / or absolute value of the corresponding calculation weight.
22. The method according to claim 17, wherein: For each memory cell group, the absolute value of the difference between the corresponding positive cell current and the corresponding negative cell current is a function of the absolute value of the corresponding calculation weight and is not a function of the sign of the corresponding calculation weight.
23. The method according to claim 17, wherein: For each memory cell group, the difference between the corresponding positive cell current and the corresponding negative cell current is a monotonic function of the corresponding calculation weight.
24. The method of claim 17, wherein: For each memory cell group, when the corresponding calculated weight is equal to zero, the corresponding positive cell current is equal to the corresponding negative cell current.
