Three-dimensional memory structure for in-memory computing
A three-dimensional structure of resistive memory cells with programmable logic functions addresses the challenges of circuit complexity and energy consumption in in-memory computing, enabling efficient parallel processing and increased storage and computing capacities.
Patent Information
- Application Number
- FR2022000963
- Authority / Receiving Office
- FR · FR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2022-02-03
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2042-02-03
AI Technical Summary
Existing non-volatile memory technologies face challenges such as increased circuit complexity, area, and energy consumption due to large peripheral circuits and limitations in calculation precision, especially for parallel operations, and current solutions for in-memory computing suffer from reduced performance and programming constraints.
A three-dimensional structure of resistive, non-volatile, programmable memory cells with specific programming sequences allows for elementary logic functions and data transfer without destructive operations, reducing peripheral circuit complexity and enabling broader programming flexibility.
The solution achieves reduced circuit area and energy consumption, supports complex algorithms, and enhances calculation performance by allowing parallel data processing with increased storage and computing capacities, compatible with semiconductor manufacturing techniques.
Smart Images

Figure 00000033_0000 
Figure 00000033_0001 
Figure 00000034_0000
Abstract
Description
Title of the invention: Three-dimensional memory structure for computing in memory Scope of application
[0001] The present invention relates to the field of designing computational memory circuits or in-memory computing circuits. More particularly, the invention relates to a non-volatile resistive memory circuit. The circuit makes it possible to carry out elementary Boolean logic operations within the structure of the memories making up the circuit. Technical problem
[0002] Generally speaking, a system on a chip comprises memory circuits for storing data and calculation circuits for performing logical and arithmetic operations. Data exchange operations between the memory circuits and the calculation circuits are very energy-intensive and reduce the calculation performance of the system. In this context, emerging non-volatile memory technologies make it possible to produce memory circuits capable of performing both the storage function and the calculation function. In this case, we speak of a "computation circuit in the memory". This makes it possible to avoid data transfers between the storage means and the calculation means, and thus, makes it possible to reduce energy consumption and improve the calculation speed.
[0003] However, the development of “in-memory computing circuits” solutions presents several challenges and constraints such as: - The use of large peripheral circuits around the memories to process the results. These peripheral circuits cover, for example, a plurality of detection amplifiers. This considerably increases the circuit area and the physical implementation complexity. - Constraints on the accuracy of the calculations performed. More specifically, this problem occurs when executing algorithms requiring parallel calculations.
[0004] Thus, there is a need to design new architectures for computing circuits in memory, which are less complex and at the same time present a reduction in electrical consumption and an improvement in computing performance in terms of speed and precision. Prior Art / State of the Art Restrictions
[0005] The scientific publication [1] describes three general approaches to performing “in-memory computing” operations. All three approaches rely on programming operands within resistive memory cells that allow return to perform sums and products using Kirchhoff's laws. The first solution is called "non-stateful logic" and the second solution is called "matrix-vector-multiplication". Both solutions require the use of large peripheral circuits to the memories in order to process the raw results. They therefore have the disadvantages associated with peripheral circuits. These disadvantages consist of implementation complexity, large area and increased power consumption. The third solution is called "stateful logic". This solution does not require large peripheral circuits. However, it has the disadvantage of reducing the computational accuracy when executing algorithms requiring parallel computation operations.
[0006] The scientific publication [2] proposes an elementary calculation structure in resistive memory made by stacking a plurality of resistive memories controlled by a common FinFET transistor. This solution has the disadvantage of a limitation in the number of memories stacked in an elementary structure due to the increase in loss currents. In addition, this solution has programming constraints because the operands and the results of a logic operation must be stored within the same elementary structure. This reduces the possibilities of programming the calculation circuit in the memory and thus reduces the computational performance of the system. Response to the problem and provision of a solution
[0007] To overcome the limitations of existing solutions regarding the design of calculation circuits in memory, the invention proposes a circuit based on a three-dimensional structure composed of resistive, non-volatile and programmable memory cells. In addition, the invention proposes specific programming sequences making it possible to carry out elementary logic functions and data transfer functions in a non-destructive manner. It is thus possible for the person skilled in the art to implement more complex algorithms from the elementary logic functions and the data transfer function.
[0008] The device according to the invention makes it possible to reduce the complexity of the peripheral circuits by limiting it to a few control transistors. The calculation operations do not require the use of detection amplifier type circuits. This reduces the surface area of the circuit and its energy consumption and simplifies its physical implementation.
[0009] The device according to the invention further allows operation to be obtained in the three spatial dimensions without the constraints of storage location of the operands and results. This offers greater programming flexibility than solutions where the calculations are located in a single elementary structure.
[0010] The device according to the invention can be produced with a number of memory cells considerably higher than that of known solutions. This offers the advantage of multiplying storage and computing capacities at the same time. This makes it possible to obtain a device compatible with the execution of complex calculation algorithms within the memory structure. In addition, the device is capable of parallelizing a considerable number of data processing operations to achieve better performance.
[0011] The invention further relates to a thin-film structure for producing the three-dimensional memory structure according to the invention. This elementary structure has the advantage of compatibility with the manufacturing techniques of the semiconductor industry. Thus, the device according to the invention has reduced production costs. Summary / Claims
[0012] The subject of the invention is a computing circuit in the memory comprising a plurality of memory planes of rank i ranging from 1 to N, with N a natural integer greater than 1, each plane forming a two-dimensional matrix of non-volatile, resistive and programmable memory cells, each memory cell having a selection node, a first input / output node and a second input / output node. Said matrix comprising M rows of rank j ranging from 1 to M and K columns of rank k ranging from 1 to K, with M and K two non-null natural integers. Said computing circuit comprising: - control means for controlling storage and calculation operations in memory carried out by the calculation circuit; - at least one elementary group of memory cells comprising: • a first memory cell belonging to any one of the memory planes and intended to store a first input data item; • a second memory cell belonging to any one of the memory planes and intended to store a second input data; • a third memory cell belonging to a memory plane different from that of the first and second memory cells. The third memory cell is intended to store the result of a first logical operation having as operands the first and second input data. The first, second and third memory cells belonging to columns of the same rank k in their respective matrices. The first, second and third memory cells being interconnected by a common interconnection line, in particular in order to form a resistive voltage divider bridge.
[0013] According to a particular aspect of the invention, for each memory plane: the selection nodes of memory cells belonging to the same row are inter- connected by a selection line; the first input / output nodes of memory cells belonging to the same column are interconnected; the second input / output nodes of memory cells belonging to the same column are interconnected.
[0014] According to a particular aspect of the invention, each memory cell comprises: - a programmable resistive storage structure having: • an upper electrode connected to the second input / output node of said memory cell; • and a lower electrode; - a selection transistor having a gate connected to the selection node of said memory cell and connecting the lower electrode to the first input / output node of said memory cell.
[0015] According to a particular aspect of the invention, each column of memory cells comprises: - A stack of a plurality of selection transistors in a first direction, each selection transistor, in particular of the wrapping gate type, comprising a conduction channel, perpendicular to the first direction in a semiconductor material and having two ends; the first end of the channel corresponding to the source of the selection transistor and the second corresponding to the drain of the transistor; - a first metal pillar in the first direction connecting the sources of the different selection transistors; - a dielectric layer parallel to the first direction and laterally covering the drains of the selection transistors; - at least one metallic layer in parallel to the first direction deposited on the dielectric layer; - a second metal pillar in the first direction having lateral contact with the at least one metal layer; For each selection transistor, the assembly formed by the drain, the dielectric layer and the metal layer constitutes a resistive storage structure.
[0016] According to a particular aspect of the invention, for each memory plane: the first input / output nodes of the memory cells belonging to the odd-rank columns are interconnected forming a first source line; the first input nodes of the memory cells belonging to the columns of even ranks are interconnected forming a second source line; the first source line and the second source line being distinct.
[0017] According to a particular aspect of the invention, the elementary group of memory cells further comprises a fourth memory cell belonging to a memory plane different from that of the first, second and third memory cells. The fourth memory cell is intended to store the result of a second logical operation having as operands the first and the second input data and using the result of the first logical operation stored in the third memory cell. The fourth memory cell is connected to the first, to the second and to the third memory cell via the common interconnection line.
[0018] According to a particular aspect of the invention, for each elementary group of memory cells, the interconnection line connects the second input / output nodes of the memory cells of said elementary group to each other.
[0019] According to a particular aspect of the invention, the first logical operation is a Boolean OR function.
[0020] According to a particular aspect of the invention, the control means are configured to carry out the following steps, in the order of enumeration, in order to carry out the OR logic function: i. Initialize the third cell to a high resistive state; ii. selecting the first and third memory cells by turning on their selection transistors; iii. applying a first positive read voltage to the first input / output node of the first memory cell and simultaneously connecting the first input / output node of the third memory cell to electrical ground; iv. select the second and third memory cells by turning on their selection transistors; v. applying the first positive read voltage to the first input / output node of the first memory cell and simultaneously connecting the first input / output node of the third memory cell to electrical ground; Each memory cell having: a first voltage threshold for switching from a high resistive state to a low resistive state; and a second voltage threshold for switching from a low resistive state to a high resistive state greater than the first voltage threshold. Said first read voltage being positive with an amplitude greater than or equal to the first switching threshold and strictly less than the second switching threshold.
[0021] According to a particular aspect of the invention, the second logical operation is an exclusive NOR Boolean function.
[0022] According to a particular aspect of the invention, the control means are configured to carry out the following steps, in the order of enumeration, in order to carry out the exclusive NOR logic function: i. Initialize the fourth cell to a low resistive state; ii. select the first, third and fourth memory cells by turning on their transistors; iii. carry out simultaneously: a. applying a second positive read voltage to the first input / output node of the first memory cell; b. applying a third negative read voltage to the first input / output node of the third memory cell; c. and applying a fourth positive read voltage to the first input / output node of the fourth memory cell; iv. select the second, third and fourth memory cells by turning on their selection transistors; v. carry out simultaneously: a. applying the second positive read voltage to the first input / output node of the second memory cell; b. applying the third negative read voltage to the first input / output node of the third memory cell; c. and applying a fourth positive read voltage to the first input / output node of the fourth memory cell; The absolute value of the third reading voltage being substantially equal to the absolute value of the second reading voltage. The sum of the absolute value of the fourth reading voltage and the second reading voltage being greater than the second crossing threshold.
[0023] According to a particular aspect of the invention, the first logical operation is a Boolean AND function.
[0024] According to a particular aspect of the invention, the calculation circuit in the memory further comprises for each interconnection line a regulation transistor; each regulation transistor having a gate receiving a regulation signal and connecting the associated interconnection line to ground so as to produce a variable impedance.
[0025] According to a particular aspect of the invention, the calculation circuit in the memory comprises a plurality of elementary groups of memory cells. Each interconnection line is connected to an adjacent interconnection line via a switch.
[0026] According to a particular aspect of the invention, the control means are configured to carry out the transfer of a data item stored from a starting memory cell to an arrival memory cell belonging to a plane different from that of the cell starting memory and connected to an interconnection line adjacent to that of the starting memory cell. The transfer is carried out by performing, in the order of enumeration, the following steps: i'- initialize the arrival cell to a high resistive state; ii'- put the switch in the on state; iii' - simultaneously carry out: a. applying a first positive transfer voltage to the first input / output node of the starting memory cell; b. connecting the first input / output node of the arriving memory cell to electrical ground; c. applying a second positive transfer voltage to the first input / output node of the memory cell belonging to the same plane of the starting memory cell and connected to the same interconnection line of the arriving memory cell; d. applying the second positive transfer voltage to the first input / output node of the memory cell belonging to the same plane of the arrival memory cell and connected to the same data line (or bit line) of the departure memory cell; The amplitude of the second transfer voltage being less than or equal to half of the first transfer voltage.
[0027] According to a particular aspect of the invention, for each elementary group of memory cells, the interconnection line connects the first input / output nodes of the memory cells of said elementary group to each other.
[0028] The term "common interconnection line" means a conductive line connecting the same nodes of a plurality of memory cells distributed over at least two memory planes. The connection is made in a direct manner and without an intermediary. According to a particular aspect of the invention, the common interconnection line corresponds to the common data line (bit line in English). According to a particular aspect of the invention, the common interconnection line corresponds to the common source line (source line in English).
[0029] According to a particular aspect of the invention, the memory circuit comprises a plurality of memory planes of rank i ranging from 1 to N, with N an integer greater than or equal to 2. Detailed description
[0030] Other characteristics and advantages of the present invention will appear better on reading the description which follows in relation to the following appended drawings.
[0031] [Fig.1a] [Fig.1a] illustrates a sectional view of a resistive storage structure programmable.
[0032] [Fig. 1b] [Fig. 1b] illustrates a memory cell used to produce the memory calculation circuit according to the invention.
[0033] [Fig.2a] [Fig.2a] illustrates an elementary group of memory cells according to a first embodiment of the invention intended to carry out the logical operation A OR B.
[0034] [Fig.2b] [Fig.2b] illustrates the steps to perform the logical operation A OR B with the elementary group of memory cells according to the first embodiment of the invention.
[0035] [Fig.3a] [Fig.3a] illustrates an elementary group of memory cells according to a second embodiment of the invention intended to carry out the logical operation A XNORB.
[0036] [Fig.3b] [Fig.3b] illustrates the steps to perform the logical operation A XNOR B with the elementary group of memory cells according to the second embodiment of the invention.
[0037] [Fig.4] [Fig.4] illustrates the steps to perform the logical operation A AND B with the elementary group of memory cells according to the first embodiment of the invention.
[0038] [Fig.5a] [Fig.5a] illustrates a first embodiment of the calculation circuit in the memory according to the invention.
[0039] [Fig.5b] [Fig.5b] illustrates a sectional view of the thin-film structure of a column of the calculation circuit in the memory according to the invention.
[0040] [Fig.5c] [Fig.5c] illustrates a second embodiment of the calculation circuit in the memory according to the invention.
[0041] [Fig.6a] [Fig.6a] illustrates a third embodiment of the calculation circuit in the memory according to the invention.
[0042] [Fig.6b] [Fig.6b] illustrates an elementary group of memory cells according to a fourth embodiment of the invention intended to carry out a data transfer.
[0043] [Fig.6c] [Fig.6c] illustrates the steps to perform a data transfer with the elementary group of memory cells according to the fourth embodiment of the invention.
[0044] [Fig.7] [Fig.7] illustrates the steps executed by the calculation circuit according to the invention to calculate the hypervector product of three data vectors.
[0045] [Fig.8] [Fig.8] illustrates an elementary group of memory cells according to a third embodiment of the invention intended to carry out the logical operation A OR B.
[0046] We will begin by describing the operating principle of a programmable resistive data storage structure. [Fig.1a] illustrates a sectional view of a programmable resistive storage structure. The storage structure SI consists of the stack of thin layers in the following order: at least a first layer Cl made of an electrically conductive material forming a lower electrode EL2; a second layer C2 made of a dielectric material and at least a third layer C3 made of electrically conductive material forming an upper electrode ELI. The characteristics of the materials constituting the NVM memory structure layers make it possible to obtain operation in resistive memory with variable conductive filament. The second layer C2 is referred to in the remainder of the description as the “central layer”.
[0047] Alternatively, it is possible to produce each ELI and / or EL2 electrode by a stack of several conductive layers.
[0048] By way of example and without loss of generality, the first layer C1 (and therefore the lower electrode EL2) and the third layer C3 (and therefore the upper electrode ELI) are made of titanium nitride TiN. The thickness of each of the layers C1(EL2) and C3(ELI) is of the order of a few tens of nanometers, more particularly equal to 100nm.
[0049] By way of example and without loss of generality, the central layer C2 is made of hafnium oxide. The thickness of the central layer C2 is of the order of a few nanometers, more particularly equal to 10 nm.
[0050] The operation of a resistive SI storage structure requires the formation of a conductive filament through at least a portion of the electrically insulating central layer C2.
[0051] Initially, the storage structure SI is a MIM (metal, insulator, metal) type structure having an infinite resistance between the two electrodes ELI and EL2. First, the filament F must be formed through at least a portion of the volume of the central layer C2. The formation of the filament makes it possible to obtain a variable resistance by modulating the length 1 of the formed conductive filament. To form the filament, a positive electrical formation voltage is applied to the upper electrode ELI. The electrical formation voltage has a sufficiently high amplitude and / or duration to cause the generation of oxygen vacancies in the central layer C2. Indeed, the applied electrical formation voltage must exceed a predetermined value so as to tear oxygen ions from the crystal lattice of the central metal oxide layer.The ions will merge towards the upper electrode EL2 to form a conductive filament F through the central layer made up of oxygen vacancies.
[0052] Once the conductive filament F is formed, the behavior of a resistive element with a variable resistance R is obtained depending on the length of the conductive filament F. When the electrical potential of the upper electrode VELi is lower than that of the electrode lower electrode VEL2> the storage structure SI sees a negative voltage Vreset across its terminals. In this case, oxygen ions will fill some of the oxygen vacancies forming the conductive filament. This results in a reduction in the length of the conductive filament. Thus, the resistance of the resistive element increases. This is called a high resistive state and a RESET type write operation. Conversely, when the electrical potential of the upper electrode VELi is higher than that of the lower electrode VEL2> the storage structure SI sees a positive voltage Vset across its terminals. The length of the conductive filament F increases by the same mechanism described for the wire formation operation. Thus, the resistance of the resistive element decreases. This is called a low resistive state and a SET type write operation.
[0053] The applied voltage Vset must be greater in absolute value than a first voltage threshold to switch from a high resistive state to a low resistive state. Similarly, the applied voltage Vreset must be greater in absolute value than a second voltage threshold to switch from a low resistive state to a high resistive state.
[0054] The following convention is chosen as an example: when the storage structure SI is configured to store binary data in the high logic state (x= 1) it is in a low resistive state. Conversely, when the storage structure SI is configured to store binary data in the low logic state (x= 0) it is in a high resistive state. An inverse convention is also possible.
[0055] [Fig.lb] illustrates an electrical diagram of a memory cell CM according to the invention. The memory cell CM comprises a memory structure SI and a selection transistor TL. The memory cell CM further comprises a selection node WL, a first input / output node SL and a second input / output node BL. The upper electrode ELI of the storage structure SI is connected to the second input / output node BL of the memory cell CM. The selection node WL is intended to receive a selection signal VWL. The gate of the transistor T1 is connected to the selection node WL. The transistor T2 is mounted so as to connect the lower electrode EL2 to the first input / output node SL. Thus, the memory cell CM can only be written in SET or RESET when the transistor T1 is in the on state. In addition, the transistor T1 makes it possible to limit the current passing through the storage structure SI during a SET operation.Current limitation protects the S1 structure from the risk of internal structural destruction due to the presence of high potential currents during the transition from a high resistive state to a low resistive state.
[0056] Generally, it is possible to produce, within the framework of the invention, memory cells of the OxRAM, CbRAM or any other non-volatile resistive memory technology.
[0057] In general, the invention relates to a calculation circuit in the memory comprising a plurality of memory planes of ranks i from 1 to N, with N an integer. natural greater than 1. Each plane forms a two-dimensional matrix of memory cells CM as described above. Each matrix comprises M rows and K columns of memory cells. This gives a three-dimensional memory structure. The three-dimensional memory structure according to the invention is characterized by the implementation of at least one elementary group of several memory cells capable of performing at least one logical operation. We will describe in the following several non-limiting examples of embodiments of elementary groups of memory cells.
[0058] [Fig.2a] illustrates an elementary group of memory cells according to a first embodiment of the invention intended to carry out the logical operation A OR B. The elementary group is composed of three memory cells CMACMB, CM0R. The first memory cell CMA belongs to the memory plane Pini and is intended to store a first input data item A. The second memory cell CMB belongs to the memory plane Pini and is intended to store a second input data item B. The third memory cell CMOR belongs to a memory plane P0Ri different from the memory plane Pini. The third memory cell is intended to store the result of the logical operation A OR B.
[0059] In the example illustrated, each of the Piniet memory planes P0Ri is a matrix of memory cells of four rows (Lb L2, L3, L4) and two columns (Ci, C2). In each memory plane: - the selection nodes WL of the memory cells belonging to the same line Lj are interconnected by a selection line WLÿ; - the first input / output nodes SL of the memory cells belonging to the same column Ck are interconnected through a common source line; - the second input / output nodes BL of the memory cells belonging to the same column Ck are interconnected by a common data line (bit line in English).
[0060] Thus, each memory plane comprises M selection lines WLÿ (M being the number of lines per plane), K source lines and K data lines (K being the number of columns per plane).
[0061] In the elementary group illustrated, the first memory cell CMA, the second memory cell CMB and the third memory cell CM0R all belong to columns of the same rank k in their respective matrices. The second input / output nodes BL of the first, second and third memory cells CMACMB and CMor are interconnected by a common interconnection line BL[a. The common interconnection line corresponds in this case to the data line (bit line in English) of the memory cells CMACMB and CM0R. This specific assembly combined with the application of a well-defined voltage sequence on the nodes of said cells makes it possible to carry out the logic function A OR B in the third cell.
[0062] Alternatively, it is possible to implement the two input cells CMA and CMB in two separate memory planes, provided that they belong to columns of the same rank k.
[0063] Thus, the elementary group is controlled via the following nodes: the first input / output node common to CMA and CMB noted SLU, the first input / output node of CMOr noted SL2a and the three selection nodes (WL11WL12 WL2ji).
[0064] Advantageously, it is possible to produce another elementary group of symmetrical memory cells on the second columns C2 of the planes Pini and P0Ri. This makes it possible to carry out two logical OR operations on two different pairs of operands in parallel.
[0065] [Fig.2b] illustrates the steps for carrying out the logical operation A OR B with the elementary group of memory cells according to the first embodiment of the invention.
[0066] The first step 100 consists of initializing the third CMOR cell to a high resistive state corresponding to a low logic state x=0.
[0067] Then, a first calculation cycle 200 is triggered. The cycle 200 comprises a first step 201 consisting of selecting the first memory cell CMA and the third memory cell CM0R by turning on their selection transistors TL. The second step 202 of the cycle 200 consists of applying a first positive reading voltage to the first input / output node SLkl of the first memory cell and simultaneously connecting the first input / output node SL2a of the third memory cell to electrical ground.
[0068] The amplitude of the first positive read voltage applied to the node SL[a is strictly less than the second voltage threshold Vreset to switch from a low resistive state to a high resistive state. This ensures the preservation of the resistive state of the first memory cell CMA containing the first input data. Two cases are possible for this configuration:
[0069] In the first case, the first cell CMA is considered to be in a low resistive state (A=1). This means that the voltage drop across said memory cell is low. The first memory cell CMA is comparable to a switch in the on state. The applied voltage is propagated to the second input / output node BL of the third memory cell CM0R via the common interconnection line BL[a. Thus, the third memory cell CM0R ends up with a positive voltage on the upper electrode ELI connected to the interconnection line BL[a, and a zero voltage on the lower electrode EL2 connected to ground through the transistor Tl associated with the on state. The third memory cell CM0R is thus subjected to a positive voltage across its terminals. The amplitude of the first positive reading voltage applied to the node SLU is strictly greater than the first voltage threshold Vset to switch from a high resistive state to a low resistive state. Thus, the third memory cell CM0R switches from a high resistive state (x=0) to a low resistive state (x=l) when (A=l).
[0070] As an example, for a given resistive memory technology, the first voltage threshold Vset is equal to 1.5V and the second voltage threshold VreSet is equal to 2V. The amplitude of the first positive read voltage applied to the node SLia is equal to 1.8V. The first read voltage VSLia is in the form of a rectangular pulse with a duration of 100ns for example.
[0071] In the second case, it is considered that the first cell CMA is in a high resistive state (A=0). This means that the voltage drop across said memory cell is high. The first memory cell CMA is comparable to a switch in the blocking state. The third memory cell CM0R is found with a positive voltage close to zero on the upper electrode ELI and a zero voltage on the lower electrode EL2 connected to ground through the transistor T1 associated with the on state. The third memory cell CM0R is thus subjected to a low voltage across its terminals and not sufficient to cause a change in resistive state. Thus, the third memory cell CM0R remains in its initially high resistive state (x=0) when (A=0).
[0072] Then, a second calculation cycle 300 is triggered. Cycle 300 comprises a first step 301 consisting of selecting the second memory cell CMB and the third memory cell CM0R by turning on their selection transistors TL. The second step 302 of cycle 300 consists of applying the first positive read voltage to the first input / output node SL[a of the second memory cell and simultaneously connecting the first input / output node SL2a of the third memory cell to electrical ground. This is a reiteration of cycle 200 but by selecting the second memory cell CMB containing the second input data B. Thus, by the same mechanism described in the first cycle 200, if the second memory cell CMB is in a low resistive state (B=1), then the third memory cell CM0R switches to a low resistive state (x=1).If the second memory cell CMB is in a high resistive state (B=0), then the third memory cell CM 0R remains in a high resistive state (x=0).
[0073] We thus obtain, by the preceding sequence of steps, the following logic function: the third memory cell CM0R remains in a high resistive state (x=0) if and only if the first and second memory cells CMBjCMA are both in a high resistive state. (A=0 and B=0). This corresponds to the logic function A OR B as illustrated in the logic table of [Fig.2b].
[0074] [Fig.3a] illustrates an elementary group of memory cells according to a second embodiment of the invention intended to carry out the logic operation A XNOR B. The second embodiment uses the same structure of the elementary group according to the first embodiment and further comprises a fourth memory cell CMxnor- The fourth memory cell CMxnor belongs to a memory plane Pxnori different from that of the memory planes Pini and PORi. The fourth memory cell is intended to store the result of the logic function (A XNOR B). The first input data A stored in CMA> the second input data B stored in CM B and the result of (A OR B) previously calculated and stored in CMOR are used to calculate A XNOR B in CMxnor.
[0075] The second input / output node BL of the fourth memory cell CMxnor is connected to the second input / output nodes BL of the first, second and third memory cells CMACMB and CM0R via the common interconnection line BL[a.
[0076] The fourth memory cell CMxNOR belongs to the column Ci in the matrix of the plane Pxnori. It is the column of the same rank as the memory cells CMACMB and CMor. This makes it possible to simplify the interconnection between the second input / output nodes BL of said memory cells through the interconnection line BL[a.
[0077] The elementary group of memory cells according to the second embodiment of the invention is configurable via: the first input / output node common to CMA and CMb denoted SLia, the first input / output node of CM0R denoted SL2a, the first input / output node of CMXNor denoted SL3a and the four selection nodes (WL11WL12 WL 2,1WL3j1).
[0078] Advantageously, it is possible to produce another elementary group of symmetrical memory cells on the second columns C2 of the planes Pini and P0Ri. This makes it possible to carry out two logical OR operations on two different pairs of operands in a parallel manner.
[0079] [Fig.3b] illustrates the steps for performing the logical operation (A XNOR B) with the elementary group of memory cells according to the second embodiment of the invention. To perform this calculation, the result of the logical operation (A OR B) must be previously calculated and stored in the third memory cell CM0R. For this, it is possible to execute the steps of the method described for the elementary group according to the first embodiment of the invention described previously. It is also possible to write the value of the logical operation (A OR B) in the third memory cell CM0R from an external computer.
[0080] The first step 400 consists of initializing the fourth cell CMxnor to a low resistive state corresponding to a high logic state x=l.
[0081] Then, a first calculation cycle 500 is triggered. Cycle 500 comprises a first step 501 consisting of selecting the first memory cell CMA, the third memory cell CM0R and the fourth memory cell CMXN0R by turning on their selection transistors T1. The second step 502 of cycle 500 consists of simultaneously carrying out:
[0082] applying a second positive read voltage VSLia to the first input / output node SLU of the first memory cell CMA;
[0083] applying a third negative read voltage VSL2a to the first input / output node SL2a of the third memory cell CM0R;
[0084] and applying a fourth positive read voltage VSL3a to the first input / output node SL3a of the fourth memory cell CMXXOR;
[0085] The absolute value of the second positive reading voltage VSLia is substantially equal to that of the third negative reading voltage VSL2a. Advantageously, the absolute value of the second positive reading voltage VSLia is slightly greater than that of the third negative reading voltage VSL2a. The difference in absolute value between VSLia and VSL2a is between 0% and 40% of the absolute value of the third reading voltage VSL2a. More particularly, the difference in absolute value between VSLia and VSL2a is between 25% and 35% of the absolute value of the third reading voltage VSL2a. On the other hand, the amplitude of the fourth positive reading voltage VSL3a is strictly less than the second voltage threshold Vreset to go from a low resistive state to a high resistive state. Three cases are possible for this configuration:
[0086] In the first case, it is considered that the first cell CMA is in a low resistive state (A=1) and the third cell CM0R is in a low resistive state (A OR B=1). The first memory cell CMA can be likened to a switch in the on state. The second positive read voltage VSLia is propagated to the second input / output node BL of the fourth memory cell CMXXOR via the common interconnection line BLia. Similarly, the third memory cell CM0R can be likened to a switch in the on state. The third negative read voltage VSL2a is propagated to the second input / output node BL of the fourth memory cell CMXXOh via the common interconnection line BLia. By application of the mesh law, the node BL of the fourth cell CMxNOK has a potential equal to the algebraic sum of VSL2a and VSLia, which is a few mV. We thus obtain the following equation: VBL^ = VSL}a + VSL2a ~ OVAincess, the fourth memory cell CMxnor sees the following voltage at these terminals: v XNOR =vbi^ - VSL* = (VSL^+VSL^ - » - vsl3 <o ; with | VSL3a j < VRESET The fourth memory cell CMxnor is thus subjected to a negative voltage Vxnor across its terminals. The amplitude of the negative voltage Vxnor is not high enough to exceed the second voltage threshold Vreset to switch from a low resistive state to a high resistive state. Thus, the fourth memory cell CMxnor remains in a low resistive state (x=l) when (A=l) and (A OR B=l).
[0087] In the second case, the first cell CMA is considered to be in a high resistive state (A=0) and the third cell CM0R is in a high resistive state (A OR B=0). None of the read voltages VSL2a and VSLi is propagated to the BL node of the fourth memory cell CMxnor. The fourth memory cell CMxNOR thus remains in its initial resistive state. The fourth memory cell CMxnor remains in a low resistive state (x=1) when (A=0) and (A OR B=0).
[0088] In the third case, it is considered that the first cell CMA is in a high resistive state (A=0) and the third cell CM0R is in a low resistive state (A OR B=1). The first memory cell CMA is comparable to a switch in the blocking state. The second positive read voltage VSL1a is not propagated to the second input / output node BL of the fourth memory cell CMxnor- Similarly, the third memory cell CMOR is comparable to a switch in the passing state. The third negative read voltage VSL2a is propagated to the second input / output node BL of the fourth memory cell CMxnor via the common interconnection line BL[a. The following equation is thus obtained: VBL^ ~ < OVAince, the fourth memory cell CMxnor sees the voltage following at these terminals: Vxnor —VBL}a-VSL3a = VSL^-VSL^Q-with | = | VA'L2J + > Vreset The third memory cell CM0R is thus subjected to a negative voltage Vxnor across its terminals. The amplitude of the negative voltage Vxnor is high enough to exceed the second voltage threshold Vreset- Thus, the fourth memory cell CMxnor goes from a low resistive state to a high resistive state (x=0) when (A=0) and (A OR B=1).
[0089] Then, a second calculation cycle 600 is triggered. Cycle 600 is a reiteration of cycle 500 but by selecting the second memory cell CMB containing the second input data B instead of the first memory cell CMA. Thus, by the same mechanism described in the first cycle 600, the following three cases are obtained: the fourth memory cell CMxnor goes from a low resistive state to a high resistive state (x=0) when (B=0) and (A OR B=1), otherwise the fourth memory cell CMxnor remains in a low resistive state (x=1).
[0090] We thus obtain, by the preceding sequence of steps, the following logic function: the fourth memory cell CMxNOR goes to a low resistive state (x=1) if and only if the first and second memory cells CMB CMA have resistive states opposite (A=0 and B=1 or A=1 and B=0). This corresponds to the logic function A XNOR B as illustrated in the logic table in [Fig.3b].
[0091] For example, for a given resistive memory technology, the first voltage threshold Vset is equal to 1.8V and the second voltage threshold VreSet is equal to 2V. The second read voltage VSLia is equal to IV. The third read voltage VSL2a is equal to -0.75V. The fourth read voltage VSL3a is equal to 1.5V. The read voltages VSLia, VSL2a and VSL3a are in the form of rectangular pulses with a duration of 100ns for example.
[0092] Advantageously, the selection voltage VWL3>i applied to the gate of the transistor T1 of the fourth memory cell CMxNOR is lower than the general supply voltage VDD. This makes it possible to limit the current through the memory cell during a transition. Thus, the memory cell is protected from the risk of deterioration due to the presence of high currents during a modification of its resistive state.
[0093] Furthermore, it is possible to realize the logic function (A AND B) by using the elementary group of memory cells according to the first embodiment. [Fig.4] illustrates the steps for realizing the logic operation (A AND B) with the elementary group of memory cells according to the first embodiment of the invention. In this illustration, the third memory cell is intended to store the result of the logic operation (A AND B) and it is called CanoSu instead of C0R. The connection between the three memory cells remains unchanged compared to [Fig.2a].
[0094] The first step 100' consists of initializing the third cell CManD to a low resistive state corresponding to a high logic state x=l.
[0095] Then, a first calculation cycle 200' is triggered. Cycle 200' comprises a first step 201' consisting of selecting the first memory cell CMA and the third memory cell CManD by turning on their selection transistors TL. The second step 202' of cycle 200 consists of applying a first positive read voltage VSLia to the first input / output node SL[a of the first memory cell and applying a second positive read voltage VSL2a to the third input / output node SL2a of the third memory cell CManD.
[0096] The amplitude of the first read voltage VSLia is strictly less than the second voltage threshold Vreset to switch from a low resistive state to a high resistive state. This ensures the preservation of the resistive state of the first memory cell CMA containing the first input data. The amplitude of the second read voltage VSL2a is strictly greater than the second voltage threshold Vreset to switch from a low resistive state to a high resistive state. Two cases are possible for this configuration:
[0097] In the first case, it is considered that the first CMA cell is in a low resistive state (A=1). This implies that the voltage drop across said memory cell is low. The first memory cell CMA is similar to an on-state switch. The applied voltage is propagated to the second input / output node BL of the third memory cell CMand via the common interconnect line BL[a. Thus, the third memory cell CMand sees the following voltage at these terminals: ^AND = VBLla~ VSL2a ~ ^^2« <0 ; with | RESET
[0098] The third memory cell CMand is thus subjected to a negative voltage Vand across its terminals. The amplitude of the negative voltage Vand is not high enough to exceed the second voltage threshold Vreset and move from a low resistive state to a high resistive state. Thus, the third memory cell CMand remains in a low resistive state (x=l) when (A=l).
[0099] In the second case, the first cell CMA is considered to be in a high resistive state (A=0). This means that the voltage drop across said memory cell is high. The first memory cell CMA is comparable to a switch in the blocking state. The second input / output node BL of the third memory cell CMand has a positive voltage VBLia close to zero. Thus, the third memory cell CMand sees the following voltage at these terminals: Vand = VSL2a ~ 0- V^2a<0; with I > Vreset because VBLia 0
[0100] The third memory cell CMand is thus subjected to a negative voltage Vand across its terminals. The amplitude of the negative voltage Vand is high enough to exceed the second voltage threshold Vreset. Thus, the third memory cell CMand goes to a high resistive state (x=0) when (A=0).
[0101] Then, a second calculation cycle 300' is triggered. Cycle 300' is a reiteration of cycle 200' but by selecting the second memory cell CMB containing the second input data B instead of the first memory cell CMA. Thus, by the same mechanism described in the first cycle 200', the following two cases are obtained: the third memory cell CMand goes from a low resistive state to a high resistive state (x=0) when (B=0), otherwise the third memory cell CMand remains in a low resistive state (x=1).
[0102] We thus obtain, by the preceding sequence of steps, the following logic function: the third memory cell CMANDemains in a low resistive state (x=l) if and only if the first and second memory cells CMB CMAare both in a low resistive state. (A=l and B=l). This corresponds to the logic function A AND B as illustrated in the logic table of [Fig.4].
[0103] [Fig.5a] illustrates a first embodiment of the calculation circuit in the memory 10 according to the invention. The calculation circuit 10 comprises a plurality of planes successive memory P;derangs i from 1 to N, with N a natural integer greater than 1. Each memory plane forms a two-dimensional matrix of memory cells CM according to the invention. Each matrix comprises M rows Lj and K columns Ck. As an illustrative example, the number of rows per matrix is equal to 4.
[0104] In each memory plane P;, the selection nodes WL of the memory cells belonging to the same line Lj are interconnected by a common selection line WLÿ.
[0105] In each memory plane P, the first input / output nodes SL of the memory cells belonging to the same column Ck are interconnected. In addition, the first input / output nodes SL of the memory cells belonging to the columns Ck of odd ranks are interconnected forming a first source line SLia with i from 1 to N. In addition, the first input / output nodes SL of the memory cells belonging to the columns Ck of even ranks are interconnected forming a second source line SL ib with ide 1 to N.
[0106] In each memory plane P;, the second input / output nodes BL of the memory cells belonging to the same column Ck are interconnected. The memory cells belonging to columns of the same rank across all the memory planes Pi form a plane parallel to the geometric plane (X,Z). The second input / output nodes BL belonging to the same plane (X,Z), are interconnected via a common interconnection line. For example, the memory cells of the columns of rank 1 of all the memory planes are interconnected via the interconnection line BLU. The memory cells of the columns of rank 2 of all the memory planes are interconnected via the interconnection line BLib.
[0107] A three-dimensional structure of non-volatile memories is thus obtained comprising a plurality of elementary groups according to the invention capable of carrying out logical operations of the OR, AND type and any other logical operation achievable from the elementary OR and AND operations.
[0108] Advantageously, it is possible to dedicate a plane P; to store the input operands. It is possible to dedicate any plane Pr of the structure to store the results of a logical operation implementing the different combinations of the input operands. This offers programming flexibility that the usual memory calculation circuit structures do not have.
[0109] Generally speaking, the calculation circuit 10 according to the invention has the following advantage: All the cells of the same plane (XZ) have the same common interconnection line BLia. Thus, any memory cell CM of this plane (XZ) can be used to store an operand. The result can be stored in any other memory cell of the plane (XZ), with the exception of the columns comprising the input data. This leaves a great deal of freedom in programming and storing the data in the computing circuit 10.
[0110] In addition, when a logic operation is performed in parallel on all planes (XZ), any memory cell can be opened or closed thanks to an independent selection node WL;j. The selection lines WLÿ are independent within the same plane (XZ), all the cells unused during a logic operation will be closed. This makes it possible to limit the leakage currents in the columns of rank j across all the memory planes P;. For example, the transistors T1 are of the Gate-All-Around type if the memory cells are of the OxRAM type. This limits the leakage currents per memory cell CM to 10 11 A. It is then possible to implement 107 memory cells within the same plane (XZ). This results in a considerable increase in the storage and computing capacities immersed in the memory structure compared to state-of-the-art solutions.
[0111] [Fig.5b] illustrates a sectional view of an embodiment of the thin-film structure of a column Ck of four memory cells CM of the computing circuit in the memory 10 according to the invention. The materials used are presented for non-limiting illustration purposes.
[0112] The column Ck comprises a first metal pillar for producing the common source line SLia and a second metal pillar for producing the common intersection line BL[a. The column further comprises a stack of four independent gate-all-around transistors separated by an insulating material IS1. Each transistor corresponds to the selection transistor T1 of a memory cell. Each transistor comprises a silicon CC conduction channel having a first end corresponding to the drain and a second end corresponding to the source. The use of gate-all-around transistors makes it possible to considerably reduce the leakage currents of the structure. In addition, the gate-all-around transistors have the possibility of being stacked on top of each other in order to produce a dense structure according to the invention.The side wall on the drain side of all the transistors is covered with a succession of layers. The succession of layers comprises in this order a first layer of a dielectric material C'2 such as hafnium oxide (HfO2), then a first metallic layer C'3 of Titanium (Ti), then a second metallic layer of titanium nitride (TiN) for example. The succession of layers obtained at each drain of a transistor T1 forms a resistive storage structure SI corresponding to the stack of [Fig.1a]. The upper electrode ELI consists of the entire first metallic layer of Titanium and the second metallic layer of titanium nitride. The central layer C2 is the hafnium oxide layer. The lower electrode EL2 corresponds to the end of the channel of the transistor TL Advantageously, it is possible to deposit an additional metallic layer between the end of the . channel of transistor Tl and the first layer of a dielectric material C'2 to improve electrical contact at the lower electrode EL2.
[0113] A transistor T1 and the resistive storage structure SI connected to it together form a memory cell CM according to the invention. A stack of four memory cells CM is thus produced.
[0114] The column Ckest obtained by producing the following arrangement: the first metal pillar SLia in contact with all the second ends corresponding to the source; and the second metal pillar BLU placed in contact with all the upper electrodes ELI.
[0115] The structure presented in [Fig.5b] makes it possible to physically implement the circuit according to the invention with an improved geometric density compared to known solutions. More particularly, the computing circuit in the memory according to the invention makes it possible to implement a plurality of 1T1R type memory cells (one transistor and one resistor) with a more compact three-dimensional structure compared to known solutions.
[0116] [Fig.5c] illustrates a second embodiment of the calculation circuit in the memory according to the invention. In this embodiment, the calculation circuit in the memory 10 further comprises, for each interconnection line (BLia>BLib BL2aBL 2b), a control transistor Treg. Each regulation transistor Treg has a gate receiving a regulation signal VWLbl. Each regulation transistor Treg connects the associated interconnection line BL[a to ground so as to produce a variable impedance. The impedance control is carried out via the regulation signal VWLbl generated by the control means of the calculation circuit 10. This makes it possible to indirectly control the ratio of the divider bridge obtained by the memory cells forming an elementary group.
[0117] [Fig.6a] illustrates a third embodiment of the calculation circuit in the memory according to the invention. In this embodiment, the calculation circuit in the memory 10 is such that each interconnection line BLia is connected to the next intersection line BLib via a switch il. For example, the switch il is produced by a pass-gate formed by two complementary transistors. The opening and closing of the pass-gate is controlled by complementary signals WLtransi and WLtran2 generated by the control means of the circuit 10. All of the switches il make it possible to carry out a function of transferring the content stored in a memory cell to another memory cell of a different memory plane.
[0118] [Fig.6b] illustrates an elementary group of memory cells according to a fourth embodiment of the invention intended to carry out a data transfer.
[0119] The starting memory cell CM1 belongs to any column of rank k Ck of a first plane Pin. The arrival (destination) memory cell CM2 belongs to a column of rank k+1 or k-1 of a second different plane Pout. The common interconnection line BLU connected to the departure memory cell CMi is connected to the interconnection line BLib connected to the arrival memory cell CM2 via a pass gate.
[0120] [Fig.6c] illustrates the steps for performing a data transfer with the group ele memory cell module according to the fourth embodiment of the invention. The transfer method described allows data transfer while protecting the contents of the memory cells CMparout and CMp^jn that may be impacted by the transfer. The memory cell CMparin belongs to the same plane as the starting memory cell and is connected to the same interconnection line of the arrival memory cell CMpar_out.
[0121] The control means are configured to carry out the transfer of stored data from a starting memory cell CMi to an arrival memory cell CM2.
[0122] The first step i' consists of initializing the arrival cell (CM2) to a high resistive state.
[0123] The next step ii' consists of putting the switch il in the on state so as to electrically connect the interconnection line BLia and the interconnection line BLib.
[0124] The next step üi' consists of simultaneously carrying out the following operations: a. applying a first positive transfer voltage VSLia to the first input / output node SLia of the starting memory cell CMi; b. connecting the first input / output node SL2b of the arrival memory cell CM2 to electrical ground; c. applying a second positive transfer voltage VSLib to the first input / output node SL[b of the memory cell CMpar in; d. applying the second positive transfer voltage to the first input / output node VSL2a of the memory cell CMpar out;
[0125] The amplitude of the first transfer voltage VSLia is high enough to cause the arrival memory cell CM2 to switch to a low resistive state when the departure memory cell is in a low resistive state (x=1). The amplitude of the second positive transfer voltage VSLib is chosen so as to obtain in all cases a voltage lower than the threshold voltages Vset and Vreset at the terminals of CMpar in and CM Par_out. Thus, the content of these two cells is preserved during the transfer. For example, the first transfer voltage VSLia is chosen to be equal to the supply voltage VDD. The second transfer voltage VSLib is equal to half the supply voltage VDD.
[0126] We have thus demonstrated the possibility of carrying out a data transfer from a plan memory to another of the calculation circuit in the memory 10 according to the invention.
[0127] We will describe, in the following, an example of an algorithmic application of the computing circuit in the memory 10 according to the invention. Hyperdimensional computing algorithms present interesting applications of the computing circuit in the memory according to the invention. This computing paradigm is based on the manipulation of large data vectors called hypervectors (having D coordinates with D>1000). It is possible to use this approach for the classification of texts by language: a text is read sequentially letter by letter. Each letter is associated with a different hypervector. All the hypervectors of the letters of the text are then combined together using the elementary operations of multiplication (via the XNOR function), addition and permutation (via the transfer function) to create a single hypervector associated with the text.The text hypervector then makes it possible to recognize the language of the text by comparison with reference hypervectors.
[0128] [Fig.7] illustrates the steps executed by the calculation circuit 11 according to the invention to calculate the hypervector O resulting from the following input hypervectors: A=(A[0], A[l], A[2], A[3], A[4], A[5]) B=(B[0], B[l], B[2], B[3], B[4], B[5]) Ç=(C[0], C[l], C[2], C[3], C[4], C[5]) D=(D[0], D[l], D[2], D[3], D[4], D[5]) The resulting hypervector is defined by 0= ABDCAD = ABD+BDC+DCA+CAD. the product between two hypervectors A and B is defined as follows: AB=(A[Q1B[Q1. A[1]B[1], A[2]B[2], A[3]B[3], A[4]B[4], A[5]B[5])
[0129] As an example, the calculation circuit in the memory 11 is dimensioned as follows: it comprises four memory planes (Pb P2, P3, P4). Each plane comprises four rows and six columns. The first plane Pi is intended to store the input hypervectors. The second plane P2 is intended to store the results of the logical OR operations. The third plane P3 is intended to store the results of the logical XNOR operations. The fourth plane P4 is intended to store the coordinates of the intermediate results ABD, BDC, DCA and CAD. The logical operation A[i] OR B[i] corresponds to the binary summation A[i] + B[i]. The logical operation A[i] XNOR B[i] corresponds to the binary multiplication A[i].B[i].
[0130] Initially, each row of the first memory plane PiStores the coordinates of an input hypervector A, B, C, or D.
[0131] During step 1a, the calculation circuit 11 is configured to calculate in parallel in the first line of the second plane P2 the results (A+B)[i] = A[i] OR B[i]. Then, the calculation circuit 11 is configured to calculate in parallel in the first line of the third plane P3 the results (AB)[i] = A[i] XNOR B[i].
[0132] During the following step 1b, the calculation circuit 11 is configured to calculate in parallel in the first line of the second plane P2 the results (AB+D)[i] = AB[i] OR D[i]. Then, the calculation circuit 11 is configured to calculate in parallel in the first line of the fourth plane P4 the results (ABD)[i] = AB[i] XNOR D[i]. The coordinates of the vector AB being calculated and stored in the memory cells of the third plane P3 during step 1a.
[0133] We thus obtain the coordinates of the vector ABD in the first line of the fourth plane P4.
[0134] Then, the sequence of steps is repeated to calculate the coordinates of the vector BDC in the second line of the fourth plane P4.
[0135] Then, the sequence of steps is repeated to calculate the coordinates of the DCA vector in the third line of the fourth plane P4.
[0136] Then, the sequence of steps is repeated to calculate the coordinates of the CAD vector in the fourth line of the fourth plane P4.
[0137] Finally, on the fourth plane P4, we obtain the coordinates of the trigrams ABD, BDC, DCA and CAD. The last step consists of calculating the sum of each column to generate the resulting hypervector O as illustrated.
[0138] Generally speaking, the various embodiments of the calculation circuit in the memory according to the invention make it possible to execute any type of algorithm involving elementary Boolean logic operations of the OR and AND type. In addition, those skilled in the art can implement more complex Boolean logic operations from the results of said elementary Boolean logic operations previously described.
[0139] [Fig.8] illustrates an elementary group of memory cells according to a third embodiment of the invention intended to carry out the logic operation A OR B. The elementary group according to the third embodiment is distinguished from that of the first embodiment by the following: The first input / output nodes SL of the first, second and third memory cells CMACMB and CM0R are interconnected by a common interconnection line SLU. The common interconnection line corresponds in this case to the source line of the memory cells CMACMB and CM0R
[0140] In the illustrated elementary group, the first memory cell CMA> the second memory cell CMB and the third memory cell CM0R all belong to columns of the same rank in their respective matrices. In the same memory plane, the second input / output nodes BL of the memory cells belonging to the same column Ck are interconnected by a common data line (bit line in English). This embodiment can be used to perform logic operations of the OR, AND type by applying a set of predetermined voltages in the same way as which is described for the first embodiment.
[0141] References: [1 ] Memory devices and applications for in-memory computing, A. Sébastian, Nature nanotechnology 2020. [2] Four-Layer 3D Vertical RRAM Integrated with FinFET as a Versatile Computing Unit for Brain-Inspired Cognitive Information Processing, H.Li, et al., Symposium on VLSI Technology Digest ofTechnical Papers, 2016
Claims
Claims
1. A computing circuit in the memory (10) comprising a plurality of memory planes (P;) of rank i ranging from 1 to N, with N a natural integer greater than 1, each plane forming a two-dimensional matrix of non-volatile, resistive and programmable memory cells, each memory cell having a selection node (WL), a first input / output node (SL) and a second input / output node (BL); said matrix comprising M rows (Lj) of rank j ranging from 1 to M and K columns (Ck) of rank k ranging from 1 to K, with M and K two unnullified natural integers; said computing circuit (10) comprising: control means configured to control the nodes of the memory cells to perform storage and calculation operations in memory carried out by the calculation circuit; at least one elementary group of memory cells comprising: • a first memory cell (CMA) belonging to any one of the memory planes (Pini) and intended to store a first input data (A); • a second memory cell (CMB) belonging to any one of the memory planes (Pinl) and intended to store a second input data (B); • a third memory cell (CMor CManD) belonging to a memory plane (POri) different from that of the first and second memory cells (CMA CMb); the third memory cell being intended to store the result of a first logical operation having as operands the first and second input data; the first, second and third memory cells (CMA CMB ,CM0R ) belonging to columns (Ck) of the same rank k in their respective matrices; the first, second and third memory cells (CMACMB, CMor) being interconnected by a common interconnection line (BLiaSLia); said common interconnection line connecting one to the other the first input / output nodes (SL) of the memory cells of said elementary group or the second input / output nodes (BL) of the memory cells of said elementary group; the selection nodes (WL) of the memory cells belonging to the same row (Lj) are interconnected by a selection line (WLy); the first input / output nodes (SL) of the memory cells belonging to the same column (Ck) are interconnected; the second input / output nodes (BL) of the memory cells belonging to the same column (Ck) are interconnected.
2.
3. A computing circuit in the memory (10) according to claim 1 wherein each memory cell (CMACMB CM0R) comprises: a programmable resistive storage structure having: an upper electrode (ELI) connected to the second input / output node (BL) of said memory cell (CMaCMb cMor); and a lower electrode (EL2); a selection transistor (Tl) having a gate connected to the selection node (WL) of said memory cell and connecting the lower electrode (EL2) to the first input / output node (SL) of said memory cell (CMACMB,CMOr). Calculation circuit in the memory (10) according to claim 2 in which each column (Ck) of memory cells (CM) comprises: A stack of a plurality of selection transistors (Tl) in a first direction (A); each selection transistor, in particular of the wrapping gate type, comprising a conduction channel (CC), perpendicular to the first direction (A) in a semiconductor material and having two ends; the first end of the channel corresponding to the source of the selection transistor and the second corresponding to the drain of the transistor; a first metal pillar (SLia) along the first direction (A) connecting the sources of the different selection transistors; a dielectric layer (C'2) parallel to the first direction (A) and laterally covering the drains of the transistors selection; - at least one metal layer (C'3) in parallel to the first direction deposited on the dielectric layer; - a second metal pillar (BL[a) in the first direction (A) having lateral contact with the at least one metal layer; for each selection transistor, the assembly formed by the drain, the dielectric layer (C'2) and the metal layer (C'3) constitutes a resistive storage structure (SI).
4. In-memory computing circuit (10) according to any one of the preceding claims such that for each memory plane (Pinl, POri): the first input / output nodes (SL) of the memory cells belonging to the columns (Ck) of odd ranks are interconnected forming a first source line (SLia); the first input nodes (SL) of the memory cells belonging to the columns (Ck) of even ranks are interconnected forming a second source line (SLib); the first source line and the second source line being distinct;
5. Memory calculation circuit (10) according to any one of claims 1 to 4 wherein the elementary group of memory cells further comprises a fourth memory cell (CMxnor) belonging to a memory plane different from that of the first, second and third memory cells (CMA CMbCM0R); the fourth memory cell being intended to store the result of a second logical operation having as operands the first and second input data and using the result of the first logical operation stored in the third memory cell (CM0R); the fourth memory cell (CMXWI|J being connected to the first, second and third memory cells (CMACMB ,CM0R) via the common interconnection line (BLiaSLia).
6. In-memory computing circuit (10) according to any one of claims 1 to 5 wherein for each elementary group of memory cells, the interconnection line (BLia) connects to each other the second input / output nodes (BL) of the memory cells of said elementary group.
7. A memory computing circuit (10) according to any one of claims 1 to 6 wherein the first logical operation is a Boolean OR function.
8. A computing circuit in the memory (10) according to claim 7 wherein the control means are configured to perform the following steps in order to perform the OR logic function: i. Initialize the third cell to a high resistive state; ii. select the first and third memory cells (CMA, CMor) by turning on their selection transistors (Tl); iii. apply a first positive read voltage to the first input / output node (SLU) of the first memory cell and simultaneously connect the first input / output node (SL2a) of the third memory cell to electrical ground; iv. select the second and third memory cells (CMB, CMOr) by turning on their selection transistors (Tl); v.applying the first positive read voltage to the first input / output node (SLia) of the first memory cell and simultaneously connecting the first input / output node (SL2a) of the third memory cell to electrical ground; each memory cell having: a first voltage threshold (Vset) for switching from a high resistive state to a low resistive state; and a second voltage threshold (Vreset) for switching from a low resistive state to a high resistive state greater than the first voltage threshold (Vset); said first read voltage being positive with an amplitude greater than or equal to the first switching threshold and strictly less than the second switching threshold.
9. A memory computing circuit (10) according to any one of claims 7 or 8 in combination with claim 6 wherein the second logical operation is an exclusive NOR (XNOR) Boolean function.
10. A computing circuit in the memory (10) according to claim 9 wherein the control means are configured to carry out the steps following in order to realize the exclusive NOR (XNOR) logic function: i. Initialize the fourth cell to a low resistive state; ii. select the first, third and fourth memory cells (CMA, CMOr, CMxnor) by turning on their transistors; iii. carry out simultaneously: a. applying a second positive read voltage (VSLia) to the first input / output node (SLia) of the first memory cell (CMA); b. applying a third negative read voltage (VSL2a) to the first input / output node (SL2a) of the third memory cell (CM0R); c. and applying a fourth positive read voltage (VSL3a) to the first input / output node (SL3a) of the fourth memory cell (CMxnor); iv. select the second, third and fourth memory cell (CMB, CM0RCMxnor) by turning on their selection transistors (Tl); v. carry out simultaneously: a. applying the second positive read voltage (VSLia) to the first input / output node (SLia) of the second memory cell (CMB); b. applying the third negative read voltage (VSL2a) to the first input / output node (SL2a) of the third memory cell (CM0R); c. and applying a fourth positive read voltage (VSL3a) to the first input / output node (SL3a) of the fourth memory cell (CMxnor); the absolute value of the third reading voltage (VSL2a) being substantially equal to the absolute value of the second reading voltage (VSLla); the sum of the absolute value of the fourth reading voltage (VSL3a) and the third reading voltage (VSL2a) being greater than the second threshold (Vreset)-
11. A memory computing circuit (10) according to any one of claims 1 to 6 wherein the first logical operation is a Boolean AND function.
12. A computing circuit in the memory (10) according to any one of the preceding claims comprising for each interconnection line (BLiaBLib BL2aBL2b) a regulation transistor (Treg); each regulation transistor (Treg) having a gate receiving a regulation signal (VWLbl) and connecting the associated interconnection line to ground so as to produce a variable impedance.
13. A computing circuit in the memory (10) according to any one of the preceding claims comprising a plurality of elementary groups of memory cells; each interconnection line (BLia BLib BL2aBL2b) being connected to an adjacent interconnection line (BLia>BLib BL2aBL2b) via a switch (il).
14. A computing circuit in the memory (10) according to claim 13 wherein the control means are configured to carry out the transfer of a stored data item from a starting memory cell (CMi) to an arrival memory cell (CM2) belonging to a plane different from that of the starting memory cell and connected to an interconnection line adjacent to that of the starting memory cell; the transfer is carried out by executing the following steps: i'- initializing the arrival cell (CM2) to a high resistive state; ii'- setting the switch 11 to the on state; iii'- simultaneously carrying out: a. the application of a first positive transfer voltage to the first input / output node (SLia) of the starting memory cell (CMi); b. the connection of the first input / output node (SL2b) of the arrival memory cell to electrical ground; c.applying a second positive transfer voltage to the first input / output node (SLib) of the memory cell (CMpar in) belonging to the same plane of the starting memory cell and connected to the same interconnection line of the arriving memory cell; d. applying the second positive transfer voltage to the first input / output node (SL2a) of the memory cell. (CMpar_out) belonging to the same plane of the arrival memory cell and connected to the same bit line of the departure memory cell; the amplitude of the second transfer voltage being less than or equal to half of the first transfer voltage.
15. A computing circuit in the memory (10) according to any one of claims 1 to 5 wherein, for each elementary group of memory cells, the interconnection line (SLU) connects to each other the first input / output nodes (SL) of the memory cells of said elementary group.
