In-memory computing device for implementing at least multilayer neural networks
By using word line activation circuits and digital detectors in memory computing devices, the dependence on analog-to-digital converters is reduced, and the problems of large ADC occupancy and slow conversion time in the prior art are solved, thus achieving low cost and High-performance multi-layer neural network in-memory computing device.
Patent Information
- Application Number
- CN202410687106.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-05-28
- Filing Date
- 2024-05-30
- Publication Date
- 2024-12-03
AI Technical Summary
The analog-to-digital converters (ADCs) of existing in-memory computing devices occupy large chip area and slow conversion time, resulting in high manufacturing costs and low performance, especially when large amounts of computing are required, hindering the multilayer of low-cost and low-area occupancy. Manufacturing of computing devices in neural network memory.
A computing device in memory is designed, using word line activation circuit, bias circuit, memory array and digital detector, multiplication and accumulation operations are realized through electrical coupling and selector, and phase change materials are used to store the calculation weights, and through timers and input to the time converter generates intermediate output data, reducing dependence on the ADC.
The area occupation and power consumption of computing devices in memory are reduced, the performance is improved, and the manufacturing of computing devices in memory in low-cost multi-layer neural network memory is realized.
Smart Images

Figure CN119068949A_ABST
Abstract
Description
[0001] Priority claim
[0002] This application claims the benefit of Italian Patent Application No. 102023000010893 filed on May 30, 2023, the contents of which are incorporated herein by reference in their entirety to the maximum extent permitted by law. Technical Field
[0003] The present invention relates to an in-memory computing (IMC) device for implementing at least a multi-layer neural network, and further relates to a control method of the IMC device and a corresponding computer program product. Background Art
[0004] As is known, in-memory computing devices perform analog data processing at the cell level using a particular arrangement of memory cells of a memory array.
[0005] For example, in-memory computing devices are used to perform multiply and accumulate (MAC) operations, which are used, for example, to implement machine learning algorithms, such as, for example, neural networks.
[0006] The multiplication and accumulation operation provides the output vector y1,…,y M , as the input vector x1,…,x N And calculate the weight g ij The product of vectors or matrices of , for example:
[0007] Right now:
[0008]
[0009] The in-memory computing device will calculate the weight g ij Stored in memory cells, and multiplication and summation operations are performed at the cell level.
[0010] In detail, for each output vector y i , known in-memory devices generate indications The invention also provides a method for discretizing a current and includes a read circuit having a corresponding analog-to-digital converter (ADC) for discretizing the current.
[0011] The in-memory computing device allows avoiding the need to transfer data back and forth between the memory and the processing circuit. Therefore, the performance of the in-memory computing device is not limited by the bandwidth of data transfer between the memory and the processing circuit, and has low power consumption.
[0012] However, it has been verified that the ADC of the known in-memory computing device occupies a large chip area and has a slow conversion time, resulting in high manufacturing cost and low performance of the corresponding in-memory computing device, especially when the in-memory computing device needs to perform a large number of calculations and therefore requires a corresponding large number of ADCs.
[0013] This problem currently impedes, among other things, the fabrication of in-memory computing devices suitable for implementing neural networks (NNs) comprising multiple layers at low cost and with reduced area footprint.
[0014] There is a need in the art to overcome the deficiencies of the prior art. Summary of the invention
[0015] Embodiments herein relate to an IMC device, a control method of an IMC device, and a corresponding computer program product.
[0016] In an embodiment, an in-memory computing (IMC) device is configured to receive an input signal including a plurality of input data and provide a plurality of intermediate output data at least indicative of a corresponding final output signal as a function of the input signal, the IMC device comprising: a word line activation circuit configured to receive the input signal and provide a plurality of word line activation signals, each word line activation signal being a function of a corresponding input data of the plurality of input data; a bias circuit configured to provide a bias voltage; a memory array comprising a plurality of memory cells having a matrix arrangement, the matrix arrangement having M columns and N' rows, and each memory cell being coupled to a corresponding bit line and a corresponding word line in a manner such that each of the M bit lines is electrically coupled to N' corresponding memory cells and each of the N' word lines is electrically coupled to M corresponding memory cells, the bit lines being configured to each receive a bias voltage, and the memory cells being configured to each store The memory cells are configured to be respectively traversed by a corresponding cell current, which is a function of a bias voltage, a corresponding word line activation signal and a corresponding calculated weight, and each bit line is configured to be traversed by a corresponding bit line current, which is the sum of the cell currents of the memory cells connected to the bit line; a plurality of K selectors, each selector is coupled to a corresponding portion of the bit lines and each selector is configured to select one of the corresponding bit lines or not select any bit line; and a digital detector for each selector, each digital detector is coupled to the corresponding selector in a manner electrically connected to the corresponding bit line selected by the corresponding selector through the corresponding selector, the digital detector is configured to sample the corresponding bit line current passing through the corresponding bit line to which the digital detector is electrically coupled, and provide corresponding intermediate output data in response to the sampled bit line current.
[0017] In an embodiment, a method for controlling an in-memory computing device as described above includes the following steps: generating a bias voltage and applying the bias voltage to a bit line; providing a plurality of activation signals to a memory cell by a word line activation circuit, each activation signal being a function of a corresponding input value; selecting one of the corresponding bit lines or not selecting the corresponding bit line by a selector; and sampling a corresponding bit line current passing through a corresponding bit line electrically coupled to the digital detector by a digital detector, and providing corresponding intermediate output data in response to the sampled bit line current. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] For a better understanding of the present invention, preferred embodiments will now be described by way of non-limiting examples with reference to the accompanying drawings, in which:
[0019] Figure 1 A block diagram of the in-memory computing device according to an embodiment is shown;
[0020] Figure 2 Shows Figure 1 A circuit diagram of a circuit activation circuit of an in-memory computing device;
[0021] Figure 3 It shows that according to the embodiment Figure 2 Detailed circuit diagram of the timer of the line activation circuit;
[0022] Figure 4 Shows Figure 3 A detailed circuit diagram of a portion of a timer;
[0023] Figure 5A-5C It shows that according to the embodiment Figure 2 Detailed circuit diagram of the line activation circuit and the corresponding input to the time converter;
[0024] Fig. 6A and Figure 6B Shows Figure 5A-5C Detailed circuit diagram of the corresponding parts of the input to the time converter;
[0025] Fig. 7A Shown by Figure 5A-5C An example of an activation signal provided by the input to the time converter;
[0026] Figure 7B Shown by Figure 1 An example of an electrical signal generated by a computing device within a memory;
[0027] Figure 8 It shows that according to the embodiment Figure 2 Detailed circuit diagram of the digital detector of the line activation circuit;
[0028] Fig. 9 Shows Figure 8 A detailed circuit diagram of a portion of a digital detector;
[0029] Fig.10 Shown by Figure 8 An example of an electrical signal generated by a digital detector;
[0030] Fig.11 , Fig.13 and Fig.14 Schematically shows Figure 1 corresponding embodiments of an in-memory computing device;
[0031] Fig.12 It shows that according to the embodiment Fig.11 , Fig.13 and Fig.14 Detailed circuit diagram of the multiplication and accumulation components of the in-memory computing device;
[0032] Fig.15 It shows the Fig.11 , Fig.13 and Fig.14 Detailed circuit diagrams of the multiplication and accumulation components of the in-memory computing device; and
[0033] Fig.16 Shown according to Fig.15 A detailed circuit diagram of an input-to-time converter of an embodiment of an in-memory computing device. DETAILED DESCRIPTION
[0034] In the following description, elements common to different embodiments are indicated by the same reference numerals.
[0035] Figure 1 An in-memory computing device 10, also referred to hereinafter as an IMC device 10, is shown, which includes a computing memory array (or matrix) (hereinafter referred to as a memory array) 12, a word line activation circuit 14, a plurality of selectors 15, a plurality of digital detectors 16, and a bias circuit 18.
[0036] The memory array 12 is configured to perform a multiplication and accumulation (MAC) operation starting from an input vector (or signal) X, where the input data is x1, ..., x N (generally identified by index n=1, ..., N) in order to generate intermediate output data dy useful for calculating the final output vector (or signal) Y k (where k=1,...,K).
[0037] The memory array 12 is of the non-volatile type and comprises a plurality of memory cells 20 organized according to a matrix arrangement having M columns and N' rows.
[0038] The memory cells 20 arranged in the same column are connected to the same memory cell 20 through corresponding bit lines BL. i interconnected, where i = 1, ..., M. In the following, the reference numeral BL will also be used. m to indicate the bit lines, where m=i=1,…,M.
[0039] The memory cells 20 arranged in the same row are connected to the memory cells 20 by corresponding word lines WL j Two adjacent word lines WL are connected to each other, where j=1, ..., N', and N' is an even number. j and WL j+1 Each word line WL j All are single symbol word line groups WL n+,- The corresponding symbol word line group WL is formed in a manner as part of n+,- (where n=1, ..., N, where N=N' / 2).
[0040] In practice, the corresponding word line WL j and the corresponding bit line BL i Associated with each memory cell 20 .
[0041] Along each bit line BL i , the memory cells 20 are operatively grouped to form a memory bank 22, as better described below. In this way, the input data x may also be processed n The notation of (n=1, ..., N) is better described below.
[0042] The memory cells 20 are programmed to each store a corresponding calculated weight g ij , which can be used as a weight for performing in-memory computations (e.g., multiply and accumulate (MAC) operations).
[0043] The word line activation circuit 14 provides a plurality of word line activation signals 21. Each word line WL j There is a word line activation signal, which is configured to activate the corresponding word line WL j Each memory cell 20, as discussed in detail below.
[0044] The word line activation circuit 14 receives an input vector X, which includes a plurality of input values x1, . . . , x N , each symbol word line group WL n+,- There is an input value.
[0045] The word line activation signal 21 is a pulse, in particular a rectangular pulse, each having a corresponding input value x n The time duration (i.e., width) of a function.
[0046] The bias circuit 18 generates a bias voltage Vr in a manner known per se and supplies the bias voltage Vr to the bit lines BL1, . . . , BL M , as discussed in detail below. In particular, and as Figure 1 As shown in FIG. 1 , the bias circuit 18 is electrically coupled to the bit lines BL1, . . . , BL M , for example between the memory cell 20 and the selector 15 .
[0047] In this embodiment, the bias circuit 18 biases all bit lines BL1, . . . , BL M However, the bias circuit 18 can provide the same voltage Vr for the bit lines BL1, . . . , BL according to the specific application starting from the bias voltage Vr. M Bias voltages different from each other are provided.
[0048] The selectors 15 are each coupled to the bit lines BL in a manner better described below. i corresponding parts so that between these bit lines BL i and the corresponding digital detector 16. In detail, each selector 15 k (k=1, ..., K, where K is the number of selectors 15 and digital detectors 16) receiving corresponding selection signals SEL k , and based on the signal, the corresponding digital detector 16 k With the selection signal SEL k The selected bit lines BL i One electrical connection.
[0049] The digital detectors 16 are analog-to-digital converters (ADCs), and each digital detector 16 is coupled to a corresponding selector 15. k Based on the selection signal SEL k , each digital detector 16 k Optionally with selector 15 k The connected bit line BL i The digital detectors 16 are each electrically connected to the corresponding current I SEL,k Sampling is performed to provide the corresponding intermediate output data dy k , the current passes through the corresponding selector 15 k flows into the selected corresponding bit line BL i (That is, by controlling the current I flowing through the digital detector 16 to the corresponding bit line BL connected to the corresponding selector 15 BL for sampling).
[0050] Generally speaking, each digital detector 16 k , the corresponding selector 15 k and electrically connected to the corresponding selector 15k The bit line BL i The plurality of memory cells 20 together form a corresponding multiplication and accumulation component (MAC component, e.g. Fig.11 24. k As discussed better below, the various MAC components 24 are located along word lines WL j Arranged laterally, so they share the word line WL j , while having independent bit lines BL i .
[0051] Each memory cell 20 includes a storage element 25 and a selection element 26 .
[0052] The storage element 25 of each memory cell 20 is a variable resistive element, in particular here based on a phase change material (PCM), such as for example a chalcogenide.
[0053] In detail, the weight g is calculated ij φ indicates the transconductance value of the storage element 25 of the corresponding memory cell 20 , ie, it indicates the programmed resistance of the storage element 25 .
[0054] Phase change materials have at least two phases, such as an amorphous phase and a crystalline phase, each phase having a corresponding resistivity.
[0055] Phase change materials can be transformed from one phase to another by means of heat transfer (eg, using a pulse of electric current).
[0056] The resistance of each storage element 25 associated with a corresponding phase state is used to distinguish between two or more logic states of the corresponding memory cell 20 .
[0057] For example, the amorphous phase may have a higher resistance than the crystalline phase. A logic state "0" or reset state may be associated with the amorphous phase of storage element 25. A logic state "1" or set state may be associated with the crystalline phase of storage element 25.
[0058] The storage element 25 has a first terminal and a second terminal, wherein the first terminal is coupled to the corresponding bit line BL i The node 28 of the embodiment of the present invention is coupled to the reference potential node (here, ground 29) through the selection element 26.
[0059] The selection element 26 is a switch, such as a BJT transistor, 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 The word line activation signal 21 controls.
[0060] In this embodiment, the NMOS transistor forming the selection element 26 has a source coupled (here directly connected) to ground 29; a drain coupled (here directly connected) to the second terminal of the storage element 25; and a drain coupled (here directly connected) to the corresponding word line WL j of the gate.
[0061] In practice, the storage element 25 and the selection element 26 form a current path of the corresponding memory cell 20; the selection element 26 closes the corresponding current path in response to receiving the corresponding activation signal 21, thereby allowing the cell current i cell Flows from the common node 28 to ground 29.
[0062] The IMC device 10 may further include a plurality of bits coupled to the bit lines BL1, . . . , BL M The interface circuit 30 can be used, for example, to modify the transconductance value g stored in the storage element 25 in a manner known per se. ij Programming.
[0063] In addition, the IMC device 10 may further include a control circuit 31, which is operably coupled to the word line activation circuit 14, the selector 15, the digital detector 16, and the interface circuit 30 to control these components of the IMC device 10. In particular, the control circuit 31 may generate an address signal ADR and a selection signal SEL, may control the interface circuit 30, and may receive the intermediate output data dy k And generates the final output vector Y, as better described below.
[0064] Alternatively, the control circuit 31 may be located external to the IMC device 10 and operatively coupled to the IMC device 10 in a similar manner as has been described above. In this case, the IMC device 10 and the control circuit 31 are included in an in-memory computing device (not shown), for example.
[0065] In the exemplary embodiment considered here, along each bit line BL i, The memory cells 20 are operatively coupled two by two to form a memory group 22. The memory cells 20 of each memory group 22 are connected to the same bit line BL i and the corresponding word line WL j These word lines are continuous with each other and form a corresponding symbol word line group WL n+,- .
[0066] In particular, along the bit line BL considered iEach pair of memory cells 20 that are consecutive to each other forms a respective memory group 22 in such a way that each memory cell 20 is part of a single memory group 22 (ie, the memory groups 22 do not share the same memory cells 20 with each other). In the following, a particular memory group is also denoted by the reference numeral 22. m,n Indicates, where m=1,…,M and n=1,…,N.
[0067] For example, define the calculation weight g 11 and g 12 The memory cells 20 form a memory group 22 1,1 , and are connected to word lines WL 1+ and WL 1- , thus forming a symbol word line group WL 1+,- ; Define the calculation weight g 13 and g 14 The memory cells 20 form a memory group 22 1,2 , and are connected to word lines WL 2+ and WL 2- , thus forming a symbol word line group WL 2+,- ; Define the calculation weight g M(N'-1) and g MN' The memory cells 20 form a memory group 22 M,N , and are connected to word lines WL N+ and WL N- , thus forming a symbol word line group WL N+,- .
[0068] In other words, in the present embodiment, the memory array 12 includes a memory group 22 of memory cells 20. 1,1 to 22 M,N , where each memory bank 22 m,n The memory bank 22 includes two memory cells 20 arranged in a 1x2 matrix configuration, where m is an integer from 1 to M and n is an integer from 1 to N. With this arrangement, the memory bank 22 m,n There are N rows and M columns (where N = N' / 2).
[0069] Each memory bank 22 m,n The corresponding signed computation weights for the computation operations in the memory are stored. In fact, each memory cell 20 can be programmed to store data equal to one bit (g ij , where i is an integer from 1 to M and j is an integer from 1 to N'). Each calculated weight g ij has a logic “1” value or a logic “0” value, for example, which is represented by a programmable transconductance in the memory cell 20 .
[0070] In particular, a signed computation weight, such as "+1", is calculated by the memory bank 22 under consideration. m,n The programming logic "1" in the upper memory cell 20 and the programming logic "0" in the lower memory cell 20 represent (for example, g 11 =1 and g 12 =0), thus forming the matrix A signed calculated weight, such as "-1", is calculated by the memory bank 22 under consideration. m,n The programming logic "0" in the upper memory cell 20 and the programming logic "1" in the lower memory cell 20 represent (for example, g 11 = 0 and g 12 =1), thus forming a matrix and a signed calculated weight, such as "0", is calculated by the memory bank 22 under consideration. m,n The programming logic "0" in the upper memory cell 20 and the programming logic "0" in the lower memory cell 20 represent (for example, g 11 = 0 and g 12 =0), thus forming the matrix
[0071] In this way, for each symbol word line group WL n+,- , there are 22 connected to the memory group under consideration m,n The positive word line WL of the upper memory cell 20 n+ and connected to the memory bank 22 under consideration m,n The negative word line WL of the lower memory cell 20 n- .
[0072] Figure 2 A detailed and exemplary embodiment of the word line activation circuit 14 is shown, which includes a timer (or master counter) 45 providing a timer signal TM and a plurality of inputs to the time converter 46, one input to the time converter 46 for each symbol word line group WL n+,- For example, the timer signal TM is Gray coded.
[0073] The plurality of input-to-time converters 46 each receive a timer signal TM and a corresponding input value x n , and in response provides two corresponding word line activation signals 21 (one for the symbol word line group WL n+,- The corresponding positive word line WL j+ , one for the corresponding negative word line WL j- ).
[0074] The word line activation circuit 14 also receives an address signal ADR, which indicates which word lines WL are activated. jFor example, if the input vector X has a lower symbol word line group WL than the memory array 12 n+,- If the value of the number N is greater than the number of words, then the address signal ADR can be used to activate only the symbol word line group WL in use. n+,- Some of them.
[0075] For example, the address signal ADR may be an N-bit digital signal ADR n , where n=1, ..., N (i.e., each symbol word line group WL n+,- Each bit of the address signal ADR is associated with a corresponding symbol word line group WL n+,- Correspondingly, and for example, when the corresponding symbol word line group WL is to be enabled n+,- can be equal to logic 1 in the case of, or to disable the corresponding symbol word line group WL n+,- For example, ADR = '111 ... 1000' means that the last three symbol word line groups WL n+,- are not used, and the remaining symbol word line group WL n+,- is used.
[0076] According to one embodiment, the timer 45 is fed from the supply current (hereinafter referred to as the reference current I REF ) starts to provide the timer signal TM. For example, the reference current I REF A current source (of known type and in Figure 3 32) is generated in a manner known per se.
[0077] The timer signal TM is an L-bit digital signal, also indicated as the timer signal TM below and in the drawings. <l:1>, which is used as the reference current I REF The update frequency of the function f u Increase.
[0078] In practice, the timer signal TM is a counter signal.
[0079] The timer 45 may reset the timer signal TM to a starting value, eg, to zero, when a new calculation to be performed by the IMC device 10 begins (eg, in response to receiving a start signal from a user of the IMC device 10 ).
[0080] According to the embodiment, again referring to Figure 2 , the word line activation circuit 14 may further include a calculation completion comparator 170 .
[0081] The calculation is completed. The comparator 170 receives the timer signal TM. <l:1>and the maximum count signal MAX_COUNT <l:1>, and provides a counting end signal END as a response.
[0082] Maximum count signal MAX_COUNT <l:1>may be configured by the user of the IMC device 10 and indicates the maximum duration of a calculation performed by the IMC device 10. For example, when the corresponding bit line BL i When all associated memory cells 20 are activated, the maximum count signal MAX_COUNT <l:1>Can indicate equal to or greater than any intermediate output data dy k The maximum duration of time it takes to reach the corresponding maximum value (e.g., all F bits are equal to 1). However, the maximum count signal MAX_COUNT <l:1>A lower maximum duration may be indicated, for example, if it is desired that the IMC device 10 obtain a shorter computation time.
[0083] refer to Figure 3 and Figure 4 An exemplary and non-limiting embodiment of the timer 45 is described in detail.
[0084] exist Figure 3 In the embodiment, the timer 45 includes a current mirror 180 and a counting part 181. The current mirror 180 is obtained from the reference current I REF Start generating oscillator current I OSC , the counting section 181 receives the oscillator current I OSC Start providing timer signal TM .
[0085] The mirror ratio of current mirror 180 is 1:p, so the oscillator current I OSC is p·I REF .
[0086] 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 part 181 ).
[0087] The sources of the PMOS transistors 183 and 184 are coupled to a supply node 185, where the supply node is at a supply voltage V DD The gates of the PMOS transistors 183 and 184 are coupled to each other and 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.
[0088] 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 and second integration circuits 191, 192) and a counter stage 195 coupled to the integration stage 190 and providing a timer signal TM.
[0089] 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 .
[0090] The first integrator circuit 191 includes a first inverter 197 having an output terminal 198, a capacitor C′ coupled to the output terminal 198 of the first inverter 197, and a A A capacitor 199 and a second inverter 200 whose input terminal is coupled to the output terminal 198 of the first inverter 197.
[0091] The first inverter 197 has a coupling to the counting portion 181 ( Figure 4 ) is a power supply node of an input node 187 and receives a first oscillator control signal OS at an input terminal A .
[0092] In practice, the first inverter 197 is driven by the oscillator current I OSC Bias.
[0093] 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).
[0094] The output node 198 of the first inverter 197 is at the first oscillator integrated voltage V′ A , the voltage drops across capacitor 199.
[0095] The second inverter 200 has a first oscillator threshold V' th1 (hereinafter referred to as the first threshold V' th1 ), which receives the first oscillator integrated voltage V' at the input terminal A The first oscillator switching signal S'1 is provided at the output terminal, and the first oscillator switching signal S'1 is a first threshold value V' th1 and the first oscillator integrated voltage V' A function.
[0096] 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.
[0097] 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 B A capacitor 204 and a second inverter 205 whose input terminal is coupled to the output terminal 203 of the first inverter 202.
[0098] The first inverter 202 has a circuit coupled to the counting portion 181 ( Figure 4 ) and receives the second oscillator control signal OS at the input terminal. B .
[0099] In practice, the first inverter 202 is driven by the oscillator current I OSC Bias.
[0100] 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).
[0101] The output node 203 of the first inverter 202 is at the second oscillator integrated voltage V′ B , the voltage drops across capacitor 204 .
[0102] The second inverter 205 has a second oscillator threshold V' th2 (hereinafter referred to as the second threshold V' th2 ), which receives the second oscillator integrated voltage V' at the input terminal B The second oscillator switching signal S'2 is provided at the output terminal, and the second oscillator switching signal S'2 is a second threshold value V' th2 and the second oscillator integrated voltage V B function.
[0103] 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.
[0104] The switch circuit 193 is a latch formed by two inverters 208, 209 arranged in a ring configuration, the first switch 210 is controlled by the first oscillator switching signal S'1 and the second switch 211 is controlled by the second oscillator switching signal S'2.
[0105] 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 .
[0106] The first node 213 provides a first oscillation control signal OS A The second node 214 provides a second oscillation control signal OS B .
[0107] 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.
[0108] exist Figure 3 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 maintain the switch circuit 193 turned off when not in use, thereby allowing optimization of power consumption. In addition, the enable signal EN' can be used to set the switch circuit 193 to a defined state, such as when the IMC device 10 is turned on.
[0109] The counter stage 195 is coupled to a first node 213 and a second node 214 of the switch circuit 193 .
[0110] In detail, the charge counter stage 195 includes an inverter 216 having its input coupled to the second node 214 and a counter including an inverter 217 having its input coupled to the first node 213 and a plurality of D-type 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.
[0111] In practice, the counter of the charge counter stage 195 has L−1 flip-flops 218 .
[0112] The output terminal of the inverter 216 provides the first bit TM(1), ie, the least significant bit, of the timer signal TM.
[0113] The flip-flops 218 are cascaded to each other, from the first flip-flop 218 . 2 to the last flip-flop 218 .L in sequence.
[0114] Each flip-flop 218 has a clock input (CK input), a data input (D input), a Q output, and Output.
[0115] The CK input of the first flip-flop 218.2 is coupled to the output of the inverter 217. The output is fed back to the D input of the first flip-flop 218.2. The Q output of the first flip-flop 218.2 is the timer signal TM <l:1>The second TM(2).
[0116] 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 is the timer signal TM <l:1>The third TM(3).
[0117] 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).
[0118] 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 fed back to the D input of the last flip-flop 218.L. The Q output of the last flip-flop 218.L is the timer signal TM <l:1>The most significant bit TM(L).
[0119] refer to Figure 4 , the first inverter 197 of the first integrating circuit 191 is a CMOS inverter formed by a series circuit of a PMOS transistor 220 and an NMOS transistor 221 coupled to an output node 198. The PMOS transistor 220 and the NMOS transistor 221 receive a first oscillator control signal OS at respective gate terminals. A .
[0120] The source of PMOS transistor 220 is coupled to input node 187 of integrating stage 190 of timer 45 .
[0121] 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 mutually coupled to a node 224 providing the first oscillator switching signal S'1.
[0122] 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 of the PMOS transistor 222 and the NMOS transistor 223, for example, on the threshold or on-resistance. In practice, the switching threshold may be the input voltage for which the output of the inverter has a high logic value, or may be the input voltage for which the output of the inverter has a low logic value.
[0123] For example, the switching threshold of the second inverter 200 may 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).
[0124] 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 an output node 203. The PMOS transistor 225 and the NMOS transistor 226 receive the second oscillator control signal OS at respective gate terminals. B .
[0125] The source of PMOS transistor 225 is coupled to input node 187 of count portion 190 of timer 45 .
[0126] 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 mutually coupled to a node 229 providing the second oscillator switching signal S'2.
[0127] 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 of the PMOS transistor 227 and the NMOS transistor 228, for example, on the threshold or on-resistance. In practice, the switching threshold may be the input voltage at which the output of the inverter has a high logic value, or may be the input voltage at which the output of the inverter has a low logic value.
[0128] 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).
[0129] like Figure 4 As shown in the detailed embodiment, the first inverter 208 and the second inverter 209 of the switch circuit 193 are cross-coupled CMOS inverters, each of which includes a power supply node (here, voltage V') coupled in series with each other. DD ) and ground.
[0130] 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.
[0131] The PMOS transistor 232 is controlled by the oscillator enable signal EN′.
[0132] In this embodiment, reference Figure 4 The timer 45 includes a first stop switch and a second stop switch configured to stop updating the timer signal TM by the timer 45 , here a first NMOS transistor 234 and a second NMOS transistor 235 .
[0133] 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 END generated by the calculation end comparator 170 at the corresponding gate terminals.
[0134] When the calculation is completed, the comparator 170 ( Figure 2 ) switches the end signal END 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 the ground.
[0135] Therefore, the timer 45 stops updating the timer signal TM.
[0136] Thus, the timer 45 generates the timer signal TM by executing a plurality of successive timing iterations. <l:1>In each timing iteration, for example, the reference oscillator current I OSC The timing iteration of the first integration circuit 191 generates a first oscillator integrated voltage V' by the integration stage 190. A As the oscillator current I OSC The time integral of the first oscillator 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 by switching the first oscillator control signal OS 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>.
[0137] In this embodiment, the least significant bit of the timer signal TM is the second oscillator control signal OS B The value of .
[0138] In other words, timer 45 sets 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 causes the switching of the second inverter 200 , 205 .
[0139] Therefore, the timer signal TM <l:1>The update frequency f u The update frequency f is given by the switching event frequency of the first oscillator control signal OSA. u Depends on the oscillator current I OSC (ie, the reference current I REF ) depends on the mirror factor p of the current mirror 180, the capacitance C' A , C' B and the first threshold V′ of the first inverter 200 and the second inverter 205 th,1 and the second threshold V' th,2 .
[0140] In practice, the integrating stage 190 of the timer 45 acts as a current controlled oscillator.
[0141] refer to Figure 5A-5C , an embodiment of the input to the time converter 46 will now be described. In particular and purely by way of example, Figure 5A-5C The inputs to the time converters 461, 462 and 46 N , which are configured to be connected to the corresponding symbol word line group WL 1+,- , WL 2+,- and WL N+,- ; Clearly, the rest is input to the time converter 46 n (not shown so as not to unnecessarily obscure the view and understanding) having the same input to the time converters 461, 462 and 46 as shown here N Similar structure.
[0142] Each input to the time converter 46 n are connected to the corresponding symbol word line group WL n+,- Word line WL j , as mentioned before.
[0143] Input to Time Converter 46 n Including data interface circuit 252 n , which identifies the input data x n The symbol S n Sum value D n . Data interface circuit 252 n An example of Fig. 6A This is shown in and will be better discussed below.
[0144] Input to Time Converter 46 n It may also include a logic circuit 250 n , the logic circuit 250 n Generates a corresponding start signal START at the beginning of each MAC processing step n , for example, in response to a global start signal received by the IMC device 10. Start signal START n Indicates the start of a MAC processing step and thus the current computation iteration. n The generation of can for example depend on the input data x n The corresponding signed value of n Generates START when not zero n ) and occurs in a manner known per se. Alternatively, the start signal START n It can be generated, for example, by the control circuit 31 .
[0145] Input to Time Converter 46 n Also includes a data interface circuit 252 coupled to n Comparison circuit 256 n . Comparison circuit 256 n Receive input data x at the input terminal n The value of D n and timer signal TM <l:1>And compare them with each other to generate the corresponding matching signal MTC n When the timer signal TM <l:1>becomes equal to the input data x n The value of D n When the comparison circuit 256 n The match signal MTC is determined to be a high logic value. Comparison circuit 256 n An example of Figure 6B and better discussed below.
[0146] Input to Time Converter 46 n Also includes a set-reset circuit 258 n (particularly, an SR-type flip-flop) having a logic circuit 250 coupled to it. n To receive the start signal START n The set input (S) of the comparator circuit 256 is coupled to the comparator circuit 256. n To receive the matching signal MTC n Based on these inputs, the set-reset circuit 258 n Generates a word line enable signal EN_WL at the output terminal (Q) n When the start signal START n When the signal assumes a high logic value, the signal assumes a high logic value, and the signal is maintained as long as the matching signal MTC does not become a high logic value.
[0147] Input to Time Converter 46 n Also includes an enable logic circuit 260 n , the enabling logic circuit 260 n Coupled to data interface circuit 252 n , receiving input data x at the input terminal n The symbol S n and refinement signal ELAB n (eg, generated by the control circuit 31) and logically combine them with each other. In particular, the refinement indicator signal ELAB n The logic state of the enable logic circuit 260 indicates whether the first (positive) refinement is being performed or whether the second (negative) refinement is being performed, as better described below. n It is an exclusive-OR (XOR) logic gate.
[0148] Input to Time Converter 46 n Also includes a first combinational logic circuit 262 n (particularly, an AND logic gate), which is coupled to the set-reset circuit 258 n The output terminal Q and the enable logic circuit 260 n The output terminal receives the word line enable signal EN_WL at the input terminal. n and enable logic circuit 260 n The outputs of the word line WL are logically combined with each other to generate the word line group WL n+,- The positive word line WL n+ The intermediate drive signal.
[0149] Input to Time Converter 46 n Also includes a second combinational logic circuit 264 n (particularly, an AND logic gate), which is coupled to the set-reset circuit 258 n The output terminal Q of the enable logic circuit 260 is coupled to the enable logic circuit 260 in a logically negated manner. n The output terminal receives the word line enable signal EN_WL at the input terminal. n and enable logic circuit 260 n The outputs of the signal are logically inverted and logically combined with each other to generate the signal for the symbol word line group WL n+,- Negative word line WL n- The intermediate drive signal.
[0150] Optionally, input to the time converter 46 n A third combinational logic circuit 266 may also be included n and the fourth combinational logic circuit 268 n (Particularly, AND logic gate). Third combinational logic circuit 266 n Coupled to the first combinational logic circuit 262 n The output terminal receives the word line group WL for the corresponding symbol at the input terminal. n+,- The corresponding intermediate drive signal and address signal ADR n (ie, the nth bit of the address signal ADR), and logically combine them with each other to generate a word line for the symbol word line group WL n+,- The positive word line WL n+ The fourth combinational logic circuit 268 n Coupled to the second combinational logic circuit 264 n The output terminal receives the word line group WL for the corresponding symbol at the input terminal. n+,- The corresponding intermediate drive signal and address signal ADR n , and logically combine them with each other to generate a word line group WL for the symbol n+,- Negative word line WL n- In this way, it is possible to select the symbol word line group WL to be used. n+,- , when not all symbol word line groups WL are required n+,- Some of the symbol word line groups are disabled in the event of a situation (for example, if the number of values of the input vector X is lower than the symbol word line group WL of the memory array 12 n+,- In this case, the symbol word line group WL to be disabled n+,- The address signal ADR n The third combinational logic circuit 266 may be set to a low logic value in such a manner that the corresponding drive signal exhibits a low logic value. n and the fourth combinational logic circuit 268 n In this case, the intermediate driving signal is consistent with the driving signal.
[0151] Input to Time Converter 46 n Also includes a first word line driver circuit 270 n and the second word line driver circuit 272 n The first word line driver circuit 270 n Coupled to the third combinational logic circuit 266 n and symbol word line group WL n+,- The positive word line WL n+ The second word line driver circuit 272 n Coupled to the fourth combinational logic circuit 268 n and symbol word line group WL n+,- Negative word line WL n- In this way, each word line WL j The word line driver circuit 270 is driven by a corresponding word line driver circuit which generates a corresponding word line activation signal 21 based on a corresponding drive signal in a manner known per se. n and 272 n Each may be implemented as a CMOS driver circuit (eg, a circuit having two pairs of p- and n-channel MOSFET transistors cascaded to each other forming a buffer circuit).
[0152] Fig. 6A The data interface circuit 252 is shown in detail n .
[0153] In this exemplary and non-limiting embodiment, where the timer signal TM is Gray coded, the data interface circuit 252 n Includes Gray encoder 280 n , the Gray encoder 280 n Receive input data x at the input terminal n Gray input data x (in a format other than Gray format) and converted into Gray format in a manner known per se n,G .
[0154] Data interface circuit 252 n The combinational logic circuit 282n (particularly, an AND logic gate) is further included, which has a first input terminal receiving the clock signal CK and a second input terminal receiving the write enable signal SEL_W. n A second input terminal (eg, generated by control circuit 31).
[0155] Data interface circuit 252 n Also includes <l:0>Type memory register 284 n , which includes L D-type flip-flops. In detail, the memory register 284 n Each flip-flop has a connection to a combinational logic circuit 282 n The output terminal of the clock input terminal (CK) is provided so as to receive the Gray input data x n,G and use it as register data x n,REG Stored in storage register 284 n Connect to Gray encoder 280 n The data input terminal (D).
[0156] In particular, the combinational logic circuit 282 n The received write enable signal SEL_W n Instructions for memory register 284 n addressing, thus indicating whether this memory register 284 is enabled n This allows the memory register 284 to be controlled relative to n The number of electrical signals generated by the control circuit 31 and managed by the word line activation circuit 14 is reduced.
[0157] Data interface circuit 252 n Also includes <l:0>Type latch module 286 n , which includes L D-type flip-flops. In detail, the latch module 286 n Each flip-flop has a latch signal LATCH_D configured to receive n The clock input terminal (CK) (generated by the control circuit 31, for example) is provided so that the receiving register data x n,REG and store it in latch module 286 n Connect to memory register 284 in the manner n Specifically, latch module 286 n The first flip-flop stores register data x n,REG (Therefore, the input data x n ) n , while the latch module 286 n The remaining flip-flops store the same data as register x n,REG (Therefore, the input data x n ) value D n The corresponding bit. Due to the latch signal LATCH_D n , register data x n,REG In the latch module 286 n The storage in is such that the register data x required for the next calculation iteration is saved in the current calculation iteration. n,REG Therefore, the latch signal LATCH_D n allows the appropriate update of the data stored in the latch module 286 n Register data x in n,REG .
[0158] Figure 6B shows a comparison circuit 256 n Exemplary and non-limiting embodiments.
[0159] exist Figure 6B In the comparison circuit 256 n The first comparison block 290 is included n , first comparison block 290 n The first input terminal of the timer 45 is coupled to receive the timer signal TM <l:1>and the second input terminal is coupled to the data interface circuit 252 n (In particular, coupled to latch module 286 n A flip-flop that stores input data x n The value of D n ) to receive input data x n The value of D n The first comparison block 290 n A comparison signal CMP is generated at the output terminal, which is indicative of the timer signal TM <l:1>With input data x n The value of D n In particular, the first comparison block 290 n It is an exclusive-OR (XOR) logic gate.
[0160] Comparison circuit 256 n Also includes a second comparison block 292 n , the second comparison block 292 n The input terminal is connected to the first comparison block 290 n The output terminal of the second comparison block 292 is configured to receive the comparison signal CMP. n A match signal MTC is generated at the output based on the comparison signal CMP n In particular, the second comparison block 292 n It is a bitwise negation OR logic gate.
[0161] Fig. 7A The first positive word line WL is shown as an example 1+ And the last positive line WL N+ When the IMC device 10 starts counting (time t0), the timer signal TM is reset to a starting value, and the inputs to the time converters 46 each switch the corresponding word line activation signal 21 to a high value.
[0162] exist Fig. 7A In the example, with the first positive word line WL 1+ The associated input value x1 is lower than the last positive word line WL N+ The associated input value x N Therefore, with the first positive word line WL 1+ The associated input to the time converter 46 is relative to the last positive word line WL N+ The associated input to the time converter 46 maintains the corresponding word line activation signal 21 at a high value for a shorter time. This means that the first positive word line WL 1+ The word line activation signal 21 has an activation length T1, which is lower than the last positive word line WL N+ The activation length T of the word line activation signal 21 N .
[0163] The operation of the wordline activation circuit 14 is now described. In particular, only one iteration of the in-memory calculations associated with a corresponding input vector X to be processed is described herein; however, the following steps may be similarly repeated for each new input vector X to be processed, thereby performing a corresponding number of calculation iterations.
[0164] At the start of an in-memory computation iteration, timer 45 is initialized and decoding of address signal ADR is used to select the input to time converter 46 to be used simultaneously for data elaboration.
[0165] For each input to the time converter 46 n , if the corresponding input data x n is not zero, then the logic circuit 250 n The start signal START is activated at the beginning of each positive / negative refinement. n is set to a high logic value, and the set-reset circuit 258 n Sets output Q to a high logic value. Refines signal ELAB n The logical state of indicates whether positive refinement (logical 1) or negative refinement (logical 0) is being performed.
[0166] In fact, each computation iteration includes a first (positive) and a second (negative) data refinement step. During each data refinement step, the word line for each symbol WL n+,- The word line activation signal 21 activates a single word line WL n+ / WL n- In other words, for each symbol word line group WL enabled by the address signal ADR n+,- , there is only one word line WL in each data refinement step n+ / WL n- is active, so the same symbol word line group WL will never appear n+,- The positive word line WL n+ and negative word line WL n- The word line WL to be used in a given data refinement step n+ / WL n- The choice of occurs based on a) which (positive or negative) refinement has been performed in the data refinement step immediately preceding the one under consideration, and b) the input data x n The symbol S n .
[0167] One of four things can happen.
[0168] Case 1: If the symbol S n is logic 0 (i.e., input data x n is positive) and refines the signal ELAB n is logic 1 (i.e., the data refinement step immediately preceding the refinement step under consideration was positive), then enable logic circuit 260 n The input of the enable logic circuit 260 is different. n The output of the first combinational logic circuit 262 presents a high logic value. n Both inputs of the first combinational logic circuit 262 present a high logic value and n The output presents a high logic value to positive word line WL n+ A word line activation signal 21 is provided.
[0169] Case 2: If the symbol S n is logic 1 (i.e., input data x n is negative) and refines the signal ELAB n is logic 1 (i.e., the data refinement step immediately preceding the refinement step under consideration was positive), then enable logic circuit 260 n The inputs are equal and are logic 1, and enable logic circuit 260 n The output of the second combinational logic circuit 264 exhibits a low logic value. n Both inputs of the second combinational logic circuit 264 present a high logic value and n The output presents a high logic value to the negative word line WL n- A word line activation signal 21 is provided.
[0170] Case 3: If the symbol S n is logic 0 (i.e., input data x n is positive) and refines the signal ELAB n is logic 0 (i.e., the data refinement step immediately preceding the refinement step under consideration was negative), then enable logic circuit 260 n The inputs are equal and are logic 0, and the logic circuit 260 is enabled. n The output of the second combinational logic circuit 264 exhibits a low logic value. n Both inputs of the second combinational logic circuit 264 present a high logic value and n The output presents a high logic value to the negative word line WL n- A word line activation signal 21 is provided.
[0171] Case 4: If the symbol S n is logic 1 (i.e., input data x n is negative) and refines the signal ELAB n is logic 0 (i.e., the data refinement step immediately preceding the refinement step under consideration was negative), then enable logic circuit 260 n The inputs are different and enable logic circuit 260 n The output of the first combinational logic circuit 262 presents a high logic value. n Both inputs of the first combinational logic circuit 262 present a high logic value and n The output presents a high logic value to positive word line WL n+ A word line activation signal 21 is provided.
[0172] The timer 45 starts to increment the value of the timer signal TM and thus executes the data refinement step until the increment count value of the timer signal TM equals the value of the data interface circuit 252. n The input data x stored in n The value of .
[0173] When the increment count value of the timer signal TM is equal to the input data x stored in the data interface circuit 252n, n When the value of n The output of the set-reset circuit 258n has a high logic value, while the output Q of the set-reset circuit 258n has a low logic value. This low logic output is applied to the first combinational logic circuit 262 n and the second combinational logic circuit 264 n , so they all switch to a low logic value, thereby n+,- The end of the word line activation signal 21 is generated at . Therefore, the duration of the pulse of the word line activation signal 21 (ie, the activation length T n ) depends on the value of the timer signal TM being equal to the input data x n The amount of time required for the value.
[0174] For each input vector x, one calculation iteration is performed, thus two refinement steps (one positive and one negative, as better described below) are performed consecutively. n Each step ends when the value of . When both refinement steps are completed, the corresponding computation iteration also ends.
[0175] Finally, when the count end signal END is generated in response to the count of the timer signal TM being equal to the count of the maximum count signal MAX_COUNT, the data processing ends.
[0176] Reference now Figure 7B , which shows a simplified timing diagram of the operation of the wordline activation circuit 14, involving an exemplary in-memory computation iteration including two separate elaboration steps.
[0177] At time t1, the latch signal LATCH_D n Switching makes the data interface circuit 252 n Store input data x n The signed value of is read and the in-memory computation operation begins.
[0178] At time t2, as the first (positive) refinement of the in-memory computation operation begins, the refinement signal ELAB n Switches to logic 1. For simplicity, it is assumed here that during the first (positive) elaboration of the in-memory computation operation, in response to the input data x n The fact that it is non-zero selects all symbol word line groups WL n+,- In addition, for simplicity, assume that the input data x1 and x2 are positive, and the input data x N is negative, and for simplicity of description, other word lines are not considered below.
[0179] At time t3, the start signal START n The switching of word line WL occurs, and thus the first (positive) elaboration step of the in-memory computation operation begins. More specifically, word line WL 1+ , WL 2+ and WL N- is activated simultaneously with the word line WL 1+ and WL 2+ Case 1 and for word line WL N- 2). Still at time t3, the value of the previously reset timer signal TM starts to increase.
[0180] At time t4, the increment value of the timer signal TM is equal to the digital value of the input data X1, and thus the positive word line WL 1+ The pulse of the word line activation signal 21 ends.
[0181] At time t5, the increment value of the timer signal TM is equal to the digital value of the input data x2, and thus the positive word line WL 2+ The pulse of the word line activation signal 21 ends.
[0182] At time t6, the increment value of the timer signal TM is equal to the input data x N The digital value of the negative word line WL N- The pulse of the word line activation signal 21 ends.
[0183] At time t7, the start signal START n is switched to a low logic value and the value of the timer signal TM is reset. In addition, the intermediate output data dy1 to dy K The analog signal is sampled for analog-to-digital conversion.
[0184] At time t8, with the end of the first (positive) refinement step of the in-memory computation operation and the beginning of the second (negative) refinement step of the in-memory computation operation, the refinement signal ELAB n Switches to logic 0. For simplicity, it is assumed here that during the second (negative) elaboration of the in-memory computation operation, in response to the input data x n Non-zero, all symbol word line groups WL are selected at the same time n+,- .
[0185] At time t9, the start signal START n The switching of word line WL occurs, and thus the second (negative) elaboration step of the in-memory computation operation begins. More specifically, word line WL 1- , WL 2- and WL N+ Simultaneous activation (with the activation of word line WL 1+ and WL 2+ Case 3 and for word line WL N- The situation 4 corresponds to) occurs. Still at time t3, the value of the previously reset timer signal TM begins to increase. In addition, at time t9, the value of the previously reset timer signal TM begins to increase.
[0186] At time t10, the increment value of the timer signal TM is equal to the digital value of the input data x1, and thus the negative word line WL 1- The pulse of the word line activation signal 21 ends.
[0187] At time t11, the increment value of the timer signal TM is equal to the digital value of the input data x2, and thus the negative word line WL 2- The pulse of the word line activation signal 21 ends.
[0188] At time t12, the increment value of the timer signal TM is equal to the input data x N The digital value of the positive word line WL N+ The pulse of the word line activation signal 21 ends.
[0189] At time t13, the start signal START is n is switched to a low logic value and the value of the timer signal TM is reset. In addition, the intermediate output data dy1 to dy K The analog signal is sampled for analog-to-digital conversion.
[0190] At time t14, with the completion of the second (negative) refinement of the in-memory computation operation and the completion of the overall in-memory computation operation, the refinement signal ELAB is n Switches to logic 1.
[0191] Subsequently, the intermediate output data dy1 to dy obtained during the first positive data refinement step are calculated (e.g., by a digital processing circuit not shown and coupled to the control circuit 31 of the digital detector 16). K The digital signals (hereinafter also referred to as dy1' to dy K ') and the corresponding intermediate output data dy1 to dy obtained during the second negative data refinement step K The digital signals (hereinafter also referred to as dy1" to dy K This difference allows obtaining, for each calculation iteration and based on the corresponding input vector X, the corresponding final output vector Y=Y1, ..., Y K , where Y k =dy k' -dy k " is the kth final output data of the final output vector Y, and dy k 'anddy k " and by selector 15 k Selected bit line BL i Related.
[0192] In the examples considered so far, dy k 'Depends on the cell current i of the memory cell 20 that is active in the first refinement step cell The sum of (for example, it depends on g i1 xX1+g i2 x0+g i3 xX2 i2 x0+,…,+g i(N-1) x0+g iN xX N ) and dy k ” depends on the cell current i of the memory cell 20 that is active in the second elaboration step cell The sum of (for example, it depends on g i1 x0+g i2 xX1+g i3 x0+g i2 xX2+,…,+g i(N-1) xX N +g iN x0).
[0193] Therefore, from a time point of view, it is possible to refine each bit line BL in two successive refinement steps. i Get along this bit line BL i The corresponding bit line current I BL,i To perform signed computations. In these two refinement steps, for each sign word line group WL n+,- , using different word lines WL n+ / WL n- (That is, first, the positive word line WL n+ , then the negative word line WL n- , or vice versa, as input data x n In this way, the two bits of information contained in the memory group 22 are obtained and processed.
[0194] Figure 8 An embodiment of the digital detector 16 is shown. In detail, the following reference is made to the corresponding selector 15 coupled to the k and thus through the corresponding selector 15 k Coupled to multiple bit lines BL i is coupled to the selector 15 k Any bit line BL i Digital detector 16 k The digital detector 16 will be described.
[0195] Digital detector 16 k It includes an integration stage 110 and a counter stage 111 .
[0196] In this embodiment, the integrating stage 110 includes a current mirror 115 which mirrors the current through the selector 15 in the input node 116 of the corresponding integrating stage 111. k Connect to digital detector 16 k The bit line BL i The bit line current I BL,i .
[0197] The current mirror 115 has a mirror ratio of 1:k, so the mirrored bit line current k·I BL,i through an input node 116 of the integrating stage 110 .
[0198] In detail, the current mirror 115 has a first branch and a second branch. The first branch is formed by a selector 15. k Coupling to the bit line BL i The second branch is formed here by a corresponding PMOS transistor 117 coupled to the corresponding integrating stage 110 .
[0199] The sources of the PMOS transistors 117 and 118 are coupled to a supply node 120, which is a supply voltage V DD 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.
[0200] 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 .
[0201] 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 .
[0202] 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 and a second inverter 128 whose input terminal is coupled to the output terminal 125 of the first inverter 124.
[0203] The first inverter 124 has a phase shifter coupled to the integrating stage 110 ( Fig. 9 ) and receives a first control signal IN at the input terminal. A .
[0204] In practice, the first inverter 124 is provided by mirroring the bit line current k·I BL,i Bias.
[0205] Capacitor 127 has a first terminal coupled to the output node 125 of the first inverter 124 and a second terminal coupled to a reference node (here, to ground).
[0206] The output node 125 of the first inverter 124 is at the first integrated voltage V A , the voltage drops across capacitor 127.
[0207] The second inverter 128 has a first sampling threshold (hereinafter referred to as the first threshold V th1 ), which receives a first integrated voltage V at an 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 .
[0208] 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
[0209] 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.
[0210] The first inverter 130 has a phase shifter coupled to the integrating stage 110 ( Fig. 9 ) and receives a second control signal IN at the input terminal. B .
[0211] In practice, the first inverter 130 is provided by mirroring the bit line current k·I BL,i Bias.
[0212] The capacitor 132 has a first terminal coupled to the output node 131 of the first inverter 130 and a second terminal coupled to a reference node (here, to ground).
[0213] The output node 131 of the first inverter 130 is at the second integrated voltage V B , the voltage drops across capacitor 131.
[0214] The second inverter 133 has a second sampling threshold V th2 (hereinafter referred to as the second threshold V th2 ), which receives a second integrated voltage V at an 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 .
[0215] 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
[0216] 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 th1 can be compared with the second threshold V th2 different.
[0217] 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 .
[0218] 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 .
[0219] The first node 140 provides a first control signal IN A The second node 141 provides a second control signal IN B .
[0220] 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.
[0221] Voltage V' DD can be equal to the supply voltage V DD Or different from it. For example, if the voltage V' DD With voltage V DD different, especially lower than the supply 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 .
[0222] 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 maintain the switch circuit 123 in an off state when the switch circuit 123 is not in use, thereby allowing energy consumption to be optimized. 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.
[0223] The charge counter stage 111 is coupled to a first node 140 and a second node 141 of the switch circuit 123 .
[0224] 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 output signal dy k The number of digits.
[0225] In practice, the counter of the charge counter stage 111 has F−1 flip-flops 147 .
[0226] The output of inverter 144 provides an output signal dy k The first dy k (1), which is the least significant bit.
[0227] 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 output signal dy k The least significant bit dy k (1).
[0228] The flip-flops 147 are cascaded to each other, from the first flip-flop 147.2 to the last flip-flop 147.F in sequence.
[0229] Each flip-flop 147 has a clock input (CK input), a data input (D input), a Q output, and Output.
[0230] The CK input of the first flip-flop 147.2 is coupled to the output of the inverter 145. The output is fed back to the D input of the first flip-flop 147.2. The Q output of the first flip-flop 147.2 is the output signal dy k The second dy k (2).
[0231] 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 fed back to the D input of the second flip-flop 147.3. The Q output of the second flip-flop 147.3 is the output signal dy k The third dy k (3).
[0232] 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).
[0233] 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 fed back to the D input of the last flip-flop 147.F. The Q output of the last flip-flop 147.F is the output signal dy k The most significant bit dy k (F).
[0234] refer to Fig. 9 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 an output node 125. The PMOS transistor 150 and the NMOS transistor 151 receive a first control signal IN at respective gate terminals. A .
[0235] The source of the PMOS transistor 150 is coupled to the input node 116 of the integrating stage 110 .
[0236] 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 mutually coupled to a node 154 providing the first switching signal S1.
[0237] 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 of the PMOS transistor 152 and the NMOS transistor 153, for example on the threshold or on-resistance. 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.
[0238] For example, the switching threshold of the second inverter 128 may 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).
[0239] 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 an output node 131. The PMOS transistor 155 and the NMOS transistor 156 receive a second control signal IN at respective gate terminals. B .
[0240] The source of the PMOS transistor 155 is coupled to the input node 116 of the integrating stage 110 .
[0241] 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 to a node 159 providing the second switching signal S2.
[0242] The second threshold V of the second inverter 133 th2 is the switching threshold of the second inverter 133, i.e., it 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 allows current to flow through the source-drain path of the PMOS transistor 157 and the NMOS transistor 158.
[0243] like Fig. 9 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 voltage V' at a power supply node (here DD ) and the ground, and a corresponding PMOS transistor 160 and a corresponding NMOS transistor 161 are coupled in series with each other.
[0244] 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 and the PMOS transistor 160 of the corresponding inverter.
[0245] The PMOS transistor 162 is controlled by the enable signal EN.
[0246] In use, via the selector 15 k coupled to the digital detector 16 k of the bit line BL i of the bit line current I BL,i is mirrored in the integration stage 110 of the corresponding digital detector 16 k
[0247] Fig.10 shows Figure 8 the first control signal IN A of the digital detector 16, A the first integration voltage V B and the second integration voltage V
[0248] as an example of the trend over time. A For t0 < t < t1, the first integration voltage V th,1 is below the first threshold V. A 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
[0249] has a low value.
[0250] Therefore, referring to the first inverter 124 of the first integration circuit 121, for t0 < t < t1, the PMOS transistor 150 is turned on, while the NMOS transistor 151 is turned off. B At the same time, for t0 < t < t1, the second control signal IN
[0251] 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, while the NMOS transistor 156 is turned on. BL,i Therefore, the mirrored bit line current k·I
[0252] flows out from the input node 116 and only flows through the first inverter 124 of the first integration circuit 121, and does not flow through the first inverter 130 of the second integration circuit 122. BL,i Specifically, the mirrored bit line current k·I A flows through the PMOS transistor 150 and charges the capacitor 127. Therefore, for t0 < t < t1, the first integration voltage V
[0253] Specifically, in the Fig.10 example, for t0 < t < t1, the first integrated voltage V A increases linearly with time; however, the trend of the first integrated voltage V A depends on the specific trend of the bit line current I BL,i within the time interval t0 < t < t1.
[0254] When the first integrated voltage V A equals the first threshold voltage V th,1 , the NMOS transistor 153 of the second inverter 128 conducts and the PMOS transistor 152 is cut off.
[0255] In this embodiment, the first control signal IN A presents a high value at time t2.
[0256] The time delay between times t1 and t2 can correspond, for example, to the propagation delay of the second inverter 128 of the first integration circuit 121 and / or the switching time of the first switch 137.
[0257] For t1 < t < t2, the mirrored bit line current I BL,i continues to charge the capacitor 127; thus, the first integrated voltage V A increases to the maximum value (time t2).
[0258] 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 A ) at the input.
[0259] 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 cut-off and conducting 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 conducting and cut-off states.
[0260] 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.
[0261] 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. Thus, the second integration voltage V B Starts increasing with time from time t2.
[0262] Specifically, in the Fig.10 example, for t > t2, the second integration voltage V B Increases linearly with time; however, the trend of the second integration voltage V B Depends on the specific trend of the bit line current I BL,i .
[0263] 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. The first integration voltage V A Subsequently decreases to zero.
[0264] When the second integration voltage V B Becomes equal to the second threshold voltage V th,2 (time t3), the NMOS transistor 158 of the second inverter 133 turns on, while the PMOS transistor 157 turns off.
[0265] Therefore, at time t4, the second control signal IN B Presents a high value, similar to the case of the first control signal IN A Discussed above for time t2.
[0266] 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.
[0267] The time delay between time 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.
[0268] 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 VB increases to a maximum value (time t4).
[0269] For t2 < t < t4, the switching circuit 123 maintains the first control signal IN A At a high value and maintains the second control signal IN B At a low value.
[0270] 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.
[0271] In response to the first control signal IN A presenting a low value, the mirrored bitline current k·I BL,i returns to charge the capacitor 127 of the first integration circuit 121 until time t6, similar to the case discussed for t1 < t < t2.
[0272] Thus, from time t6 to time t7, the mirrored bitline current k·I BL,i charges the capacitor 132 of the second integration circuit 122 until time t7, similar to the case discussed for t2 < t < t4.
[0273] Referring again to Figure 8 , 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, it counts the number of rising edges of the first control signal IN A .
[0274] In practice, each digital detector 16 measures the bitline current IB k through the corresponding bitline BL i connected to it via the corresponding selector 15 L,i , thus performing multiple successive sampling iterations. In each sampling iteration, for example, referring to the sampling iteration in which the mirrored bitline 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 bitline current k·I A , compares the first integration voltage V A with a first threshold V th,1 , and resets the first integration voltage V A in response to the first integration voltage V th,1 reaching the first threshold V A , specifically here by switching the first control signal IN A . The counter stage 110 updates the corresponding output signal y A in response to the first integration voltage V th,1 reaching the first threshold V i .
[0275] In this embodiment, the least significant bit of the output signal y i is determined by the second control signal IN at the end of the calculation performed by the IMC device 10 B The value of .
[0276] In other words, each of the digital detectors 16 converts the bit line current I BL,i Converts into multiple charge packets and counts the charge packets to sample each 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 .
[0277] Therefore, the bit line current is integrated at once in the capacitor 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 output signal dy k The number of digits.
[0278] Therefore, the digital detector 16 may occupy a small chip area, so that the IMC device 10 may have a low manufacturing cost.
[0279] In addition, when the bit line current I BL,i Through the corresponding bit line BL i When each digital detector 16 starts to discretize the corresponding bit line current IB L,i Therefore, the output signal dy k It can be ready immediately after the calculation performed by the IMC device 10 is completed, or when the corresponding bit line current IB L,i It is ready immediately after stopping.
[0280] For example, according to one embodiment, the digital detectors 16 may each detect the corresponding bit line current I BL,i Sampling is performed until the digital detector 16 receives a stop signal (eg, from a user of the IMC device 10 or from the wordline activation circuit 14) indicating the end of the computation performed by the IMC device 10.
[0281] Therefore, the digital detector 16 may have a fast measurement time, thereby allowing the IMC device 10 to have a low computation time.
[0282] Furthermore, according to the illustrated embodiment, in response to the first integrated signal V A Reaching the first threshold V th,1 , the switch circuit 123 disables the first integration circuit 121 and enables the second integration circuit 122, and responds to the second integration signal V B Reaching the second threshold V th,2 , the switch circuit 123 enables the first integration circuit 121 and disables the second integration circuit 122 .
[0283] This allows the bit line current I BL,i is sampled alternately by the first integration circuit 121 and the second integration circuit 122, thereby allowing the bit line current I BL,i Capacitor 127 is charged while capacitor 132 is discharged, and capacitor 132 is charged while capacitor 127 is discharged. By doing so, no charge is lost during sampling and digital detector 16 can achieve bit line current I BL,i High measurement accuracy.
[0284] Reference again Fig. 9 The integration stage 110 may further include a first stop switch and a second stop switch, which are a first NMOS transistor 171 and a second NMOS transistor 172, which prevent the corresponding digital detector 16k from responding to the corresponding bit line current I BL,i Take samples.
[0285] In detail, the first NMOS transistor 171 and the second NMOS transistor 172 have drain terminals respectively coupled to the output nodes 127, 131 of the first integration circuit 121 and the second integration circuit 122; 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 END generated by the calculation end comparator 170 at the corresponding gate terminals.
[0286] When the timer signal TM <l:1>becomes equal to the maximum count signal MAX_CNT <l:1>When the calculation is finished, the calculation end comparator 170 switches the end signal END 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.
[0287] Therefore, the first integration circuit 121 and the second integration circuit 122 stop the bit line current BL i Perform integration.
[0288] In practice, the count end signal END may be used to determine the end of the MAC calculation of the IMC device 10 .
[0289] As mentioned above Figure 3 The timer 45 and Figure 8 Digital detector 16 k The description can be determined that the timer 45 and the digital detector 16 k The circuit structure has some advantages.
[0290] For example, the first threshold V′ of the second inverter 200 of the timer 45 is th1 is equal to the first threshold V of the second inverter 128 of the digital detector 16 th1 The second threshold value V' of the second inverter 205 of the timer 45 th2 is equal to the second threshold value V of the second inverter 133 of the digital detector 16 th2 .
[0291] Furthermore, according to one embodiment, the voltage V′ of the counting portion 181 of the timer 45 is DD Can be equal to digital detector 16 k The voltage V' of the integrating stage 110 DD In practice, the timer 45 (particularly the corresponding integration stage 190) has the same circuit diagram as the circuit diagram of any digital detector 16 (particularly the circuit diagram of the corresponding integration stage 110). Therefore, the timer 45 generates the same current as any digital detector 16 from the corresponding bit line current I BL,i Generate output data dy k In the same way, the oscillator current I OSC Generates timer signal TM <l:1>.
[0292] In use, the timer 45 is used to detect the current I BL,i Generate the corresponding output data dy k 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 output data dy1, ..., dy k Therefore, global changes (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.
[0293] By changing the oscillator current I OSC , for example, by changing the reference current I through the external signal EXT REF , you can modify the timer signal TM <l:1>The update frequency f u , and thus modifies the total calculation time of the IMC device 10. In fact, for example, the oscillator current I OSC The increase in the oscillator integrated voltage V' A 、V' B increases faster; therefore, the first oscillator control signal OS A and the second oscillator control signal OS B Switches faster, thus also increasing the timer signal TM <l:1>The update frequency f u。
[0294] Fig.11 An embodiment of an IMC device 10 is shown. In particular, Fig.11 Memory array 12, selector 15, and digital detector 16 are shown; the remaining elements of IMC device 10 are not shown here for ease of viewing, but they are still present as previously described.
[0295] As previously described, the memory array 12 has M×N′ memory cells 20 , which are grouped into M×N memory banks 22 , where N′=2N.
[0296] exist Fig.11 In the embodiment, there are K MAC components 24, which are arranged laterally and aligned to share word lines WL 1+ , ..., W.L. N- .
[0297] exist Fig.11 Each MAC component has 24 k There are RxN memory banks 22 and RxN' memory cells 20, where R is an integer and, purely as a non-limiting example, R=8 (other R values may be used, such as R=3 or R=30). The memory cells 20 are connected to the bit lines BL and word lines WL as previously described. Thus, each MAC component 24 k There are R number of columns of memory groups 22 and the same number of bit lines BL as R. Furthermore, it is assumed that M is a multiple of R, and in particular M=K·R.
[0298] Generally speaking, each MAC component 24 k The number R of columns (ie, bit lines BL) of coincides with the maximum number of layers of the neural network to be implemented by the IMC device 10. In particular, the MAC component 24 k Each row (and therefore each bit line BL) corresponds to a corresponding layer of NN. In this way, the final output data Y is obtained by processing the corresponding r-th column (where r=1, ..., R) of the MAC component 24 in the considered calculation iteration and substantially simultaneously with each other. k The final output vector Y is formed, which is calculated from the input vector X through the rth layer of the NN.
[0299] In addition, the symbol word line group WL n+,- The number N is related to the input data x1, ..., x1 that can be processed by the IMC device 10 in a single computational iteration. N The maximum number N x , y ... K The maximum number N y Corresponding. Fig.11 In the embodiment, N=N x And K = N y .
[0300] Generally speaking, K ≥ N x .
[0301] When K = N x (For example, Fig.11 ), in corresponding computation iterations, the input vector X is processed in parallel by each MAC component 24 to provide corresponding final output data y1, ..., y corresponding to the considered layer of the NN substantially simultaneously. K , these final output data form the final output vector Y. In other words, in this case, all MAC components 24 are simultaneously active to generate the final output vector Y.
[0302] Instead, when K>N x When (for example, Fig.13 and Fig.14 ), in each calculation iteration, a portion of the MAC component 24 is active and contributes to the calculation of the final output vector Y, while the remaining portion of the MAC component 24 is deactivated and does not contribute to the calculation of the final output vector Y.
[0303] For example, we can consider where K = P·N x , where P is an integer (e.g., P=3). In the exemplary case where P=3, only one-third of the MAC components 24 are simultaneously active to generate the final output vector Y in one calculation iteration, and the remaining MAC components 24 are deactivated in this calculation iteration and can be activated in a calculation iteration different from the one under consideration.
[0304] This allows the overall computational power of the IMC device 10 to be increased by a factor equal to P, without having to update each time the computational weights stored in the memory array 12. For example, this may substantially allow the IMC device 10 to implement P different neural networks in its entirety before the computational weights must be updated via the interface circuit 30; this may occur using only the portion of the MAC component 24 (and therefore the NN) that is of interest at the time.
[0305] If only a part of the MAC components 24 contributes to the generation of the final output vector Y, the remaining MAC components 24 can be disabled by the corresponding selectors 15. In particular, the selectors 15 of the MAC components 24 that are active in the considered calculation iteration receive corresponding selection signals SEL with corresponding indication values of the bit lines BL to be selected, while the selectors 15 of the MAC components 24 that are disabled in the considered calculation iteration receive corresponding selection signals SEL, each of which adopts a disabled value, such as so as not to select any bit line BL and thus prevent the corresponding digital detector 16 from being electrically connected to the bit lines BL.
[0306] Fig.12 The MAC component 24 is shown k An example of , where k=1,...,K.
[0307] In particular, the MAC component 24 k Including connecting to the bit line BL (k-1)R+1 To BL kR Memory Group 22 (k-1)R+1,1 to 22 kR,N, . Bit line BL (k-1)R+1 To BL kR Connect to selector 15 k , and selector 15 k and then coupled to the digital detector 16 k . Bit line BL (k-1)R+1 To BL kR Each entry in corresponds to a corresponding layer of the NN.
[0308] Selector 15 k is reached by selector 15 k The selection signal SEL of the control input terminal k Controlled R:1 type multiplexer. In particular, selector 15 k Each of the R input terminals is connected to a bit line BL (k-1)R+1 To BL kR The corresponding bit line in the digital detector 16 is connected to the output terminal. k .
[0309] Selector 15 k Receive selection signal SEL k , whose value determines the bit line BL (k-1)R+1 To BL kR Which of the following is related to the digital detector 16 k Therefore, the digital detector 16 k According to the selection signal SEL k From the bit line BL (k-1)R+1 To BL kR The selected bit line BL receives the current I SEL,K For example, the selection signal SEL k is an R-bit digital signal, which is all logic 0 except for a logic 1 indicating the position of the bit line BL to be selected and thus indicating the layer of NN (eg, SEL k ='01000000' selects the second bit line BL (k-1)R+2 , while SEL k ='00000001' selects the last bit line BL kR ).
[0310] If instead the selection signal SEL k If the selector 15 is set to a prohibited value (predefined type, and for example, all R bits are set to 0, that is, "00000000"), then k So that the digital detector 16 k Does not receive any current I SEL,K The digital detector 16 k and bit line BL (k-1)R+1 To BL kR Electrical decoupling.
[0311] When the digital detector 16 k Electrically coupled to the bit line BL (k-1)R+1 To BL kR When one of the two SEL,K Sampling is performed to generate intermediate output data dy in the manner previously described k .
[0312] Now refer to it again Fig.11 The operation of the IMC device 10 is described.
[0313] Upon receiving the first input vector X A After that, the first computation iteration of the IMC device 10 begins and the word line activation circuit 14 converts the first input vector X A Converted into symbol word line group WL n+,- These word line activation signals 21 are then provided to the symbol word line group WL n+,- , as mentioned before.
[0314] The selector 15 receives a selection signal SEL indicating the selection of each MAC component 24. k The first bit line BL (k-1)R+1 , corresponding to the first layer of the NN formed by the IMC device 10. For example, the selection signal SEL is equal to "10000000". Therefore, during the entire duration of the first calculation iteration, the selector 15 connects the corresponding digital detector 16 to the corresponding first bit line BL of the MAC component 24. (k-1)R+1 Electrical connection.
[0315] Each digital detector 16 k The flow through the corresponding MAC component 24 k The first bit line BL (k-1)R+1 The current I SEL,k This is done during the first refinement step to determine the corresponding intermediate output data dy k ', and is completed during the second refinement step to determine the corresponding intermediate output data dy k ", as described above. This allows each MAC component 24 to be substantially simultaneously calculated at the end of the first calculation iteration and for all MAC components 24 k Based on the intermediate output data dy of the first refinement step k ' and the intermediate output data dy of the second refinement step k "Generate the corresponding final output data Y k Thus, at the end of the first computational iteration, the first final output vector Y is generated A .
[0316] Then, receiving the second input vector X B And start the second computational iteration of the IMC device 10. In particular, the second input vector X B Can be compared with the first final output vector Y A consistent, or in any case depends on the first final output vector Y A .
[0317] In the second computation iteration of the IMC device 10, the word line activation circuit 14 sends a signal to the symbol word line group WL n+,- Provides a second input vector X B A corresponding word line activation signal 21 is generated.
[0318] In addition, the selector 15 receives a selection signal SEL indicating the selection of each MAC component 24. k The second bit line BL (k-1)R+2 , corresponding to the second layer of the NN formed by the IMC device 10. For example, the selection signal SEL is equal to "01000000". Therefore, during the duration of the second computation iteration, the selector 15 connects the corresponding digital detector 16 to the corresponding second bit line BL of the MAC component 24. (k-1)R+2 Electrical connection.
[0319] Therefore, each digital detector 16 k In two refinement steps, the flow through the corresponding MAC components 24 k The second bit line BL (k-1)R+2 The current I SEL,k Sampling is performed, thereby allowing a second final output vector Y to be generated in a manner similar to the previously described manner. B 。
[0320] In a similar manner, other computational iterations are performed (one computation for each bit line BL of the MAC component 24) until the generation of the Rth final output vector Y, which corresponds to the final result processed by the considered NN.
[0321] In particular, if the final output vector Y of a layer is used as the input vector X of a successive layer, then a cascade of the layers of the NN can be formed. In this way, given the input vector X, the corresponding final output vector Y obtained by processing through this NN can be obtained.
[0322] Fig.13 In a manner similar to Fig.11 it is shown that in the implementation of the NN (and thus in the considered computational iteration) only a part of the MAC components 24 are active and only a part of the symbol word line groups WL n+,- are required to generate the output vector Y for the IMC device 10.
[0323] In this exemplary case, the number N of input data of the input vector X x is lower than the number of symbol word line groups WL n+,- available in the IMC device 10. Thus, the address signal ADR allows only the symbol word line groups WL 1+,- to WL n+,- (where n = N x < N), i.e., the symbol word line groups for each input data of the input vector X, to be selected and enabled. The remaining symbol word line groups WL n+,- to WL N+,- are disabled by the address signal ADR, as previously described.
[0324] Furthermore, in this exemplary case, the number N k of intermediate output data dy y is lower than the number K of MAC components 24 available in the IMC device 10. Thus, the selection signal SEL allows only the MAC components 241 to 24 K' (where K' = N y < K) to be enabled, while disabling the remaining MAC components 24 K'+1 to 24 K , as previously described.
[0325] In this way, only the memory banks 22 1,1 to 22 m,n (where m < M and n < N) are enabled and operative in the considered computational iteration.
[0326] Fig.14 shows the IMC device 10 in a manner similar to Fig.11 when the IMC device 10 is capable of implementing more than one NN without updating the computational weights of the memory cells 20. In particular, each NN is implemented by a corresponding set of MAC components 24, and in particular, by a corresponding set of memory banks 22. In other words, Fig.14 the IMC device 10 of
[0327] includes a plurality of MAC circuits 24, which allows for the implementation of a corresponding plurality of NNs.
[0328] In this exemplary case, the IMC device 10 includes three sets of MAC components 24 (also referred to hereinafter as MAC circuits), and thus is capable of implementing three NNs at different times. However, it is clear that the number of MAC circuits 24 can vary, for example, be equal to 2 or greater than 3.
[0328] In particular and purely for illustrative and non - limiting purposes, the IMC device 10 includes: a first MAC circuit 24a, which includes MAC components 241 to 24 1+,- to WL n+,- (where n < N), corresponding to bit lines BL1 to BL K' (where K' < K, corresponding to bit lines BL1 to BL m here m < M); a second MAC circuit 24b, which includes MAC components 241 to 24 (n+1)+,- to WLN +, -; and a third MAC circuit 24c, which includes MAC components 24 K' to 24 1+,- to WL (n+t)+,- (where n + t < N), corresponding to bit lines BL K'+1 to 24 K (corresponding to bit lines BL m+1 to BL M ). However, it is clear that the mutual arrangement of the MAC circuits and their matrix dimensions can vary according to what is described herein.
[0329] In this exemplary case, the number n of the symbol word line groups WL n+,- corresponds to the maximum number N of input data that the first MAC circuit 24a can process in a single computational iteration, the number N - n of the symbol word line groups WL x corresponds to the maximum number N of input data that the second MAC circuit 24b can process in a single computational iteration, and the number n + t of the symbol word line groups WL n+,- corresponds to the maximum number N of input data that the third MAC circuit 24c can process in a single computational iteration. x corresponds to the maximum number N of input data that the third MAC circuit 24c can process in a single computational iteration, and the number n + t of the symbol word line groups WL n+,- corresponds to the maximum number N of input data that the third MAC circuit 24c can process in a single computational iteration. x The input vector X can have different dimensions for each MAC circuit, so the corresponding maximum number N x In addition, the number K' of the MAC components 24 is related to the maximum number N of final output data that can be processed by the first MAC circuit 24a and the second MAC circuit 24b in a single computation iteration. y The number K-K' of the MAC components 24 corresponds to the maximum number N of final output data that can be processed by the third MAC circuit 24c in a single calculation iteration. y correspond.
[0330] The MAC circuits 24a-24c are used at different times from each other, i.e., they do not function at the same time. In fact, the MAC circuits 24a-24c share some elements of the IMC device 10 (e.g., the selector 15 and the digital detector 16 are shared in the case of the first MAC circuit 24a and the second MAC circuit 24b, or the word line WL is shared in the case of the third MAC circuit 24c with respect to the first MAC circuit 24a and the second MAC circuit 24b), so it is impossible to refine multiple MAC components at the same time.
[0331] For example, the first MAC circuit 24a may first be used to store the 1,1 to 22 m,n The first NN defined by the calculated weights in the first NN is used to process the first input vector XA', and then the second MAC circuit 24b can be used to store the first MAC vector XA' in the memory group 221. ,n+1 to 22 m,N The second NN defined by the calculation weights in the second input vector X is refined B' , and then the third MAC circuit 24c can be used to store the m+1,1 Up to 22M ,n+t The third NN defined by the calculation weights in the third NN is used to refine the third input vector X C' .
[0332] By appropriately designing the number of symbol word line groups and bit lines of each MAC circuit, they can also be cascaded in such a way that the final output vector generated by one of them becomes the input vector of the successive MAC circuit. In this way, a cascade of multiple NNs can be formed.
[0333] exist Fig.13 and Fig.14 In the case of Fig.12 24k examples and the previous examples for a single MAC component ( Fig.11 ) are similarly valid and applicable.
[0334] By examining the features of the invention made according to the present invention, the advantages provided will become apparent.
[0335] The IMC device 10 allows for an increase in MAC computing power without a corresponding increase in its complexity, power consumption, and overall cost.
[0336] In fact, each digital detector 16 is associated with a plurality of bit lines BL through a corresponding selector 15. Since the area consumption and the overall cost of the IMC device 10 are generally mainly affected by the digital detectors 16, reducing the ratio between the number of digital detectors 16 and the number of available bit lines BL allows improving the MAC computing capability of the IMC device 10 without proportionally increasing its complexity, power consumption and overall cost.
[0337] In particular, this solution allows reusing and sharing circuit systems that are known to be bulky and expensive, while increasing the number of memory cells 20, which, on the other hand, occupy a small space and have a low cost, thereby increasing the computing power of the IMC device 10 without increasing its area occupation and complexity in an equally significant way.
[0338] The IMC device 10 allows the implementation of one or more NNs, each NN comprising multiple layers.
[0339] In particular, the fact that each MAC component 24 comprises a bit line BL for each layer of the NN and all these bit lines BL are coupled to the same digital detector 16 through a corresponding selector 15 allows the parallel generation of the final output data Y1, ..., Y obtained by the same layer of the NN. K , and thus use the layers of the NN in a sequential manner (the output of one layer corresponds to the input of the successive layer) without unduly increasing the complexity and cost of the IMC device 10.
[0340] Furthermore, the MAC circuits may have different numbers of MAC components 24. The MAC components 24 may have different numbers of bit lines BL, and thus different numbers of layers of NN, and they may have different groups of symbol word lines WL. n+,- Even within the same MAC component, due to the different layers considered, the symbol word line group WL n+,- These features are applicable to various embodiments of the IMC device 10 described herein, either alone or in combination.
[0341] This flexibility in the use of memory array 12 is made possible by the structure of IMC device 10 and the use of address signal ADR and select signal SEL, which allow memory groups 22 to be selectively enabled / disabled in a manner that adapts the use of memory array 12 to possible applications.
[0342] For example, there may be N=512 symbol word line groups WL n+,- (ie, 1024 word lines WL) and M=512 bit lines BL. This ensures that the IMC device 10 can perform MAC operations based on the data and layers that are commonly needed and used.
[0343] Furthermore, the use of the memory bank 20 and the operation of the IMC device 10 with two elaboration steps per computation iteration allows for performing MAC operations based on signed data.
[0344] Finally, it is clear that modifications and variations may be made to the invention described and illustrated herein without departing from the scope of the invention as defined in the appended claims.For example, the different embodiments described may be combined with each other to provide further solutions.
[0345] The embodiments of the digital detector 16, the timer 45 and the input to the time converter 46 described above are provided only as non-limiting examples, and other embodiments can also be considered in an obvious manner. For example, a voltage-controlled timer, a timer based on a ring oscillator with an odd number of inverters, a digital detector each including only one integration circuit and one counter circuit, etc. can be similarly used.
[0346] Furthermore, the timer signal TM may be an analog signal and the input to the time converter 46 may be configured to convert the corresponding input value x j The analog input signal is converted into an analog signal and the analog input signal is compared with the analog timer signal. For example, the timer signal can be derived from the current (particularly from the reference current I REF ) starts generating a voltage ramp; in this case, the update frequency of the analog timer signal indicates the slope of the voltage ramp.
[0347] The address signal ADR and the selection signal SEL may be different from those described above. For example, they may have different coding (eg, Gray coding).
[0348] Furthermore, different signed data codecs than previously described may be used in a manner which is itself obvious.
[0349] For example, each signed data may be stored by four corresponding memory cells 20, which are coupled to each other through two bit lines BL and two word lines WL, as shown in FIG. Fig.15 Generally speaking, regarding the case where each signed data is stored by two corresponding memory cells 20, which are coupled to each other by one bit line BL and two word lines WL, this allows reducing the number of electrical signals generated by the control circuit 31 and managed by the IMC device 10.
[0350] In particular, Fig.15 1 and 1. The MAC components 24 are shown to include a corresponding set of memory cells 20 of the memory array 12. k , configured to operate according to this different embodiment.
[0351] In this embodiment, the IMC device 10 is substantially similar to the previously described devices except for the differences described below. Therefore, features common to the previously described embodiments will not be repeated here.
[0352] exist Fig.15 Each memory bank has 22 m,n The memory cell 112 includes four memory cells 20 arranged in a 2x2 matrix configuration in two rows and two columns of the array 112. With this arrangement, compared to the previous embodiments (e.g., Figure 1 and Fig.12 ), requires twice as many columns to store the same amount of information.
[0353] Although Fig.15 Memory cells 20 of the same group 22 are shown to be placed in adjacent columns, but it will be understood that this is for exemplary purposes and is only for ease of description; alternatively, memory cells 20 of the same group 22 may also be placed in columns that are not adjacent to each other.
[0354] Each memory bank 22 stores a respective signed computation weight for an in-memory computation operation.
[0355] Specifically, the computation weights with a sign of "+1" are represented by the programming logic "1" in the memory cells 20 on the main diagonal of the 2×2 matrix (eg, g 11 =1 and g 22 =1) and by programming logic "0" (eg, g 12 = 0 and g 21 =0) to form the matrix That is, the identity matrix; the calculation weights with the symbol "-1" are programmed with logic "0" (eg, g 11 = 0 and g 22 = 0) and through the programming logic "1" (eg, g 12 =1 and g 21 =1) to form the matrix That is, the exchange or inverse identity matrix; and the computational weights with the symbol "0" are programmed with logic "0" (eg, g 11 =0,g 22 = 0) and by programming logic "0" (eg, g 12 = 0 and g 21 =0) to form the matrix That is, the zero matrix.
[0356] In this way, for each symbol word line group WL n+,- , there exists 22 memory groups under consideration m,n The upper memory cell 20 is connected to the positive word line WL n+ and the memory group 22 under consideration m,n The lower memory cell 20 is connected to the negative word line WL n- , and there is a memory group 22 under consideration m,n The memory cell 20 on the left is connected to the positive bit line BL m+ and the memory group 22 under consideration m,n The memory cell 20 on the right is connected to the negative bit line BL m- .
[0357] In this embodiment, the input to the time converter 46 has Fig.16 The circuit structure shown in .
[0358] In particular, each input to the time converter 46n has Figure 5A-5C Similar structure, but without the enable logic circuit 260 n And the second combinational logic circuit 264 n Receive the word line enable signal EN_WL at the input terminal n and symbol S n .
[0359] In this way, the word line group WLn considered +,- Active word line WL n+ / WL n- is based only on the input data x n The symbol S n The logic state is selected based on the previously described refinement signal ELAB n In particular, if the symbol S n is a logic 0 (indicating positive input data x n ), then during the two refinement steps of the considered computation iteration, the positive word line WL is selected n+ On the contrary, if the symbol S n is a logic 1 (indicating negative input data x n ), then during the two refinement steps of the considered computation iteration, the negative word line WL is selected n- .
[0360] In addition, selector 15 k The received selection signal SEL k Make selector 15 k The positive bit line BL is selected during the first (positive) refinement step of the considered calculation iteration m+ , and during the second (negative) refinement step of the considered calculation iteration the negative bit line BL is selected m- In this way, the intermediate output data dy generated in the first refinement step k ' and the positive line BL m+ , and thus to the memory cell 20 on the left of the memory group 22, while the intermediate output data dy generated in the second refinement step k ” and negative bit line BL m- is associated with, and is therefore associated with, the memory cell 20 to the right of the memory group 22 .
[0361] The remaining operations of the IMC device 10 are similar to those described above.
[0362] Furthermore, the memory cells 20 may be resistive memory cells not based on PCM material, but on a different technology; for example, they may be magnetoresistive (MRAM), resistive (RRAM) or static (SRAM) memory cells.
[0363] In addition, the storage element 25 of each memory cell 20 can be formed by a plurality of selectable resistive elements, which are identical or different from each other and arranged in parallel with each other, for example, between the corresponding bit line and the ground, and these resistive elements can be selectively enabled or disabled when programming the memory array 12, so that the corresponding transconductance value g ij Can be a multi-bit value.
[0364] Furthermore, the IMC device 10 has the structure of a MAC component 24 and thus the possibility of implementing different layers of the NN (in particular in a sequential manner) in order to communicate with the memory array 12 with the calculated weights and input data x1, . . . , x N The case without a symbol can be realized in a similar manner. In this case, there is no symbol word line group WL n+,- , and each word line activation signal 21 is based on the corresponding input data x n generated and provided to the corresponding word line WL n , where n=1, ..., N'.
Claims
1. An in-memory computing (IMC) device configured to receive an input signal comprising a plurality of input data and to provide a plurality of intermediate output data at least indicative of respective final output signals as a function of the input signal, the IMC device comprising: a wordline activation circuit configured to receive an input signal and provide a plurality of wordline activation signals, each wordline activation signal being a function of a respective input data of the plurality of input data; a bias circuit configured to provide a bias voltage; a memory array comprising a plurality of memory cells, the memory cells having a matrix arrangement having M columns and N' rows, and each memory cell being coupled to a corresponding bit line and a corresponding word line in such a manner that each of the M bit lines is electrically coupled to the N' corresponding memory cells and each of the N' word lines is electrically coupled to the M corresponding memory cells, the bit lines being configured to each receive a bias voltage, the memory cells being configured to each store a corresponding calculation weight and to each receive a corresponding word line activation signal from a word line activation signal of the corresponding word line, the memory cells being configured to each be traversed by a corresponding cell current, the cell current being a function of the bias voltage, the corresponding word line activation signal and the corresponding calculation weight, each bit line being configured to be traversed by a corresponding bit line current, the bit line current being a sum of the cell currents of the memory cells connected to the bit line; a plurality of K selectors, each selector coupled to a corresponding portion of the bit lines and each selector configured to select one of the corresponding bit lines or none of the bit lines; as well as a digital detector for each selector, each digital detector being electrically connected through the corresponding selector in such a manner that the corresponding bit line electrical connection is selected by the corresponding selector, the digital detector being configured to sample a corresponding bit line current passing through the corresponding bit line to which the digital detector is electrically coupled, and to provide corresponding intermediate output data in response to the sampled bit line current.
2. The IMC device according to claim 1: wherein each selector has a respective input for coupling to each bit line of the selector, a control input, and a respective output coupled to a respective digital detector; wherein each selector is configured to receive a corresponding control signal through a control input terminal, the control signal indicating prohibition of selection of a bit line coupled to the selector or indicating a bit line to be selected among the bit lines coupled to the selector; and Each selector is configured to electrically couple the corresponding digital detector with the selected corresponding bit line or to electrically decouple the corresponding digital detector from the corresponding bit line according to the corresponding control signal.
3. The IMC device of claim 1, wherein the word line activation circuit is further configured to receive an address signal indicating the plurality of word lines among the N' word lines to which the word line activation signal is provided, and provide the word line activation signal to a word line selected based on the address signal.
4. The IMC device of claim 1 , further comprising a first multiplication and accumulation (MAC) circuit defined by a selector, a digital detector, a bit line coupled to the selector, and a memory cell, the first MAC circuit being configured to perform a multiplication and accumulation (MAC) operation associated with a first neural network defined by the calculated weights of the memory cell and comprising a number R1 of layers; wherein each selector, a corresponding digital detector, a corresponding bit line coupled to the selector, and a portion of the memory cell arranged along the bit line coupled to the selector form a corresponding multiply and accumulate (MAC) component of the first MAC circuit; wherein each MAC component of the first MAC circuit has a respective number R1 of said bit lines coupled to a respective selector, the number R1 of bit lines coupled to the respective selector being equal to the number R1 of layers of the first neural network in such a way that the calculated weights of the memory cells of the respective r1-th bit line of the MAC component together define a respective r1-th layer of the first neural network, wherein for each MAC component, r1=1, ..., R1; wherein the number L of MAC components is equal to the maximum number of final output data of the final output signal generated by the first MAC circuit; Wherein for each layer of the first neural network, each MAC component is configured to generate corresponding intermediate output data of the intermediate output data, and the final output data of the final output signal generated according to the layer of the first neural network depends on the corresponding intermediate output data generated by the MAC component according to the layer of the first neural network.
5. The IMC device according to claim 4: wherein the input data are in a signed format, and each input data comprises a corresponding sign bit indicating a sign of the input data and a corresponding plurality of value bits indicating a value of the input data; wherein the memory cells along each bit line are grouped in pairs in a manner such that each memory cell is part of only one of said pairs, each of said pairs of memory cells along the same bit line defining a respective memory group configured to store respective signed computation weights; wherein for each memory group, a corresponding word line coupled to one of the memory cells of the memory group defines a positive word line, and a corresponding word line coupled to another memory cell of the memory group defines a negative word line, the positive word line and the negative word line connected to the memory group together forming a corresponding sign word line group; wherein the wordline activation circuit is configured to provide a wordline activation signal to the symbol wordline group in at least one computation iteration including a corresponding first elaboration step and a corresponding second elaboration step; The wordline activation circuit is configured to: in a first refinement step of a computation iteration, providing a first plurality of said wordline activation signals to a sign wordline group based on respective input data, each wordline activation signal of said first plurality of wordline activation signals being provided to one of a positive wordline and a negative wordline of a respective sign wordline group in accordance with respective input data of the input data, one of the positive wordline and the negative wordline of the respective sign wordline group being selected based on a sign of the respective input data and a refinement signal indicating performance of the first refinement step; as well as in a second refinement step of the computation iteration, providing a second plurality of the wordline activation signals to the sign wordline group based on the corresponding input data, each wordline activation signal of the second plurality of wordline activation signals being provided to the other of the positive wordline and the negative wordline of the corresponding sign wordline group in accordance with the corresponding input data, the other of the positive wordline and the negative wordline of the corresponding sign wordline group being selected based on the sign of the corresponding input data and the refinement signal indicating the performance of the second refinement step; wherein in the computation iteration, the selectors of the first MAC circuit are each configured to select a corresponding r1-th bit line among corresponding bit lines of the MAC components of the first MAC circuit, or not select any bit line; wherein in both the first refinement step and the second refinement step of the computation iteration, the digital detector is configured to sample a corresponding bit line current of a corresponding r1-th bit line of a MAC component passing through a first MAC circuit electrically coupled to the digital detector; wherein the digital detector is further configured to provide corresponding intermediate output data of a first refinement step in response to the bit line current sampled in the first refinement step of the computation iteration, and to provide corresponding intermediate output data of a second refinement step in response to the bit line current sampled in the second refinement step of the computation iteration; and The final output data of the final output signal generated by the first MAC circuit in the calculation iteration according to the r1-th layer of the first neural network depends on both the intermediate output data generated by the MAC component of the first MAC circuit in the first refinement step and the intermediate output data generated by the MAC component of the first MAC circuit in the second refinement step.
6. The IMC device of claim 5 , wherein the word line activation circuit comprises a timer configured to provide a timer signal and a plurality of input-to-time converters coupled to the timer and each coupled to a corresponding symbol word line group, the input-to-time converters being configured to each compare the timer signal with a corresponding input data of the input data and in response provide a corresponding activation signal to the corresponding symbol word line group.
7. The IMC device of claim 6, wherein each input to the time converter comprises: a data interface circuit configured to receive corresponding input data and identify the sign and value of the input data; a comparison circuit coupled to the data interface circuit and configured to receive at input the value of the corresponding input data and the timer signal, compare them with each other and generate, based on this comparison, a corresponding match signal indicating the fact that the timer signal is equal to the value of the input data or different therefrom; a set-reset circuit having a set input configured to receive a start signal indicating the start of a computation iteration, having a reset input coupled to the comparison circuit and configured to receive the match signal, and having an output, the set-reset circuit being configured to generate a wordline enable signal at the output based on the match signal and the start signal, the wordline enable signal indicating a value of the start signal and a value of the match signal; an enable logic circuit coupled to the data interface circuit and configured to receive the sign of the input data and the refinement signal at an input terminal and combine them with each other by an exclusive-OR logic; a first combinational logic circuit coupled to an output terminal of the set-reset circuit and an output terminal of the enable logic circuit and configured to receive at an input terminal a word line enable signal and an output of the enable logic circuit, logically combine them with each other by AND logic, and generate a corresponding intermediate drive signal for a positive word line of a corresponding sign word line group; as well as A second combinational logic circuit is coupled to the output terminal of the set-reset circuit and to the output terminal of the enable logic circuit in a logically negated manner and is configured to receive the word line enable signal and the logical inversion of the output of the enable logic circuit at the input terminal, logically combine them with each other through AND logic, and generate a corresponding intermediate drive signal for the negative word line of the corresponding sign word line group.
8. The IMC device of claim 7, wherein each input to the time converter further comprises: a third combinatorial logic circuit coupled to the output terminal of the first combinatorial logic circuit and configured to receive at an input terminal the intermediate drive signal generated by the first combinatorial logic circuit and the address signal for the corresponding symbol word line group, logically combine them with each other by AND logic, and generate a drive signal for the positive word line of the symbol word line group; as well as a fourth combinatorial logic circuit, coupled to the output terminal of the second combinatorial logic circuit and configured to receive at the input terminal the intermediate drive signal generated by the second combinatorial logic circuit and the address signal for the corresponding symbol word line group, logically combine them with each other through AND logic, and generate a corresponding drive signal for the negative word line of the symbol word line group.
9. The IMC device according to claim 4: wherein the input data are in a signed format, and each input data comprises a corresponding sign bit indicating a sign of the input data and a corresponding plurality of value bits indicating a value of the input data; wherein the memory cells of each MAC component are grouped into groups of two respective pairs in such a manner that each memory cell is part of only one of the groups, each pair being arranged along a respective bit line of the MAC component and the memory cells of the two pairs being arranged along two respective word lines of the word lines, each of said groups of memory cells defining a respective memory group configured to store a respective signed computation weight; wherein for each memory group, respective word lines coupled to two of the memory cells of the memory group define a positive word line and respective word lines coupled to another two of the memory cells of the memory group define a negative word line, the positive word lines and the negative word lines connected to the memory group together forming a respective sign word line group; wherein the wordline activation circuit is configured to provide a wordline activation signal to the symbol wordline group in at least one computation iteration including a corresponding first elaboration step and a corresponding second elaboration step; The wordline activation circuit is configured to: In a first refinement step of a computation iteration, providing a first plurality of said wordline activation signals to a sign wordline group based on corresponding input data, each wordline activation signal of said first plurality of wordline activation signals being provided to one of a positive wordline and a negative wordline of a corresponding sign wordline group according to a corresponding input data of the input data, one of the positive wordline and the negative wordline of the corresponding sign wordline group being selected based on a sign of the corresponding input data; as well as In a second refinement of the computation iteration, a second plurality of the wordline activation signals are provided to the sign wordline group based on the corresponding input data, each wordline activation signal of the second plurality of wordline activation signals being provided to the other of the positive wordline and the negative wordline of the corresponding sign wordline group according to the corresponding input data, the other of the positive wordline and the negative wordline of the corresponding sign wordline group being selected based on the sign of the corresponding input data; and wherein in the computation iterations, the selectors of the first MAC circuit are each configured to select r1′-th and r1″-th bit lines or not select any bit lines among corresponding bit lines of the MAC component of the first MAC circuit in the first refinement step and the second refinement step, respectively, wherein the r1′-th and r1″-th bit lines are bit lines coupled to the same memory group implementing the r1-th layer of the first neural network; wherein in a first refinement step of the computation iteration, the digital detector is configured to sample a respective bit line current of a respective r1′-th bit line of a MAC component of a first MAC circuit electrically coupled to the digital detector, and in a second refinement step of the computation iteration, the digital detector is configured to sample a respective bit line current of a respective r1″-th bit line of a MAC component of the first MAC circuit electrically coupled to the digital detector; wherein the digital detector is further configured to provide corresponding intermediate output data of a first refinement step in response to the bit line current sampled in the first refinement step of the computation iteration, and to provide corresponding intermediate output data of a second refinement step in response to the bit line current sampled in the second refinement step of the computation iteration; and The final output data of the final output signal generated by the first MAC circuit in the calculation iteration according to the r1-th layer of the first neural network depends on both the intermediate output data generated by the MAC component of the first MAC circuit in the first refinement step and the intermediate output data generated by the MAC component of the first MAC circuit in the second refinement step.
10. The IMC device of claim 9, wherein the word line activation circuit comprises a timer configured to provide a timer signal and a plurality of input-to-time converters coupled to the timer and each coupled to a corresponding symbol word line group, the input-to-time converters being configured to each compare the timer signal with a corresponding input data of the input data and in response each provide a corresponding activation signal to the corresponding symbol word line group.
11. The IMC device of claim 10, wherein each input to the time converter comprises: a data interface circuit configured to receive corresponding input data and identify the sign and value of the input data; a comparison circuit coupled to the data interface circuit and configured to receive at input the value of the corresponding input data and the timer signal, compare them with each other and generate, based on this comparison, a corresponding match signal indicating the fact that the timer signal is equal to the value of the input data or different therefrom; a set-reset circuit having a set input configured to receive a start signal indicating the start of a computation iteration, having a reset input coupled to the comparison circuit and configured to receive the match signal, and having an output, the set-reset circuit being configured to generate a wordline enable signal at the output based on the match signal and the start signal, the wordline enable signal indicating a value of the start signal and a value of the match signal; a first combinational logic circuit coupled to the output terminal of the set-reset circuit and coupled to the data interface circuit in a logically negated manner and configured to receive at the input terminal a word line enable signal and a logical inversion of a sign of input data, logically combine them with each other by AND logic and generate a corresponding intermediate drive signal for a positive word line of a corresponding sign word line group; as well as A second combinational logic circuit is coupled to the output terminal of the set-reset circuit and the data interface circuit and is configured to receive the word line enable signal and the sign of the input data at the input terminal, logically combine them with each other through AND logic, and generate a corresponding intermediate drive signal for the negative word line of the corresponding sign word line group.
12. The IMC device of claim 11, wherein each input to the time converter further comprises: a third combinatorial logic circuit coupled to the output terminal of the first combinatorial logic circuit and configured to receive at an input terminal the intermediate drive signal generated by the first combinatorial logic circuit and the address signal for the corresponding symbol word line group, logically combine them with each other by AND logic, and generate a drive signal for the positive word line of the symbol word line group; as well as a fourth combinatorial logic circuit, coupled to the output terminal of the second combinatorial logic circuit and configured to receive at the input terminal the intermediate drive signal generated by the second combinatorial logic circuit and the address signal for the corresponding symbol word line group, logically combine them with each other through AND logic, and generate a corresponding drive signal for the negative word line of the symbol word line group.
13. The IMC device according to claim 1, comprising: a first multiply and accumulate (MAC) circuit defined by a first portion of a selector, a corresponding digital detector, a bit line coupled to the first portion of the selector, and memory cells along the bit line coupled to the first portion of the selector, the first MAC circuit configured to perform a multiply and accumulate (MAC) operation associated with a first neural network defined by computed weights of the memory cells along the bit line coupled to the first portion of the selector and including a number R1 of layers; as well as at least one second multiply and accumulate (MAC) circuit defined by the second portion of the selector, the corresponding digital detector, the bit line coupled to the second portion of the selector, and the memory cells along the bit line coupled to the second portion of the selector, the at least one second MAC circuit configured to perform a multiply and accumulate (MAC) operation associated with at least one second neural network defined by the computed weights of the memory cells along the bit line coupled to the second portion of the selector and including a number R2 of layers; wherein each selector, a corresponding digital detector, a corresponding bit line coupled to the selector, and a portion of a memory cell arranged along the bit line coupled to the selector form a corresponding multiply and accumulate (MAC) component of the first MAC circuit or the at least one second MAC circuit; wherein each MAC component of the first MAC circuit has a respective number R1 of said bit lines coupled to respective selectors in such a way that the calculated weights of the memory cells of the respective r1-th bit lines of the MAC components of the first MAC circuit together define a respective r1-th layer of the first neural network, the number R1 of bit lines coupled to the respective selectors being equal to the number R1 of layers of the first neural network, wherein for each MAC component of the first MAC circuit, r1=1, ..., R1; wherein each MAC component of the at least one second MAC circuit has a respective number R2 of said bit lines coupled to respective selectors in such a way that the calculated weights of the memory cells of the respective r2-th bit lines of the MAC components of the at least one second MAC circuit together define a respective r2-th layer of the second neural network, the number R2 of bit lines coupled to the respective selectors being equal to the number R2 of layers of the second neural network, wherein for each MAC component of the at least one second MAC circuit, r2=1, ..., R2; wherein the number K of MAC components is equal to or greater than the larger of a maximum number of final output data of a final output signal generated by the first MAC circuit and a maximum number of final output data of a final output signal generated by the at least one second MAC circuit; wherein for each layer of the first neural network, each MAC component of the first MAC circuit is configured to generate respective intermediate output data of the intermediate output data, and the final output data of the final output signal generated by the first MAC circuit according to the layer of the first neural network depends on the respective intermediate output data generated by the MAC component of the first MAC circuit according to the layer of the first neural network; as well as Wherein for each layer of the at least one second neural network, each MAC component of the at least one second MAC circuit is configured to generate corresponding intermediate output data of intermediate output data, and the final output data of the final output signal generated by the at least one second MAC circuit according to the layer of the at least one second neural network depends on the corresponding intermediate output data generated by the MAC components of the at least one second MAC circuit according to the layer of the at least one second neural network.
14. The IMC device according to claim 13: wherein the input data are in a signed format and each input data comprises a respective sign bit indicating a sign of the input data and a respective plurality of value bits indicating a value of the input data; wherein the memory cells along each bit line are grouped in pairs in such a manner that each memory cell is part of only one of the pairs, each of said pairs of memory cells along the same bit line defining a respective memory group configured to store respective signed computation weights; wherein for each memory group, a corresponding word line coupled to one of the memory cells of the memory group defines a positive word line and a corresponding word line coupled to another memory cell of the memory group defines a negative word line, the positive word line and the negative word line connected to the memory group together forming a corresponding sign word line group; wherein the wordline activation circuit is configured to provide a wordline activation signal to the symbol wordline group in at least one computation iteration including a corresponding first elaboration step and a corresponding second elaboration step; The wordline activation circuit is configured to: in a first refinement step of a computation iteration, providing a first plurality of said wordline activation signals to a sign wordline group based on respective input data, each wordline activation signal of said first plurality of wordline activation signals being provided to one of a positive wordline and a negative wordline of a respective sign wordline group in accordance with respective input data of the input data, one of the positive wordline and the negative wordline of the respective sign wordline group being selected based on a sign of the respective input data and a refinement signal indicating performance of the first refinement step; as well as in a second refinement step of the computation iteration, providing a second plurality of the wordline activation signals to the sign wordline group based on the corresponding input data, each wordline activation signal of the second plurality of wordline activation signals being provided to the other of the positive wordline and the negative wordline of the corresponding sign wordline group in accordance with the corresponding input data, the other of the positive wordline and the negative wordline of the corresponding sign wordline group being selected based on the sign of the corresponding input data and the refinement signal indicating the performance of the second refinement step; wherein in the computation iteration, the selectors of the first MAC circuit are each configured to select a corresponding r1-th bit line among corresponding bit lines of the MAC components of the first MAC circuit, or not select any bit line; wherein in both the first refinement step and the second refinement step of the computation iteration, the digital detector is configured to sample a corresponding bit line current of a corresponding r1-th bit line of a MAC component passing through a first MAC circuit electrically coupled to the digital detector; wherein the digital detector is further configured to provide corresponding intermediate output data of a first refinement step in response to the bit line current sampled in the first refinement step of the computation iteration, and to provide corresponding intermediate output data of a second refinement step in response to the bit line current sampled in the second refinement step of the computation iteration; and The final output data of the final output signal generated by the first MAC circuit in the calculation iteration according to the r1-th layer of the first neural network depends on both the intermediate output data generated by the MAC component of the first MAC circuit in the first refinement step and the intermediate output data generated by the MAC component of the first MAC circuit in the second refinement step.
15. The IMC device of claim 14 , wherein the word line activation circuit comprises a timer configured to provide a timer signal and a plurality of input-to-time converters coupled to the timer and each coupled to a corresponding symbol word line group, the input-to-time converters being configured to each convert the timer signal to a corresponding input data (x n ) are compared and in response each provides a corresponding activation signal to the corresponding symbol word line group.
16. The IMC device of claim 15, wherein each input to the time converter comprises: a data interface circuit configured to receive corresponding input data and identify the sign and value of the input data; a comparison circuit coupled to the data interface circuit and configured to receive at input the value of the corresponding input data and the timer signal, compare them with each other and generate, based on this comparison, a corresponding match signal indicating the fact that the timer signal is equal to the value of the input data or different therefrom; a set-reset circuit having a set input configured to receive a start signal indicating the start of a computation iteration, having a reset input coupled to the comparison circuit and configured to receive the match signal, and having an output, the set-reset circuit being configured to generate a wordline enable signal at the output based on the match signal and the start signal, the wordline enable signal indicating a value of the start signal and a value of the match signal; an enable logic circuit coupled to the data interface circuit and configured to receive the sign of the input data and the refinement signal at an input terminal and combine them with each other by an exclusive-OR logic; a first combinational logic circuit coupled to an output terminal of the set-reset circuit and an output terminal of the enable logic circuit and configured to receive at an input terminal a word line enable signal and an output of the enable logic circuit, logically combine them with each other by AND logic, and generate a corresponding intermediate drive signal for a positive word line of a corresponding sign word line group; as well as A second combinational logic circuit is coupled to the output terminal of the set-reset circuit and to the output terminal of the enable logic circuit in a logically negated manner and is configured to receive the word line enable signal and the logical inversion of the output of the enable logic circuit at the input terminal, logically combine them with each other through AND logic, and generate a corresponding intermediate drive signal for the negative word line of the corresponding sign word line group.
17. The IMC device of claim 16, wherein each input to the time converter further comprises: a third combinatorial logic circuit coupled to the output terminal of the first combinatorial logic circuit and configured to receive at an input terminal the intermediate drive signal generated by the first combinatorial logic circuit and the address signal for the corresponding symbol word line group, logically combine them with each other by AND logic, and generate a drive signal for the positive word line of the symbol word line group; as well as a fourth combinatorial logic circuit, coupled to the output terminal of the second combinatorial logic circuit and configured to receive at the input terminal the intermediate drive signal generated by the second combinatorial logic circuit and the address signal for the corresponding symbol word line group, logically combine them with each other through AND logic, and generate a corresponding drive signal for the negative word line of the symbol word line group.
18. The IMC device according to claim 13: wherein the input data are in a signed format, and each input data comprises a corresponding sign bit indicating a sign of the input data and a corresponding plurality of value bits indicating a value of the input data; wherein the memory cells of each MAC component are grouped into groups of two respective pairs in such a manner that each memory cell is part of only one of the groups, each pair being arranged along a respective bit line of the MAC component and the memory cells of the two pairs being arranged along two respective word lines of the word lines, each of said groups of memory cells defining a respective memory group configured to store a respective signed computation weight; wherein for each memory group, respective word lines coupled to two of the memory cells of the memory group define a positive word line and respective word lines coupled to another two of the memory cells of the memory group define a negative word line, the positive word lines and the negative word lines connected to the memory group together forming a respective sign word line group; wherein the wordline activation circuit is configured to provide a wordline activation signal to the symbol wordline group in at least one computation iteration including a corresponding first elaboration step and a corresponding second elaboration step; The wordline activation circuit is configured to: In a first refinement step of a computation iteration, providing a first plurality of said wordline activation signals to a sign wordline group based on corresponding input data, each wordline activation signal of said first plurality of wordline activation signals being provided to one of a positive wordline and a negative wordline of a corresponding sign wordline group according to a corresponding input data of the input data, one of the positive wordline and the negative wordline of the corresponding sign wordline group being selected based on a sign of the corresponding input data; as well as In a second refinement of the computation iteration, a second plurality of the wordline activation signals are provided to the sign wordline group based on the corresponding input data, each wordline activation signal of the second plurality of wordline activation signals being provided to the other of the positive wordline and the negative wordline of the corresponding sign wordline group according to the corresponding input data, the other of the positive wordline and the negative wordline of the corresponding sign wordline group being selected based on the sign of the corresponding input data; and wherein in the computation iterations, the selectors of the first MAC circuit are each configured to select r1′-th and r1″-th bit lines or not select any bit lines among corresponding bit lines of the MAC component of the first MAC circuit in the first refinement step and the second refinement step, respectively, wherein the r1′-th and r1″-th bit lines are bit lines coupled to the same memory group implementing the r1-th layer of the first neural network; wherein in a first refinement step of the computation iteration, the digital detector is configured to sample a respective bit line current of a respective r1′-th bit line of a MAC component of a first MAC circuit electrically coupled to the digital detector, and in a second refinement step of the computation iteration, the digital detector is configured to sample a respective bit line current of a respective r1″-th bit line of a MAC component of the first MAC circuit electrically coupled to the digital detector; wherein the digital detector is further configured to provide corresponding intermediate output data of a first refinement step in response to the bit line current sampled in the first refinement step of the computation iteration, and to provide corresponding intermediate output data of a second refinement step in response to the bit line current sampled in the second refinement step of the computation iteration; and The final output data of the final output signal generated by the first MAC circuit in the calculation iteration according to the r1-th layer of the first neural network depends on both the intermediate output data generated by the MAC component of the first MAC circuit in the first refinement step and the intermediate output data generated by the MAC component of the first MAC circuit in the second refinement step.
19. The IMC device of claim 18, wherein the word line activation circuit comprises a timer configured to provide a timer signal and a plurality of input-to-time converters coupled to the timer and each coupled to a corresponding symbol word line group, the input-to-time converters being configured to each compare the timer signal with a corresponding input data of the input data and in response each provide a corresponding activation signal to the corresponding symbol word line group.
20. The IMC device of claim 19, wherein each input to the time converter comprises: a data interface circuit configured to receive corresponding input data and identify the sign and value of the input data; a comparison circuit coupled to the data interface circuit and configured to receive at input the value of the corresponding input data and the timer signal, compare them with each other and generate, based on this comparison, a corresponding match signal indicating the fact that the timer signal is equal to the value of the input data or different therefrom; a set-reset circuit having a set input configured to receive a start signal indicating the start of a computation iteration, having a reset input coupled to the comparison circuit and configured to receive the match signal, and having an output, the set-reset circuit being configured to generate a wordline enable signal at the output based on the match signal and the start signal, the wordline enable signal indicating a value of the start signal and a value of the match signal; a first combinational logic circuit coupled to the output terminal of the set-reset circuit and coupled to the data interface circuit in a logically negated manner and configured to receive at the input terminal a word line enable signal and a logical inversion of a sign of input data, logically combine them with each other by AND logic and generate a corresponding intermediate drive signal for a positive word line of a corresponding sign word line group; as well as A second combinational logic circuit is coupled to the output terminal of the set-reset circuit and the data interface circuit and is configured to receive the word line enable signal and the sign of the input data at the input terminal, logically combine them with each other through AND logic, and generate a corresponding intermediate drive signal for the negative word line of the corresponding sign word line group.
21. The IMC device of claim 20, wherein each input to the time converter further comprises: a third combinatorial logic circuit coupled to the output terminal of the first combinatorial logic circuit and configured to receive at an input terminal the intermediate drive signal generated by the first combinatorial logic circuit and the address signal for the corresponding symbol word line group, logically combine them with each other by AND logic, and generate a drive signal for the positive word line of the symbol word line group; as well as a fourth combinatorial logic circuit, coupled to the output terminal of the second combinatorial logic circuit and configured to receive at the input terminal the intermediate drive signal generated by the second combinatorial logic circuit and the address signal for the corresponding symbol word line group, logically combine them with each other through AND logic, and generate a corresponding drive signal for the negative word line of the symbol word line group.
22. The IMC device of claim 1, further comprising a control circuit operably coupled to the word line activation circuit, the selector, and the digital detector, and configured to control the word line activation circuit and the selector, receive intermediate output data from the digital detector, and generate a final output signal based on the intermediate output data.
23. The IMC device of claim 1, wherein the memory cells are non-volatile memory cells, and wherein each memory cell has a current path including a storage element and a selection element and extending between a common node and a reference potential node, the selection element being configured to selectively close a corresponding current path according to a corresponding activation signal.
24. The IMC device according to claim 1, further comprising: a bias circuit configured to generate a bias voltage to be applied to the bit line; A control circuit system is configured to: controlling the word line activation circuit to provide a plurality of activation signals to the memory cell, each activation signal being a function of a corresponding input value; Controlling the selector to select one of the corresponding bit lines or not select any bit line; The digital detector is controlled to sample a corresponding bit line current passing through a corresponding bit line electrically coupled to the digital detector, and to provide corresponding intermediate output data in response to the sampled bit line current.
Citation Information
Cited By
In-memory computation device with at least an improved digital detector for a more accurate current measurement
US20250046371A1
