An in-memory computing circuit for non-volatile random access memory

By designing the in-memory computing circuit of non-volatile random access memory and employing parallel reading and sparse sensing techniques, the problem of low tunnel magnetoresistance ratio of magnetic tunnel junction devices was solved, achieving efficient multiplication and accumulation operations and reducing circuit delay and energy consumption.

CN117877553BActive Publication Date: 2026-11-17BEIHANG UNIV
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202311480237.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-08
Publication Date
2026-11-17
Estimated Expiration
2043-11-08

AI Technical Summary

Technical Problem

Existing in-memory computing architectures based on magnetic random access memory (MRAM) are difficult to implement in-memory simulation due to the low tunnel magnetoresistance ratio of magnetic tunnel junction devices. This results in low energy efficiency and insufficient computation signal margin, which cannot meet the needs of analog computing.

Method used

Design an in-memory computation circuit for non-volatile random access memory (MTJ), including a register accumulation module, a voltage follower read module, and an input sparsity sensing module. By parallel reading and segment-by-segment sequential computation, the read error rate of MTJ is reduced, and weight sparsity sensing and input sparsity sensing are accelerated.

Benefits of technology

It significantly improves access time in multiplication and accumulation operations, effectively reduces circuit latency and energy consumption, and improves computational efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117877553B_ABST
    Figure CN117877553B_ABST
Patent Text Reader

Abstract

This application provides an in-memory computation circuit for non-volatile random access memory (RAM), belonging to the field of integrated circuit technology. The in-memory computation circuit includes a register accumulation module, a voltage follower read module, and an input sparsity sensing module. The input sparsity sensing module detects whether the input data of each memory cell group is 0 and skips 0-valued input data. The voltage follower read module reads the computation weight of each memory cell group in each column of the memory cell array. The register accumulation module determines the multiplication results of memory cell groups with non-zero input data in each column of the memory cell array based on the computation weights and performs accumulation calculations to obtain the in-memory computation result of that column of memory cell array. This application achieves parallel reading of segmented bit lines through the design of the voltage follower read module and accelerates the process through the design of the input sparsity sensing module, thereby significantly improving access time in multiplication and accumulation operations and effectively reducing circuit latency and power consumption.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of integrated circuits, and more specifically, relates to an in-memory computing circuit for non-volatile random access memory. Background Technology

[0002] Currently, almost all in-memory computing architectures based on Magnetic Random Access Memory (MRAM) are digital in-memory computing schemes, where algorithm operations are executed through scheduling Boolean logic. Although digital in-memory computing schemes achieve highly parallel and reliable computation, their energy efficiency is lower than that of analog in-memory computing schemes (i.e., highly parallel analog multiplication and accumulation operations based on Ohm's law of crossbar arrays) due to more complex timing and control. The main reason why MRAM cannot be used for analog in-memory computing schemes is that the tunnel magnetoresistance ratio (TMR) of magnetic tunnel junction (MTJ) devices is low, which makes it difficult to generate a computational signal margin that meets the requirements of analog computing in traditional crossbar arrays.

[0003] In view of this, there is an urgent need to design a new in-memory computing circuit based on non-volatile memory to solve the above-mentioned technical problems. Summary of the Invention

[0004] To address the problems existing in the prior art, this application provides an in-memory computing circuit for non-volatile random access memory, which can significantly improve access time in multiplication and accumulation operations and effectively reduce circuit latency and power consumption.

[0005] According to a first aspect of this application, an in-memory computing circuit for a non-volatile random access memory (RAM) is provided. The RAM comprises an N-column array of memory cells, each array comprising sixteen groups of memory cells, and each group comprising eight bit cells. The in-memory computing circuit includes a register accumulation module, a voltage follower read module, and an input sparsity sensing module, wherein:

[0006] The input sparsity sensing module is used to detect whether the input data of each of the storage unit groups is 0 and skip the input data that is 0. The input data is eight-bit data, and the eight-bit data corresponds one-to-one with the eight bit units.

[0007] The voltage follower read module is used to read the calculated weight of each group of storage cells on each column of the storage cell array, wherein the voltage follower read module corresponds one-to-one with the storage cell group;

[0008] The register accumulation module is used to determine the multiplication result of the storage cell group whose input data is not 0 on each column of the storage cell array according to the calculation weight, and to perform accumulation calculation to obtain the in-memory calculation result of the column of the storage cell array. The register accumulation module corresponds one-to-one with the storage cell array.

[0009] This application provides an in-memory computing circuit for non-volatile random access memory. On the one hand, by designing a voltage follower read module and applying it to each part of the segmented bit line, parallel reading is achieved, and the voltage follower read module can reduce the read error rate of MTJ (magnetic tunnel junction). On the other hand, an ultra-fast all-digital MAC operation similar to analog CIM is proposed for the bit line, which calculates segment by segment sequentially and realizes weight sparsity sensing. Furthermore, an input sparsity sensing module is designed to accelerate input sparsity sensing, thereby significantly improving access time in multiplication and accumulation operations and effectively reducing circuit latency and power consumption. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is one of the schematic diagrams of the in-memory computing circuit according to an embodiment of this application;

[0012] Figure 2 This is a second schematic diagram of the in-memory computing circuit according to an embodiment of this application;

[0013] Figure 3a This is a schematic diagram illustrating the calculation principle of traditional simulated CIM.

[0014] Figure 3b This is a schematic diagram illustrating the calculation principle of AladCIM in this application;

[0015] Figure 4 This is a schematic diagram of the structure of a storage cell group according to an embodiment of this application;

[0016] Figure 5 This is a schematic diagram of the input sparse sensing module according to an embodiment of this application;

[0017] Figure 6 The transient simulation diagrams of each signal in the input sparse sensing module according to the embodiments of this application are shown below.

[0018] Figure 7This is a third schematic diagram of the in-memory computing circuit according to an embodiment of this application;

[0019] Figure 8 This is a schematic diagram of the structure of a voltage follower reading unit according to an embodiment of this application;

[0020] Figure 9 The transient simulation results are for a voltage follower readout unit according to an embodiment of this application.

[0021] Figure 10 This is a schematic diagram showing the activation time of different parasitic capacitances according to embodiments of this application;

[0022] Figure 11 This is a transient simulation diagram of the MAC operation of the AladCIM circuit structure according to an embodiment of this application. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] In the embodiments of this application, the terms "first," "second," "upper," "lower," etc., are used to distinguish different elements by their names, but do not indicate the spatial arrangement or temporal order of these elements, and these elements should not be limited by these terms. The term "and / or" includes any one or more of the terms listed in association and all combinations thereof. The terms "comprising," "including," "having," etc., refer to the presence of the stated features, elements, components, or assemblies, but do not exclude the presence or addition of one or more other features, elements, components, or assemblies.

[0025] The abbreviations and key terms defined in this application are as follows: IMC: In-Memory Computing; CMOS: Complementary Metal Oxide Semiconductor; SRAM: Static Random-Access Memory; DRAM: Dynamic Random Access Memory; MAC: Multiply-and-Accumulate; NVM: Non-volatile Memory; MTJ: Magnetic Tunnel Junction; STT-CIM: Spin-Transfer Torque Computing-In-Memory; DFF: D Flip-Flop; STT-MRAM: Spin-Transfer-Torque-Magnetic Random Access Memory; MAC: Multiply Accumulate; TMR: Tunnel Magnetoresistance Ratio.

[0026] With the exponential growth of data volume, limited data bandwidth has become a significant factor restricting computing performance. Limited bandwidth not only increases transmission latency but also leads to substantial energy consumption due to the frequent movement of data between memory and the processor. This combination of limited bandwidth and high power consumption during data transfer forms the "Von Neumann bottleneck." In recent decades, the storage speed of memory cells has increased by only about 10%, while the computing performance of processing units has increased by a staggering 60%. This speed difference between memory cells and processing units creates a "memory wall problem." When the processor's computing power reaches a certain level, the speed of accessing memory cannot keep up with the processor's data processing speed, preventing the full utilization of additional processor components.

[0027] To address the aforementioned issues, in-memory computing (IMC) platforms were proposed, which are hardware designs that perform logical operations within a memory array. Currently, various IMC architectures utilize mature memory technologies such as SRAM (Static Random-Access Memory) and DRAM (Dynamic Random-Access Memory). SRAM-based IMC architectures have successfully demonstrated high efficiency in performing multiplication and accumulation (MAC) operations and have become dominant in neural network computing. However, the low density of standard 6-transistor SRAM limits its capacity, restricting its application in large-scale data operations. DRAM-based IMC architectures, benefiting from higher bit density, avoid these drawbacks, but at the cost of increased energy consumption due to their constantly repeated refresh mechanism.

[0028] Emerging non-volatile memory (NVM) utilizes the inherent physical characteristics of devices to store data, allowing its bit cells to consist of a single transistor and a single device, resulting in high bit density similar to DRAM. The non-volatility of NVM enables near-zero leakage and static power consumption, significantly improving performance. Among different types of NVM, spintronic memory offers even higher read / write speeds and has already been commercialized in fields such as automotive electronics and wearable devices. Therefore, it is considered the most promising candidate for next-generation memory.

[0029] In view of this, this application provides an in-memory computing circuit for non-volatile random access memory, namely, a quasi-analog all-digital in-memory computing circuit based on non-volatile memory, which can significantly improve access time and save energy in multiplication and accumulation operations.

[0030] Analog-like all-digital in-memory computing circuits based on non-volatile memory, such as Figure 1 As shown, AladCIM (analog-like in-memory computation) uses completely independent row and column decoders to perform memory and in-memory computation functions respectively; two timing control circuits (memory time domain control and in-memory computation time domain control) generate timing signals for memory and in-memory computation functions respectively, which helps to reduce the complexity of timing circuit design; the input register, input detector and input filter form a sparse sensing module; it also includes a register accumulation circuit, a shift full-add tree and read / write driver and sensing amplifier.

[0031] This application focuses on describing the register accumulation module, the voltage follower reading module, and the input sparse sensing module. The remaining components are existing technologies and will not be described in detail here.

[0032] like Figure 2 As shown, the non-volatile random access memory includes an N-column array of memory cells, each array comprising sixteen groups of memory cells, and each group comprising eight bit cells. The in-memory computing circuitry includes: a register accumulation module, a voltage follower read-through (VRCF) module, and an input sparsity sensing module, wherein:

[0033] The input sparsity sensing module is used to detect whether the input data of each of the storage unit groups is 0 and skip the input data that is 0. The input data is eight-bit data, and the eight-bit data corresponds one-to-one with the eight bit units.

[0034] The voltage follower read module is used to read the calculated weight of each group of storage cells on each column of the storage cell array, wherein the voltage follower read module corresponds one-to-one with the storage cell group;

[0035] The register accumulation module is used to determine the multiplication result of the storage cell group whose input data is not 0 on each column of the storage cell array according to the calculation weight, and to perform accumulation calculation to obtain the in-memory calculation result of the column of the storage cell array. The register accumulation module corresponds one-to-one with the storage cell array.

[0036] This application provides an in-memory computing circuit for non-volatile random access memory. On the one hand, by designing a voltage follower read module and applying it to each part of the segmented bit line, parallel reading is achieved, and the voltage follower read module can reduce the read error rate of MTJ (magnetic tunnel junction). On the other hand, an ultra-fast all-digital MAC operation similar to analog CIM is proposed for the bit line, which calculates segment by segment sequentially and realizes weight sparsity sensing. Furthermore, an input sparsity sensing module is designed to accelerate input sparsity sensing, thereby significantly improving access time in multiplication and accumulation operations and effectively reducing circuit latency and power consumption.

[0037] It should be noted that, as Figure 3aAs shown, the traditional analog in-memory computing principle is introduced. In the traditional CIM analog computing architecture, the computation latency of MAC is closely related to the data precision. For example, when the input precision is 8 bits, it usually takes 8 cycles (i.e., T0 to T7) to complete the MAC operation. In each cycle, the MAC operation is first performed in each column of the cross array; then, the MAC of each column is weighted according to the weight, and they are added together to obtain the LMAC (Local MAC) result; finally, 8 LMAC results are obtained in 8 cycles, which are weighted again according to the input, and they are added together to obtain the GMAC (Global MAC).

[0038] Among them, Figure 3a In this context, Bit-cell represents a storage unit group; Weight0[0] represents the calculated weight of the first storage unit group in the first column; Weight0[7] represents the calculated weight of the first storage unit group in the eighth column; Weight1[0] represents the calculated weight of the second storage unit group in the first column; Weight2[0] represents the calculated weight of the third storage unit group in the first column; Weight i-1[0] represents the calculated weight of the (i-1)th storage unit group in the first column; Weight i[0] represents the calculated weight of the ith storage unit group in the first column; Weight1[7] represents the calculated weight of the second storage unit group in the eighth column; Weight2[7] represents the calculated weight of the third storage unit group in the eighth column; Weight i-1[7] represents the calculated weight of the (i-1)th storage unit group in the eighth column; Weight i[7] represents the calculated weight of the ith storage unit group in the eighth column; Input 0[0]~[7] represents the input data of the first storage unit group in the (1 to 8)th column; Input Input 1[0]~[7] represents the input data of the second storage cell group in the (1 to 8)th column; Input 2[0]~[7] represents the input data of the third storage cell group in the (1 to 8)th column; Input i-1[0]~[7] represents the input data of the (i-1)th storage cell group in the (1 to 8)th column; Input i[0]~[7] represents the input data of the ith storage cell group in the (1 to 8)th column; ADC represents analog-to-digital converter; TDC represents time-to-digital converter.

[0039] Among them, Figure 3aIn the first cycle T0, MAC00 represents the product of the weight of the 0th bit of input data Input0 with the weight of the first storage unit group in the first column, plus the product of the weight of the 0th bit of input data Input1 with the weight of the second storage unit group in the first column, and so on, plus the weight of the 0th bit of Input7 with the weight of the eighth storage unit group in the first column; MAC70 represents the product of the weight of the 0th bit of input data Input0 with the weight of the first storage unit group in the seventh column, plus the product of the weight of the 0th bit of input data Input1 with the weight of the second storage unit group in the seventh column, and so on, plus the weight of the 0th bit of Input7 with the weight of the eighth storage unit group in the seventh column. Figure 3a The same applies to MAC10 to MAC60 (not shown); LMAC0 represents the sum of MAC00 to MAC70 according to a certain weight; in the second cycle T1, MAC01 represents the product of the first bit of input data Input0 with the calculation weight of the first storage unit group of the first column, plus the product of the first bit of input data Input1 with the calculation weight of the second storage unit group of the first column, ... plus the calculation weight of the first bit of Input7 with the calculation weight of the eighth storage unit group of the first column; MAC71 represents the product of the first bit of input data Input0 with the calculation weight of the first storage unit group of the seventh column, plus the product of the first bit of input data Input1 with the calculation weight of the second storage unit group of the seventh column, ... plus the calculation weight of the first bit of Input7 with the calculation weight of the eighth storage unit group of the seventh column; LMAC1 represents the sum of MAC01 to MAC71 according to a certain weight; GMAC represents the sum of LMAC0 to LMAC7 according to a certain weight; Figure 3a The other parameters are explained similarly, and will not be repeated here.

[0040] However, the principle of in-memory computation in this application is as follows: Figure 3b As shown, the computation latency of the AladCIM architecture's MAC is related to the cumulative number of iterations. Figure 3b As shown, in the first cycle T0, only one bit cell in a column is activated, which is multiplied by the input data using logic circuitry. One cycle is the computation time for one group of memory cells. Therefore, for a column of memory cells, the computation from the first group to the sixteenth group comprises 16 computation cycles, i.e., T0 to T15.

[0041] In each remaining cycle, the MAC result obtained in the previous cycle is added to the MAC result of that cycle to obtain the LMAC result. Finally, the LMAC results of the last cycle are weighted based on the weights represented by columns, and they are summed to obtain the GMAC result.

[0042] exist Figure 3bIn the table, Bit-cell represents a storage unit group; Weight0[0] represents the calculation weight of the first storage unit group in the first column; Weight0[7] represents the calculation weight of the first storage unit group in the eighth column; Weight1[0] represents the calculation weight of the second storage unit group in the first column; Weight2[0] represents the calculation weight of the third storage unit group in the first column; Weight i-1[0] represents the calculation weight of the (i-1)th storage unit group in the first column; Weight i[0] represents the calculation weight of the ith storage unit group in the first column; Weight1[7] represents the calculation weight of the second storage unit group in the eighth column; Weight2[7] represents the calculation weight of the third storage unit group in the eighth column; Weight i-1[7] represents the calculation weight of the (i-1)th storage unit group in the eighth column; Weight i[7] represents the calculation weight of the ith storage unit group in the eighth column; MAC Logic represents the multiplication-accumulation logic operation unit; Input 0[7:0] represents the input data of the first storage unit group in the (1 to 8)th column; Input 1[7:0] represents the input data of the second storage cell group in the (1 to 8)th column; Input i 0[7:0] represents the input data of the i-th storage cell group in the (1 to 8)th column.

[0043] exist Figure 3b In the first cycle T0, MAC00 is the product of the first group of memory cells in the first column; MAC70 is the product of the first group of memory cells in the eighth column. Figure 3b Similarly, MAC10 to MAC60 (not shown) are also described in this application and will not be repeated here. In the second cycle T1, MAC01 is the product result of the second group of memory cells in the first column; MAC71 is the product result of the second group of memory cells in the eighth column. Figure 3b Similarly, MAC11 to MAC61 (not shown) are also described in this application. LMAC01 is the sum of the product of the first storage cell group in the first column and the product of the second storage cell group in the first column; LMAC71 is the sum of the product of the first storage cell group in the eighth column and the product of the second storage cell group in the eighth column. Figure 3b The same applies to LMAC11 to LMAC61, which are not shown in the figure, and will not be described again in this application.

[0044] In the i-th period Ti, MAC0i is the product of the i-th memory cell group in the first column; MAC7i is the product of the i-th memory cell group in the eighth column; LMAC0i is equivalent to the sum of the products of the first memory cell group in the first column to the i-th memory cell group in the first column; LMAC7i is equivalent to the sum of the products of the first memory cell group in the eighth column to the i-th memory cell group in the eighth column; GMAC is the sum of the products of each memory cell group in the eight-column memory cell array, rearranged into... Figure 3b The formula shown in the figure.

[0045] Furthermore, the data input precision of the storage cell group in this application is 8 bits, thus, as Figure 4 As shown, a memory cell group consists of 8 memory cells (i.e., bit cells). Each memory cell contains a transistor controlled by a word line (WL) signal and an MTJ (magnetic tunnel junction). One bit of data is stored in the MTJ in the form of an MTJ resistance state, where a high resistance state represents the data "1" and a low resistance state represents the digit "0". In a memory cell group, the first memory cell is called the "0-input converter" and its MTJ is permanently set to a low resistance state, while the other 7 memory cells each store 1 bit of weighted data.

[0046] In some optional embodiments of this example, such as Figure 2 As shown, the memory cell array further includes bit lines and source lines, each of which is divided into sixteen equal parts, wherein:

[0047] The memory cell array further includes bit lines and source lines, each of which is divided into sixteen equal parts, wherein:

[0048] The bit lines include the first segment bit line BL0[0], the second segment bit line BL0[1], the third segment bit line, the fourth segment bit line, the fifth segment bit line, the sixth segment bit line, the seventh segment bit line, the eighth segment bit line, the ninth segment bit line, the tenth segment bit line, the eleventh segment bit line, the twelfth segment bit line, the thirteenth segment bit line, the fourteenth segment bit line, the fifteenth segment bit line, and the sixteenth segment bit line BL0

[15] .

[0049] The source lines include the first segment source line SL0[0], the second segment source line SL0[1], the third segment source line, the fourth segment source line, the fifth segment source line, the sixth segment source line, the seventh segment source line, the eighth segment source line, the ninth segment source line, the tenth segment source line, the eleventh segment source line, the twelfth segment source line, the thirteenth segment source line, the fourteenth segment source line, the fifteenth segment source line, and the sixteenth segment source line;

[0050] The sixteen storage unit groups include the first storage unit group, the second storage unit group, the third storage unit group, the fourth storage unit group, the fifth storage unit group, the sixth storage unit group, the seventh storage unit group, the eighth storage unit group, the ninth storage unit group, the tenth storage unit group, the eleventh storage unit group, the twelfth storage unit group, the thirteenth storage unit group, the fourteenth storage unit group, the fifteenth storage unit group, and the sixteenth storage unit group.

[0051] In this configuration, one end of the nth memory cell group is connected to the bit line of the nth segment, and the other end of the nth memory cell group is connected to the source line of the nth segment. n is an integer greater than or equal to 1 and less than or equal to 16.

[0052] For example, one end of the first storage cell group is connected to the first segment bit line, and the other end of the first storage cell group is connected to the first segment source line; one end of the second storage cell group is connected to the second segment bit line, and the other end of the second storage cell group is connected to the second segment source line; one end of the third storage cell group is connected to the third segment bit line, and the other end of the third storage cell group is connected to the third segment source line; one end of the fourth storage cell group is connected to the fourth segment bit line, and the other end of the fourth storage cell group is connected to the fourth segment source line; one end of the fifth storage cell group is connected to the fifth segment bit line, and the other end of the fifth storage cell group is connected to the fifth segment source line; one end of the sixth storage cell group is connected to the sixth segment bit line, and the other end of the sixth storage cell group is connected to the sixth segment source line... one end of the sixteenth storage cell group is connected to the sixteenth segment bit line, and the other end of the sixteenth storage cell group is connected to the sixteenth segment source line. The connections of the remaining storage cell groups to the segment source lines and segment bit lines are not described in detail here.

[0053] To further reduce the access time of the AladCIM circuit structure when performing all-digital MAC operations, this application designs an input sparsity sensing acceleration scheme, such as... Figure 2 As shown, the input sparse sensing module includes an input register unit, an input detection unit, and an input filtering unit.

[0054] In some optional embodiments of this example, such as Figure 2As shown, the input terminals of the input sparse sensing module include: the first storage unit group data input terminal IN0[7:0], the second storage unit group data input terminal IN1[7:0], the third storage unit group data input terminal IN2[7:0], the fourth storage unit group data input terminal IN3[7:0], the fifth storage unit group data input terminal IN4[7:0], the sixth storage unit group data input terminal IN5[7:0], the seventh storage unit group data input terminal IN6[7:0], and the eighth storage unit group data input terminal IN7[7:0]. IN8[7:0], IN9[7:0], IN10[7:0], IN11[7:0], IN12[7:0], IN13[7:0], IN14[7:0], IN15[7:0], and IN15[7:0], the data input terminals of the 16th and 17th memory groups, respectively.

[0055] The output terminals of the input sparse sensing module include a first input data detection signal output terminal DE0, a second input data detection signal output terminal DE1, a third input data detection signal output terminal, a fourth input data detection signal output terminal, a fifth input data detection signal output terminal, a sixth input data detection signal output terminal, a seventh input data detection signal output terminal, an eighth input data detection signal output terminal, a ninth input data detection signal output terminal, a tenth input data detection signal output terminal, an eleventh input data detection signal output terminal, a twelfth input data detection signal output terminal, a thirteenth input data detection signal output terminal, a fourteenth input data detection signal output terminal, a fifteenth input data detection signal output terminal, and a sixteenth input data detection signal output terminal DE15, wherein the nth input data detection signal output terminal is used to output the nth input data detection signal.

[0056] For example, the first input data detection signal output terminal DE0 is used to output the first input data detection signal; the second input data detection signal output terminal DE1 is used to output the second input data detection signal; the third input data detection signal output terminal is used to output the third input data detection signal; the fourth input data detection signal output terminal is used to output the fourth input data detection signal; the fifth input data detection signal output terminal is used to output the fifth input data detection signal; the sixth input data detection signal output terminal is used to output the sixth input data detection signal; the seventh input data detection signal output terminal is used to output the seventh input data detection signal; and the eighth input data detection signal output terminal is used to output the eighth input data detection signal. The ninth input data detection signal output terminal is used to output the ninth input data detection signal; the tenth input data detection signal output terminal is used to output the tenth input data detection signal; the eleventh input data detection signal output terminal is used to output the eleventh input data detection signal; the twelfth input data detection signal output terminal is used to output the twelfth input data detection signal; the thirteenth input data detection signal output terminal is used to output the thirteenth input data detection signal; the fourteenth input data detection signal output terminal is used to output the fourteenth input data detection signal; the fifteenth input data detection signal output terminal is used to output the fifteenth input data detection signal; and the sixteenth input data detection signal output terminal is used to output the sixteenth input data detection signal.

[0057] In some optional embodiments of this example, such as Figure 2 As shown, the input sparse sensing module includes an input register unit, an input detection unit, and an input filtering unit.

[0058] First, the input register unit of the input sparse sensing module will be introduced:

[0059] The input terminals of the input register unit are the data input terminals IN0[7:0] of the first storage unit group, IN1[7:0] of the second storage unit group, IN2[7:0] of the third storage unit group, IN3[7:0] of the fourth storage unit group, IN4[7:0] of the fifth storage unit group, IN5[7:0] of the sixth storage unit group, IN6[7:0] of the seventh storage unit group, IN7[7:0] of the eighth storage unit group, IN8[7:0] of the ninth storage unit group, and the tenth storage unit. The nth storage unit group has the following data input terminals: IN9[7:0], IN10[7:0], IN11[7:0], IN12[7:0], IN13[7:0], IN14[7:0], IN15[7:0], and IN15[7:0]; wherein, the nth storage unit group data input terminal is used to input the input data of the nth storage unit group; the output terminal of the input register unit is connected to the input detection unit.

[0060] Secondly, the input detection unit of the input sparsity sensing module will be introduced:

[0061] The input detection unit includes: a first input detection subunit, a second input detection subunit, a third input detection subunit, a fourth input detection subunit, a fifth input detection subunit, a sixth input detection subunit, a seventh input detection subunit, an eighth input detection subunit, a ninth input detection subunit, a tenth input detection subunit, an eleventh input detection subunit, a twelfth input detection subunit, a thirteenth input detection subunit, a fourteenth input detection subunit, a fifteenth input detection subunit, and a sixteenth input detection subunit;

[0062] The first input terminal of the nth input detection subunit is used to input the input data of the nth storage unit group. The input data of the nth storage unit group includes: the nth first bit data, the nth second bit data, the nth third bit data, the nth fourth bit data, the nth fifth bit data, the nth sixth bit data, the nth seventh bit data, and the nth eighth bit data.

[0063] The output terminal of the nth input detection subunit is the output terminal of the nth input data detection signal.

[0064] In some alternative embodiments of this example, the second input terminal of the nth input detection subunit is an ENCS signal terminal, which is used to input the ENCS signal.

[0065] In a specific example, such as Figure 2 As shown, when n=1:

[0066] The first input terminal of the first input detection subunit is used to input the input data of the first storage unit group, and the second input terminal of the first input detection subunit is the ENCS signal terminal; the input data of the first storage unit group includes: first first bit data IN0[0], first second bit data IN0[1], first third bit data IN0[2], first fourth bit data, first fifth bit data, first sixth bit data, first seventh bit data and first eighth bit data IN0[7]; the output terminal of the first input detection subunit is the first input data detection signal output terminal DE0.

[0067] In yet another specific example, when n = 2:

[0068] The first input terminal of the second input detection subunit is used to input the input data of the second storage unit group, and the second input terminal of the second input detection subunit is the ENCS signal terminal; the input data of the second storage unit group includes: second first bit data IN1[0], second second bit data IN1[1], second third bit data IN1[2], second fourth bit data, second fifth bit data, second sixth bit data, second seventh bit data and second eighth bit data IN1[7]; the output terminal of the second input detection subunit is the second input data detection signal output terminal DE1.

[0069] Similarly, other cases of n are similar, and will not be elaborated upon here.

[0070] It should be noted that in this application, the circuit structures of the first input detection subunit to the sixteenth detection subunit are identical, that is, they include the same structure and the same number of components, such as... Figure 2 The structure of the first input detection subunit is described as follows:

[0071] The first input detection subunit includes: a first three-input OR gate, a second three-input OR gate, a first two-input OR gate, a third three-input OR gate, a first two-input AND gate, a second two-input AND gate, a first NOT gate, a third two-input AND gate, a first NMOS transistor, a second NMOS transistor, a first PMOS transistor, a third NMOS transistor, and a first delay subunit.

[0072] Wherein, the first input terminal of the first three-input OR gate is used to input the first first bit data IN0[0]; the second input terminal of the first three-input OR gate is used to input the first second bit data IN0[1]; the third input terminal of the first three-input OR gate is used to input the first third bit data IN0[2]; the output terminal of the first three-input OR gate is used to output the first OR result of the first first bit data IN0[0], the first second bit data IN0[1], and the first third bit data IN0[2];

[0073] The first input terminal of the second three-input OR gate is used to input the first fourth bit data; the second input terminal of the second three-input OR gate is used to input the first fifth bit data; the third input terminal of the second three-input OR gate is used to input the first sixth bit data; the output terminal of the second three-input OR gate is used to output the second OR result of the first fourth bit data, the first fifth bit data, and the first sixth bit data.

[0074] The first input terminal of the first two-input OR gate is used to input the first seventh bit data; the second input terminal of the first two-input OR gate is used to input the first eighth bit data IN0[7]; the output terminal of the first two-input OR gate is used to output the third OR result of the first seventh bit data and the first eighth bit data IN0[7];

[0075] The first input terminal of the third three-input OR gate is connected to the output terminal of the first three-input OR gate; the second input terminal of the third three-input OR gate is connected to the output terminal of the second three-input OR gate; the third input terminal of the third three-input OR gate is connected to the output terminal of the first two-input OR gate; the output terminal of the third three-input OR gate is used to output the fourth OR result of the first OR result, the second OR result, and the third OR result.

[0076] Wherein, the first input terminal of the first two-input AND gate is connected to the output terminal of the third three-input OR gate, the second input terminal of the first two-input AND gate is connected to the ENCS signal terminal, and the output terminal of the first two-input AND gate is used to output the fourth OR result and the first AND result of the ENCS signal;

[0077] The first input terminal of the second two-input AND gate is connected to the first ENID signal terminal, wherein the first ENID signal terminal is used to input the first ENID signal; the second input terminal of the second two-input AND gate is connected to the output terminal of the first two-input AND gate, and the output terminal of the second two-input AND gate is used to output the first ENID signal and the second AND result of the first AND result;

[0078] The first input terminal of the third two-input AND gate is connected to the output terminal of the first NOT gate, wherein the input terminal of the first NOT gate is connected to the output terminal of the first two-input AND gate, and the output terminal of the first NOT gate is used to output the first NOT result of the first AND result; the second input terminal of the third two-input AND gate is connected to the first ENID signal terminal, and the output terminal of the third two-input AND gate is used to output the third AND result of the first NOT result and the first ENID signal.

[0079] Wherein, the gate of the first NMOS transistor is connected to the output terminal of the first NOT gate, the first terminal of the first NMOS transistor is connected to the output terminal of the third two-input AND gate, and the second terminal of the first NMOS transistor is connected to the first ENID signal terminal;

[0080] The gate of the second NMOS transistor is connected to the gate of the third NMOS transistor, the first end of the second NMOS transistor is connected to the output end of the third AND gate, and the second end of the second NMOS transistor is connected to the first ENID signal terminal.

[0081] The gate of the first PMOS transistor is connected to the output terminal of the first NOT gate, the first terminal of the first PMOS transistor is connected to the first terminal of the first delay sub-unit, and the second terminal of the first PMOS transistor is connected to the first ENID signal terminal.

[0082] The gate of the third NMOS transistor is connected to the gate of the second NMOS transistor, the first end of the third NMOS transistor is connected to the first end of the first delay sub-unit, and the second end of the third NMOS transistor is connected to the first ENID signal terminal.

[0083] The second terminal of the first delay subunit is connected to the output terminal of the second two-input AND gate, wherein the output terminal of the second two-input AND gate is the first input data detection signal output terminal DE0, and the first input data detection signal is the second AND result.

[0084] Finally, the input filtering unit of the input sparse sensing module is introduced, such as... Figure 2 As shown:

[0085] The input filtering unit includes a fourth NMOS transistor, a fifth NMOS transistor, a sixth NMOS transistor, a seventh NMOS transistor, an eighth NMOS transistor, a ninth NMOS transistor, a tenth NMOS transistor, an eleventh NMOS transistor, a twelfth NMOS transistor, a thirteenth NMOS transistor, a fourteenth NMOS transistor, a fifteenth NMOS transistor, a sixteenth NMOS transistor, a seventeenth NMOS transistor, an eighteenth NMOS transistor, and a nineteenth NMOS transistor, a fourth three-input OR gate, a fifth three-input OR gate, a sixth three-input OR gate, a seventh three-input OR gate, a second two-input OR gate, a third two-input OR gate, an eighth three-input OR gate, a ninth three-input OR gate, and a fourth two-input OR gate.

[0086] Wherein, the gate of the m-th NMOS transistor is connected to the output terminal of the n-th input data detection signal, the first terminal of the m-th NMOS transistor is connected to the first input terminal of the n-th input detection subunit, and the second terminal of the m-th NMOS transistor is connected to the register accumulation module, wherein m = n + 3.

[0087] In a specific example, when n=1, the gate of the fourth NMOS transistor is connected to the output terminal of the first input data detection signal, the first terminal of the fourth NMOS transistor is connected to the first input terminal of the first input detection subunit, and the second terminal of the fourth NMOS transistor is connected to the register accumulation module.

[0088] When n=2, the gate of the fifth NMOS transistor is connected to the output terminal of the second input data detection signal, the first terminal of the fifth NMOS transistor is connected to the first input terminal of the second input detection subunit, and the second terminal of the fifth NMOS transistor is connected to the register accumulation module.

[0089] When n=16, the gate of the nineteenth NMOS transistor is connected to the output terminal of the sixteenth input data detection signal, the first terminal of the nineteenth NMOS transistor is connected to the first input terminal of the sixteenth input detection subunit, and the second terminal of the nineteenth NMOS transistor is connected to the register accumulation module.

[0090] Wherein, the first input terminal of the fourth three-input OR gate is connected to the first input terminal of the first input detection subunit, the second input terminal of the fourth three-input OR gate is connected to the first input terminal of the second input detection subunit, the third input terminal of the fourth three-input OR gate is connected to the first input terminal of the third input detection subunit, and the output terminal of the fourth three-input OR gate is used to output the first input data detection signal, the second input data detection signal, and the fifth OR result of the second input data detection signal;

[0091] The first input terminal of the fifth three-input OR gate is connected to the first input terminal of the fourth input detection subunit, the second input terminal of the fifth three-input OR gate is connected to the first input terminal of the fifth input detection subunit, the third input terminal of the fifth three-input OR gate is connected to the first input terminal of the sixth input detection subunit, and the output terminal of the fifth three-input OR gate is used to output the sixth OR result of the fourth input data detection signal, the fifth input data detection signal, and the sixth input data detection signal;

[0092] The first input terminal of the sixth three-input OR gate is connected to the first input terminal of the seventh input detection subunit, the second input terminal of the sixth three-input OR gate is connected to the first input terminal of the eighth input detection subunit, the third input terminal of the sixth three-input OR gate is connected to the first input terminal of the ninth input detection subunit, and the output terminal of the sixth three-input OR gate is used to output the seventh OR result of the seventh input data detection signal, the eighth input data detection signal, and the ninth input data detection signal;

[0093] The first input terminal of the 73-input OR gate is connected to the first input terminal of the 10th input detection subunit, the second input terminal of the 73-input OR gate is connected to the first input terminal of the 11th input detection subunit, the third input terminal of the 73-input OR gate is connected to the first input terminal of the 12th input detection subunit, and the output terminal of the 73-input OR gate is used to output the eighth OR result of the 10th input data detection signal, the 11th input data detection signal, and the 12th input data detection signal;

[0094] The first input terminal of the second two-input OR gate is connected to the first input terminal of the thirteenth input detection subunit, the second input terminal of the second two-input OR gate is connected to the first input terminal of the fourteenth input detection subunit, and the output terminal of the second two-input OR gate is used to output the ninth OR result of the thirteenth input data detection signal and the fourteenth input data detection signal;

[0095] The first input terminal of the third two-input OR gate is connected to the first input terminal of the fifteenth input detection subunit, the second input terminal of the third two-input OR gate is connected to the first input terminal of the sixteenth input detection subunit, and the output terminal of the third two-input OR gate is used to output the tenth OR result of the fifteenth input data detection signal and the sixteenth input data detection signal.

[0096] Wherein, the first input terminal of the eighth three-input OR gate is connected to the output terminal of the fourth three-input OR gate, the second input terminal of the eighth three-input OR gate is connected to the output terminal of the fifth three-input OR gate, the third input terminal of the eighth three-input OR gate is connected to the output terminal of the sixth three-input OR gate, and the output terminal of the eighth three-input OR gate is used to output the fifth OR result, the sixth OR result, and the eleventh OR result of the eighth OR result;

[0097] The first input terminal of the 93rd input OR gate is connected to the output terminal of the 73rd input OR gate, the second input terminal of the 93rd input OR gate is connected to the output terminal of the 22nd input OR gate, the third input terminal of the 93rd input OR gate is connected to the output terminal of the 33rd input OR gate, and the output terminal of the 93rd input OR gate is used to output the 8th OR result, the 9th OR result, and the 12th OR result of the 11th OR result;

[0098] The first input terminal of the fourth two-input OR gate is connected to the output terminal of the eighth three-input OR gate, the second input terminal of the fourth two-input OR gate is connected to the output terminal of the ninth three-input OR gate, and the output terminal of the fourth two-input OR gate is used to output the eleventh OR result and the thirteenth OR result of the twelfth OR result. The output terminal of the fourth two-input OR gate is connected to the register accumulation module and uses the thirteenth OR result as the DESUM signal.

[0099] The above describes the input sparse sensing module of this application. Its working principle will be further explained below:

[0100] The input register circuit consists of DFFs and is used to store 16 8-bit input data. The input detection circuit generates DEi and ENIDI signals based on the input data. The 8 input data perform OR logic. If all 8 input data are "0", the result of the OR logic is "0", otherwise the result is "1". The result is ANDed with EDID. When the result is "1", DEi outputs the result of ENID. When the result is "0", DEi outputs "0". For example, when the input data IN0[7:0] of the first storage cell group is not all "0", DE0 is "1", and the MAC operation of IN0[7:0] and Weigh0[0] is performed. At the same time, ENID1 outputs the delay signal delay of ENID. For example, when all the input data IN1[7:0] of the second memory cell group is 0, then DE1 is "0". VFRC cannot transfer the data to CL, so MAC operation cannot be performed. At the same time, ENID2 will directly output the data of ENID1, thus skipping the MAC operation of IN1[7:0] and Weigh1[0]. Figure 5As shown, input data consisting entirely of zeros is skipped directly, requiring no time or energy and reducing computation time. From Figure 5 As can be seen, the input data IN1, IN2, IN7, IN8, and IN14 are all 0, so they can be skipped during calculation. Figure 5 The portion circled in the dashed box is not calculated, thus not consuming additional computing resources. This significantly improves access time in multiplication and accumulation operations, effectively reducing circuit latency and energy consumption.

[0101] The input filtering unit transmits non-zero input data to the register accumulator circuit. The corresponding input data is connected through NMOS transistors (i.e., the fourth to the nineteenth NMOS transistors). For example, the IN0[0] data is connected to the 0 bit of all input data through the NMOS transistor controlled by DE0, avoiding the input of 0 input data, thus making the calculation accurate. At the same time, OR logic is performed on all DEi (i.e., DE0 to DE15) to generate the DESUM signal.

[0102] Reference Figure 6 It shows a transient simulation diagram of the input sparse sensing module, from Figure 6 As can be seen, the execution time of the DESUM signal is about 2.5ns. Compared with the traditional time of 3.2ns, the technical solution of this application can skip the input data with an input of 0. Therefore, compared with the traditional method of calculating the product of each input data, the execution time is greatly reduced.

[0103] To further improve read performance, this application also includes a voltage follower read module. Next, as... Figure 7 The voltage follower reading module (VFRC) of this application is described below, specifically:

[0104] The voltage follower reading module includes: a first voltage follower reading unit VFRC0[0], a second voltage follower reading unit VFRC0[1], a third voltage follower reading unit, a fourth voltage follower reading unit, a fifth voltage follower reading unit, a sixth voltage follower reading unit, a seventh voltage follower reading unit VFRC0[1], an eighth voltage follower reading unit, a ninth voltage follower reading unit, a tenth voltage follower reading unit, an eleventh voltage follower reading unit, a twelfth voltage follower reading unit, a thirteenth voltage follower reading unit, a fourteenth voltage follower reading unit, a fifteenth voltage follower reading unit, and a sixteenth voltage follower reading unit VFRC0

[15] ;

[0105] In this context, one end of the nth voltage follower reading unit is connected to the nth segment bit line, where n is an integer greater than or equal to 1 and less than or equal to 16.

[0106] Furthermore, the input terminal of the nth voltage follower reading unit includes: an nth segment bit line input terminal and an nth reference voltage input terminal, wherein the nth segment bit line input terminal is used to connect the nth segment bit line, and the nth reference voltage input terminal is used to input the nth reference voltage;

[0107] The output of the nth voltage follower read unit is used to output the calculated weight of the nth memory cell group (e.g., Weight n[0]).

[0108] In a specific example, when n = 1:

[0109] The input terminals of the first voltage follower reading unit include: the ENCS signal terminal, the first segment bit line input terminal, and the first reference voltage input terminal, wherein the first segment bit line input terminal is used to connect to the first segment bit line, and the first reference voltage input terminal is used to input the first reference voltage;

[0110] The output of the first voltage follower read unit is used to output the calculated weight of the first storage unit group.

[0111] In yet another specific example, when n = 2:

[0112] The input terminals of the second voltage follower reading unit include: the ENCS signal terminal, the second segment bit line input terminal, and the second reference voltage input terminal, wherein the second segment bit line input terminal is used to connect to the second segment bit line, and the second reference voltage input terminal is used to input the second reference voltage;

[0113] The output of the second voltage follower read unit is used to output the calculated weights of the second storage cell group.

[0114] Similarly, other cases of n are similar, and will not be elaborated upon here.

[0115] It should be noted that the circuit structures of the first voltage follower reading unit VFRC0[0] to the sixteenth voltage follower reading unit VFRC0

[15] are the same, that is, the types and quantities of circuit components are the same.

[0116] like Figure 7 As shown, the circuit structure is introduced using the first voltage follower reading unit as an example. Specifically:

[0117] The first voltage follower readout unit includes a second PMOS transistor, a third PMOS transistor, a fourth PMOS transistor, a fifth PMOS transistor, a twentieth NMOS transistor, and a twenty-first NMOS transistor, wherein:

[0118] The gate of the second PMOS transistor is connected to the gate of the third PMOS transistor. The first end of the second PMOS transistor is used to connect to the power supply, and the second end of the second PMOS transistor is used to connect to the first end of the fourth PMOS transistor.

[0119] The first end of the third PMOS transistor is used to connect to the power supply, the second end of the third PMOS transistor is used to connect to the first end of the fifth PMOS transistor, and the gates of the second PMOS transistor and the third PMOS transistor are connected to the ENCS signal terminal.

[0120] The gate of the fourth PMOS transistor is connected to the gate of the fifth PMOS transistor, and the second end of the fourth PMOS transistor is connected to the first end of the twentieth NMOS transistor.

[0121] The second end of the fifth PMOS transistor is connected to the first end of the twentieth NMOS transistor, and the second end of the fifth PMOS transistor is used as the output end of the first voltage follower readout unit. The gates of the fourth PMOS transistor and the fifth PMOS transistor are connected to the first end of the twentieth NMOS transistor.

[0122] The gate of the twentieth NMOS transistor is the input terminal of the first segmented bit line, and the second terminal of the twentieth NMOS transistor is grounded;

[0123] The gate of the 21st NMOS transistor is the input terminal of the first reference voltage, and the second terminal of the 21st NMOS transistor is grounded.

[0124] Here, according to Figure 8 The circuit structure of the first voltage-following read unit described above is summarized below. The Voltage-Following Read Cell represents the voltage-following read circuit. The first voltage-following read module consists of two parts: one part is a conventional amplifier circuit composed of two NMOS transistors and two PMOS transistors (i.e., the 20th NMOS transistor, the 21st NMOS transistor, the 4th PMOS transistor, and the 5th PMOS transistor); the other part is a control circuit composed of two PMOS transistors (i.e., the 2nd PMOS transistor and the 3rd PMOS transistor).

[0125] When the voltage follower read circuit begins reading data from the bit cell MTJ, the bit line BL is first precharged to VDD, and EN (equivalent to the aforementioned ENCS signal) is also activated. Because the input voltage is greater than the reference voltage (Vref), the output is "1" at this time. Subsequently, the word line is activated, and the voltage on BL (VDD) increases. BL Discharge begins, and the discharge rate varies depending on the state of the MTJ (see reference). Figure 9(Transient simulation results of VFRC). When MTJ is in a low-impedance state, that is, when the state of MTJ is "0", the voltage of the bit line drops faster and is set to V. L Otherwise, it is set to V. H The reference voltage is set at V H —V L Distinguish between them.

[0126] In other words, during the period when the word line is activated, if the input voltage is less than the reference voltage, the output voltage of the voltage follower circuit, Output, drops to a low voltage state "0" as the bit line voltage decreases; if the input voltage is greater than the reference voltage, the output voltage of the voltage follower circuit, Output, increases to a high voltage state "1" as the bit line voltage increases.

[0127] It should be understood that in this application, the calculation weight of each memory cell group is determined according to the output voltage of the voltage follower read circuit. When the output voltage of the voltage follower read unit corresponding to the memory cell group is in a high voltage state "1", it means that the calculation weight of the memory cell group is 1; when the output voltage of the voltage follower read unit corresponding to the memory cell group is in a low voltage state "0", it means that the calculation weight of the memory cell group is 0. Thus, the calculation weight of each memory cell group on each column of memory cell array can be determined. For example, the calculation weights of the first memory cell group to the sixteenth memory cell group of the first column of memory cell array are determined as: Weight 0[0] to Weight 15[0], etc.

[0128] In addition, C BL The smaller the parasitic capacitance, the lower the V BL The faster the voltage drops, the more likely it is to fall below the reference voltage during amplification if the parasitic capacitance is small. This can lead to reading errors. Therefore, if... Figure 10 As shown, the VFRC circuit of this application can adapt to C circuits with different parasitic capacitances by changing the activation time of WL. BL This makes the reading more accurate. Among them, in Figure 10 As can be seen from this, it lists four types of C with different parasitic capacitances. BL The corresponding activation time is also shown, which will not be detailed here.

[0129] In some optional embodiments of this example, such as Figure 4 As shown, the nth segment bit line and the nth segment source line also include an nth first segment transistor and an nth second segment transistor, wherein:

[0130] The in-memory calculation signal ENCIM is transmitted between the gate of the nth first segment transistor and the gate of the nth second segment transistor; the separation and connection of the nth segment bit line and the nth segment source line are controlled by the nth first segment transistor and the nth second segment transistor.

[0131] In some optional embodiments of this example, the in-memory computing circuit further includes a pre-charge transistor, wherein a first terminal of the pre-charge transistor is used to input a PL signal, a second terminal of the pre-charge transistor is used to connect to a power supply, and a third terminal of the pre-charge transistor is used to connect to the bit line, wherein:

[0132] The bit line is precharged to a high potential by the precharge transistor controlled by the PL signal;

[0133] When the bit line is not precharged, the nth first segment transistor and the nth second segment transistor are all turned on under the control of the memory calculation signal ENCIM, connecting the nth segment bit line and the nth segment source line;

[0134] Once the bit line pre-charge is complete, the PL signal disconnects the power supply and the bit line. At least one of the nth first segment transistor and the nth second segment transistor is turned off under the control of the memory calculation signal ENCIM, thus separating the nth segment bit line from the nth segment source line.

[0135] In some optional embodiments of this example, such as Figure 4 As shown, the in-memory computing circuit further includes an nth control transistor disposed between the nth voltage follower read unit and the nth first segment transistor, wherein:

[0136] The gate of the nth control transistor is connected to the output terminal of the nth input data detection signal, and the gate of the nth control transistor is used to input the nth input data detection signal;

[0137] The first terminal of the nth control transistor is connected to the output terminal of the nth voltage follower readout unit;

[0138] The second terminal of the nth control transistor is connected to the CL calculation line of the memory cell array, wherein the CL calculation line is used to transmit the CL signal.

[0139] In a specific example, when n = 1:

[0140] The in-memory computing circuit further includes a first control transistor disposed between the first voltage follower read unit and the first segmented transistor, wherein:

[0141] The gate of the first control transistor is connected to the output terminal of the first input data detection signal, and the gate of the first control transistor is used to input the first input data detection signal;

[0142] The first terminal of the first control transistor is connected to the output terminal of the first voltage follower read unit;

[0143] The second terminal of the first control transistor is connected to the CL calculation line of the memory cell array, wherein the CL calculation line is used to transmit the CL signal.

[0144] In yet another specific example, when n=2: the in-memory calculation circuit further includes a second control transistor disposed between the second voltage follower read unit and the second first segment transistor, wherein:

[0145] The gate of the second control transistor is connected to the output terminal of the second input data detection signal, and the gate of the second control transistor is used to input the second input data detection signal;

[0146] The first terminal of the second control transistor is connected to the output terminal of the second voltage follower read unit;

[0147] The second terminal of the second control transistor is connected to the CL calculation line of the memory cell array, wherein the CL calculation line is used to transmit the CL signal.

[0148] Similarly, the cases for the other n are similar, and will not be elaborated upon here.

[0149] Next, the register accumulation module of this application will be introduced, such as... Figure 7 As shown, the register accumulator module includes thirteen input terminals and twelve output terminals.

[0150] The thirteen input terminals are: the first accumulated data input terminal, the second accumulated data input terminal, the third accumulated data input terminal, the fourth accumulated data input terminal, the fifth accumulated data input terminal, the sixth accumulated data input terminal, the seventh accumulated data input terminal, the eighth accumulated data input terminal, the CL signal input terminal, the ENCS signal terminal, the R input terminal, the S input terminal, and the DESUM signal input terminal.

[0151] The in-memory calculation result is a 12-bit binary number, and the twelfth output terminal is used to output one bit of the in-memory calculation result.

[0152] In some optional embodiments of this example, the first accumulated data input terminal is used to input the nth first bit data; the second accumulated data input terminal is used to input the nth second bit data; the third accumulated data input terminal is used to input the nth third bit data; the fourth accumulated data input terminal is used to input the nth fourth bit data; the fifth accumulated data input terminal is used to input the nth fifth bit data; the sixth accumulated data input terminal is used to input the nth sixth bit data; the seventh accumulated data input terminal is used to input the nth seventh bit data; and the eighth accumulated data input terminal is used to input the nth eighth bit data.

[0153] The CL signal input terminal is connected to the CL calculation line, and the CL signal is input to the CL signal input terminal; the DESUM signal input terminal is connected to the output terminal of the fourth two-input OR gate, and the DESUM signal input terminal is used to input the DESUM signal, which is actually the OR result of DE0 to DE15.

[0154] In some optional embodiments of this example, the register accumulation module includes eight data selectors, twelve output registers, and twelve full adders, wherein: the second terminals of the eight data selectors are all grounded, and the third terminals of the eight data selectors are all connected to the SC signal terminal;

[0155] When the SC signal is 1, the first terminals of the eight data selectors are respectively used to input the nth first bit data, the nth second bit data, the nth third bit data, the nth fourth bit data, the nth fifth bit data, the nth sixth bit data, the nth seventh bit data and the nth eighth bit data;

[0156] When the SC signal is 0, the first terminals of the eight data selectors are used to input 0; the fourth terminals of the eight data selectors are connected to the full adder.

[0157] In some alternative embodiments of this example, the twelve full adders include a first full adder to a twelfth full adder, and the eight data selectors include a first data selector to an eighth data selector, wherein:

[0158] The fourth terminal of the p-th data selector is connected to the B input terminal of the p-th full adder, where p is an integer greater than or equal to 1 and less than or equal to 8; the B input terminals of the ninth full adder and the twelfth full adder are grounded.

[0159] In a specific example, the fourth terminal of the first data selector is connected to the B input terminal of the first full adder; the fourth terminal of the second data selector is connected to the B input terminal of the second full adder; the fourth terminal of the third data selector is connected to the B input terminal of the third full adder... and the fourth terminal of the eighth data selector is connected to the B input terminal of the eighth full adder.

[0160] The twelve output registers include a first output register to a twelfth output register, wherein the first output register is used to output the most significant bit of the in-memory calculation result, and the twelfth output register is used to output the least significant bit of the in-memory calculation result.

[0161] In some optional embodiments of this example, the register accumulation module further includes a first three-input AND gate and a second NOT gate, wherein: the first input terminal of the first three-input AND gate is the input terminal of the CL signal, the second input terminal of the first three-input AND gate is the input terminal of the DESUM signal, the third input terminal of the first three-input AND gate is the input terminal of the ENCS signal, and the output terminal of the first three-input AND gate is used to output the fourth AND result of the CL signal, the DESUM signal, and the ENCS signal, and the fourth AND result is used as the value of the SC signal;

[0162] The input of the second NOT gate is connected to the output of the first three-input AND gate. The output of the second NOT gate is used to output the second NOT result of the fourth AND result and use the second NOT result as the value of the SCB signal. The SCB signal is connected to the clock input of the first output register to the clock input of the twelfth output register. The R input is connected to the first output register to the twelfth output register, and the S input is connected to the first output register to the twelfth output register.

[0163] in, Figure 11 The transient simulation timing diagram of the bit line BL signal, DE0-DE15 signal, CLI signal, SC signal, SCB signal, word line WL signal, S signal, R signal, precharge PL signal, ENCS signal, and ENCIM signal during MAC operation is shown.

[0164] Therefore, combined Figure 11Similar to the aforementioned bit-line separation technique, this application proposes an in-memory computing circuit structure based on bit-line separation technology. The bit line is divided into 16 parts, each with eight bit units. The bit line and the source line (SL) are connected via NMOS transistors and controlled by the ENCIM signal. Each part of the BL reads the result of the unit performing the calculation via VFRCs, and the output of VFRCs is connected to the computing line (CL) via an NMOS transistor controlled by the DEi signal. Before performing the multiply-and-accumulate (MAC) operation, ENCIM and ENCS are high, making the segmented BL connected. At this time, the BL is pre-charged to VDD via the PLi signal through the PMOS transistor. Then, ENCIM goes low, separating the bit line and entering the computing mode. The segmented bit-line technique also makes the parasitic capacitance on the BL evenly distributed. Afterwards, the VFRCi[0] output of the first segment on the BL is transmitted to CLi via the activated DE0 signal.

[0165] For example, when the output "0" of VFRC[0] is transferred to CLI, an AND operation is performed between CLI, DE0, and ENCS in the register accumulator circuit. If CLI is "0" during DEi activation, the SC signal is "0", which turns off the full adder FA and puts DFF in its original state, meaning the register accumulator circuit does not work. Conversely, if CLI is "1", the output "1" of VFRC[1] is transferred to the CLI signal during DEi activation, and the SC signal is "1", causing the register accumulator circuit to work.

[0166] At this point, the FA chain performs addition on the input data and the data in the DFF. SCB is the signal of SC after passing through an inverter (i.e., the second NOT gate). When SC is a falling edge, the SCB signal generates a rising edge, which is used as the Clk signal to trigger the DFF on the rising edge. When DE0 activation ends, the DFF will record the result generated on the FA chain. Through the above principle, the output data of the segmented BL's VFRC is sequentially transmitted to CLI, where addition and register operations are performed on the input data in the register accumulator circuit, thereby achieving 16 weighted accumulations of the 8-bit input data after activating DE0 to DE15.

[0167] In some alternative embodiments of this example, the precharge transistor is a PMOS, the control transistor and the segment transistor are NMOS, the first terminal of the precharge transistor is the gate, the second terminal of the precharge transistor is either the source or the drain, and the third terminal of the precharge transistor is either the source or the drain.

[0168] It should be noted that, due to space limitations in the accompanying drawings, some repeated logic gate components and MOS transistor components mentioned above have been omitted from the drawings.

[0169] In summary, the in-memory computing circuit for non-volatile random access memory provided in this application achieves parallel reading by designing a voltage-following read module and applying it to each segment of the bit line, and the voltage-following read module can reduce the read error rate of MTJ (magnetic tunnel junction). On the other hand, an ultra-fast all-digital MAC operation similar to analog CIM is proposed for the bit line, which calculates segment by segment sequentially and realizes weight sparsity sensing. Furthermore, an input sparsity sensing module is designed to accelerate input sparsity sensing, thereby significantly improving access time in multiplication and accumulation operations and effectively reducing circuit latency and power consumption.

[0170] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. An in-memory computing circuit for a non-volatile random access memory, characterized in that, The non-volatile random access memory includes an N-column array of memory cells, each array comprising sixteen groups of memory cells, and each group comprising eight bit cells; the in-memory computing circuit includes a register accumulation module, a voltage follower read module, and an input sparsity sensing module, wherein: The input sparsity sensing module is used to detect whether the input data of each of the storage unit groups is 0 and skip the input data that is 0. The input data is eight-bit data, and the eight-bit data corresponds one-to-one with the eight bit units. The voltage follower read module is used to read the calculated weight of each group of storage cells on each column of the storage cell array, wherein the voltage follower read module corresponds one-to-one with the storage cell group; The register accumulation module is used to determine the multiplication result of the storage cell group whose input data is not 0 on each column of the storage cell array according to the calculation weight, and perform accumulation calculation to obtain the in-memory calculation result of the column of the storage cell array. The register accumulation module corresponds one-to-one with the storage cell array. The memory cell array further includes bit lines and source lines, each of which is divided into sixteen equal parts: The position lines include the first segment position line, the second segment position line, the third segment position line, the fourth segment position line, the fifth segment position line, the sixth segment position line, the seventh segment position line, the eighth segment position line, the ninth segment position line, the tenth segment position line, the eleventh segment position line, the twelfth segment position line, the thirteenth segment position line, the fourteenth segment position line, the fifteenth segment position line, and the sixteenth segment position line; The source lines include a first segment source line, a second segment source line, a third segment source line, a fourth segment source line, a fifth segment source line, a sixth segment source line, a seventh segment source line, an eighth segment source line, a ninth segment source line, a tenth segment source line, an eleventh segment source line, a twelfth segment source line, a thirteenth segment source line, a fourteenth segment source line, a fifteenth segment source line, and a sixteenth segment source line; The sixteen storage unit groups include the first storage unit group, the second storage unit group, the third storage unit group, the fourth storage unit group, the fifth storage unit group, the sixth storage unit group, the seventh storage unit group, the eighth storage unit group, the ninth storage unit group, the tenth storage unit group, the eleventh storage unit group, the twelfth storage unit group, the thirteenth storage unit group, the fourteenth storage unit group, the fifteenth storage unit group, and the sixteenth storage unit group. In this configuration, one end of the nth memory cell group is connected to the bit line of the nth segment, and the other end of the nth memory cell group is connected to the source line of the nth segment. n is an integer greater than or equal to 1 and less than or equal to 16.

2. The in-memory computing circuit according to claim 1, characterized in that, The input terminals of the input sparse sensing module include: a first storage unit group data input terminal, a second storage unit group data input terminal, a third storage unit group data input terminal, a fourth storage unit group data input terminal, a fifth storage unit group data input terminal, a sixth storage unit group data input terminal, a seventh storage unit group data input terminal, an eighth storage unit group data input terminal, a ninth storage unit group data input terminal, a tenth storage unit group data input terminal, an eleventh storage unit group data input terminal, a twelfth storage unit group data input terminal, a thirteenth storage unit group data input terminal, a fourteenth storage unit group data input terminal, a fifteenth storage unit group data input terminal, and a sixteenth storage unit group data input terminal; The output terminals of the input sparse sensing module include a first input data detection signal output terminal, a second input data detection signal output terminal, a third input data detection signal output terminal, a fourth input data detection signal output terminal, a fifth input data detection signal output terminal, a sixth input data detection signal output terminal, a seventh input data detection signal output terminal, an eighth input data detection signal output terminal, a ninth input data detection signal output terminal, a tenth input data detection signal output terminal, an eleventh input data detection signal output terminal, a twelfth input data detection signal output terminal, a thirteenth input data detection signal output terminal, a fourteenth input data detection signal output terminal, a fifteenth input data detection signal output terminal, and a sixteenth input data detection signal output terminal, wherein the nth input data detection signal output terminal is used to output the nth input data detection signal; The input sparse sensing module includes an input register unit, an input detection unit, and an input filtering unit, wherein: The input terminals of the input register unit are the data input terminals of the first storage unit group, the second storage unit group, the third storage unit group, the fourth storage unit group, the fifth storage unit group, the sixth storage unit group, the seventh storage unit group, the eighth storage unit group, the ninth storage unit group, the tenth storage unit group, the eleventh storage unit group, the twelfth storage unit group, the thirteenth storage unit group, the fourteenth storage unit group, the fifteenth storage unit group, and the sixteenth storage unit group, wherein the nth storage unit group data input terminal is used to input the input data of the nth storage unit group; The output of the input register unit is connected to the input detection unit.

3. The in-memory computing circuit of claim 2, wherein, The input detection unit includes: The first input detection subunit, the second input detection subunit, the third input detection subunit, the fourth input detection subunit, the fifth input detection subunit, the sixth input detection subunit, the seventh input detection subunit, the eighth input detection subunit, the ninth input detection subunit, the tenth input detection subunit, the eleventh input detection subunit, the twelfth input detection subunit, the thirteenth input detection subunit, the fourteenth input detection subunit, the fifteenth input detection subunit, and the sixteenth input detection subunit; The first input terminal of the nth input detection subunit is used to input the input data of the nth storage unit group. The input data of the nth storage unit group includes: the nth first bit data, the nth second bit data, the nth third bit data, the nth fourth bit data, the nth fifth bit data, the nth sixth bit data, the nth seventh bit data, and the nth eighth bit data. The output terminal of the nth input detection subunit is the output terminal of the nth input data detection signal; The second input terminal of the nth input detection subunit is the ENCS signal terminal, which is used to input the ENCS signal. When n=1: The first input terminal of the first input detection subunit is used to input the input data of the first storage unit group, and the second input terminal of the first input detection subunit is the ENCS signal terminal. The input data of the first storage cell group includes: first first bit data, first second bit data, first third bit data, first fourth bit data, first fifth bit data, first sixth bit data, first seventh bit data, and first eighth bit data; The output terminal of the first input detection subunit is the output terminal of the first input data detection signal; The circuit structures of the first input detection subunit to the sixteenth detection subunit are identical, wherein: The first input detection subunit includes: a first three-input OR gate, a second three-input OR gate, a first two-input OR gate, a third three-input OR gate, a first two-input AND gate, a second two-input AND gate, a first NOT gate, a third two-input AND gate, a first NMOS transistor, a second NMOS transistor, a first PMOS transistor, a third NMOS transistor, and a first delay subunit; Wherein, the first input terminal of the first three-input OR gate is used to input the first first bit data; the second input terminal of the first three-input OR gate is used to input the first second bit data; the third input terminal of the first three-input OR gate is used to input the first third bit data; and the output terminal of the first three-input OR gate is used to output the first OR result of the first first bit data, the first second bit data, and the first third bit data. The first input terminal of the second three-input OR gate is used to input the first fourth bit data; the second input terminal of the second three-input OR gate is used to input the first fifth bit data; the third input terminal of the second three-input OR gate is used to input the first sixth bit data; the output terminal of the second three-input OR gate is used to output the second OR result of the first fourth bit data, the first fifth bit data, and the first sixth bit data. The first input terminal of the first two-input OR gate is used to input the first seventh bit data; the second input terminal of the first two-input OR gate is used to input the first eighth bit data; the output terminal of the first two-input OR gate is used to output the third OR result of the first seventh bit data and the first eighth bit data. The first input terminal of the third three-input OR gate is connected to the output terminal of the first three-input OR gate; the second input terminal of the third three-input OR gate is connected to the output terminal of the second three-input OR gate; the third input terminal of the third three-input OR gate is connected to the output terminal of the first two-input OR gate; the output terminal of the third three-input OR gate is used to output the first OR result, the second OR result, and the fourth OR result of the third OR result; Wherein, the first input terminal of the first two-input AND gate is connected to the output terminal of the third three-input OR gate, the second input terminal of the first two-input AND gate is connected to the ENCS signal terminal, and the output terminal of the first two-input AND gate is used to output the fourth OR result and the first AND result of the ENCS signal; The first input terminal of the second two-input AND gate is connected to the first ENID signal terminal, wherein the first ENID signal terminal is used to input the first ENID signal; the second input terminal of the second two-input AND gate is connected to the output terminal of the first two-input AND gate, and the output terminal of the second two-input AND gate is used to output the first ENID signal and the second AND result of the first AND result; The first input terminal of the third two-input AND gate is connected to the output terminal of the first NOT gate, wherein the input terminal of the first NOT gate is connected to the output terminal of the first two-input AND gate, and the output terminal of the first NOT gate is used to output the first NOT result of the first AND result; the second input terminal of the third two-input AND gate is connected to the first ENID signal terminal, and the output terminal of the third two-input AND gate is used to output the third AND result of the first NOT result and the first ENID signal; Wherein, the gate of the first NMOS transistor is connected to the output terminal of the first NOT gate, the first terminal of the first NMOS transistor is connected to the output terminal of the third two-input AND gate, and the second terminal of the first NMOS transistor is connected to the first ENID signal terminal; The gate of the second NMOS transistor is connected to the gate of the third NMOS transistor, the first terminal of the second NMOS transistor is connected to the output terminal of the third AND gate, and the second terminal of the second NMOS transistor is connected to the first ENID signal terminal. The gate of the first PMOS transistor is connected to the output terminal of the first NOT gate, the first terminal of the first PMOS transistor is connected to the first terminal of the first delay sub-unit, and the second terminal of the first PMOS transistor is connected to the first ENID signal terminal. The gate of the third NMOS transistor is connected to the gate of the second NMOS transistor, the first end of the third NMOS transistor is connected to the first end of the first delay sub-unit, and the second end of the third NMOS transistor is connected to the first ENID signal terminal. The second end of the first delay subunit is connected to the output end of the second two-input AND gate, wherein the output end of the second two-input AND gate is the output end of the first input data detection signal, and the first input data detection signal is the second AND result.

4. The in-memory computing circuit of claim 3, wherein, The input filtering unit includes a fourth NMOS transistor, a fifth NMOS transistor, a sixth NMOS transistor, a seventh NMOS transistor, an eighth NMOS transistor, a ninth NMOS transistor, a tenth NMOS transistor, an eleventh NMOS transistor, a twelfth NMOS transistor, a thirteenth NMOS transistor, a fourteenth NMOS transistor, a fifteenth NMOS transistor, a sixteenth NMOS transistor, a seventeenth NMOS transistor, an eighteenth NMOS transistor, and a nineteenth NMOS transistor, a fourth three-input OR gate, a fifth three-input OR gate, a sixth three-input OR gate, a seventh three-input OR gate, a second two-input OR gate, a third two-input OR gate, an eighth three-input OR gate, a ninth three-input OR gate, and a fourth two-input OR gate; Wherein, the gate of the m-th NMOS transistor is connected to the output terminal of the n-th input data detection signal, the first terminal of the m-th NMOS transistor is connected to the first input terminal of the n-th input detection subunit, and the second terminal of the m-th NMOS transistor is connected to the register accumulation module, wherein m=n+3; When n=1, the gate of the fourth NMOS transistor is connected to the output terminal of the first input data detection signal, the first terminal of the fourth NMOS transistor is connected to the first input terminal of the first input detection subunit, and the second terminal of the fourth NMOS transistor is connected to the register accumulation module. When n=2, the gate of the fifth NMOS transistor is connected to the output terminal of the second input data detection signal, the first terminal of the fifth NMOS transistor is connected to the first input terminal of the second input detection subunit, and the second terminal of the fifth NMOS transistor is connected to the register accumulation module. When n=16, the gate of the nineteenth NMOS transistor is connected to the output terminal of the sixteenth input data detection signal, the first terminal of the nineteenth NMOS transistor is connected to the first input terminal of the sixteenth input detection subunit, and the second terminal of the nineteenth NMOS transistor is connected to the register accumulation module. Wherein, the first input terminal of the fourth three-input OR gate is connected to the first input terminal of the first input detection subunit, the second input terminal of the fourth three-input OR gate is connected to the first input terminal of the second input detection subunit, the third input terminal of the fourth three-input OR gate is connected to the first input terminal of the third input detection subunit, and the output terminal of the fourth three-input OR gate is used to output the first input data detection signal, the second input data detection signal, and the fifth OR result of the second input data detection signal; The first input terminal of the fifth three-input OR gate is connected to the first input terminal of the fourth input detection subunit, the second input terminal of the fifth three-input OR gate is connected to the first input terminal of the fifth input detection subunit, the third input terminal of the fifth three-input OR gate is connected to the first input terminal of the sixth input detection subunit, and the output terminal of the fifth three-input OR gate is used to output the sixth OR result of the fourth input data detection signal, the fifth input data detection signal, and the sixth input data detection signal; The first input terminal of the sixth three-input OR gate is connected to the first input terminal of the seventh input detection subunit, the second input terminal of the sixth three-input OR gate is connected to the first input terminal of the eighth input detection subunit, the third input terminal of the sixth three-input OR gate is connected to the first input terminal of the ninth input detection subunit, and the output terminal of the sixth three-input OR gate is used to output the seventh OR result of the seventh input data detection signal, the eighth input data detection signal, and the ninth input data detection signal; The first input terminal of the 73-input OR gate is connected to the first input terminal of the 10th input detection subunit, the second input terminal of the 73-input OR gate is connected to the first input terminal of the 11th input detection subunit, the third input terminal of the 73-input OR gate is connected to the first input terminal of the 12th input detection subunit, and the output terminal of the 73-input OR gate is used to output the eighth OR result of the 10th input data detection signal, the 11th input data detection signal, and the 12th input data detection signal; The first input terminal of the second two-input OR gate is connected to the first input terminal of the thirteenth input detection subunit, the second input terminal of the second two-input OR gate is connected to the first input terminal of the fourteenth input detection subunit, and the output terminal of the second two-input OR gate is used to output the ninth OR result of the thirteenth input data detection signal and the fourteenth input data detection signal; The first input terminal of the third two-input OR gate is connected to the first input terminal of the fifteenth input detection subunit, the second input terminal of the third two-input OR gate is connected to the first input terminal of the sixteenth input detection subunit, and the output terminal of the third two-input OR gate is used to output the tenth OR result of the fifteenth input data detection signal and the sixteenth input data detection signal; Wherein, the first input terminal of the eighth three-input OR gate is connected to the output terminal of the fourth three-input OR gate, the second input terminal of the eighth three-input OR gate is connected to the output terminal of the fifth three-input OR gate, the third input terminal of the eighth three-input OR gate is connected to the output terminal of the sixth three-input OR gate, and the output terminal of the eighth three-input OR gate is used to output the fifth OR result, the sixth OR result, and the eleventh OR result of the eighth OR result; The first input terminal of the 93rd input OR gate is connected to the output terminal of the 73rd input OR gate, the second input terminal of the 93rd input OR gate is connected to the output terminal of the 22nd input OR gate, the third input terminal of the 93rd input OR gate is connected to the output terminal of the 33rd input OR gate, and the output terminal of the 93rd input OR gate is used to output the 8th OR result, the 9th OR result, and the 12th OR result of the 11th OR result; The first input terminal of the fourth two-input OR gate is connected to the output terminal of the eighth three-input OR gate, the second input terminal of the fourth two-input OR gate is connected to the output terminal of the ninth three-input OR gate, and the output terminal of the fourth two-input OR gate is used to output the eleventh OR result and the thirteenth OR result of the twelfth OR result. The output terminal of the fourth two-input OR gate is connected to the register accumulation module and uses the thirteenth OR result as the DESUM signal.

5. The in-memory computing circuit of claim 4, wherein, The voltage follower reading module includes: a first voltage follower reading unit, a second voltage follower reading unit, a third voltage follower reading unit, a fourth voltage follower reading unit, a fifth voltage follower reading unit, a sixth voltage follower reading unit, a seventh voltage follower reading unit, an eighth voltage follower reading unit, a ninth voltage follower reading unit, a tenth voltage follower reading unit, an eleventh voltage follower reading unit, a twelfth voltage follower reading unit, a thirteenth voltage follower reading unit, a fourteenth voltage follower reading unit, a fifteenth voltage follower reading unit, and a sixteenth voltage follower reading unit; Wherein, one end of the nth voltage follower reading unit is connected to the nth segment bit line, and n is an integer greater than or equal to 1 and less than or equal to 16; The circuit structures of the first voltage follower reading unit to the sixteenth voltage follower reading unit are the same; The input terminals of the nth voltage follower reading unit include: an nth segment bit line input terminal and an nth reference voltage input terminal, wherein the nth segment bit line input terminal is used to connect to the nth segment bit line, and the nth reference voltage input terminal is used to input the nth reference voltage; The output of the nth voltage follower read unit is used to output the calculated weight of the nth memory cell group; Where, when n=1: The input terminals of the first voltage follower reading unit include: the ENCS signal terminal, the first segment bit line input terminal, and the first reference voltage input terminal, wherein the first segment bit line input terminal is used to connect to the first segment bit line, and the first reference voltage input terminal is used to input the first reference voltage; The output of the first voltage follower read unit is used to output the calculated weights of the first storage unit group; When n=1, the first voltage follower readout unit includes a second PMOS transistor, a third PMOS transistor, a fourth PMOS transistor, a fifth PMOS transistor, a twentieth NMOS transistor, and a twenty-first NMOS transistor, wherein: The gate of the second PMOS transistor is connected to the gate of the third PMOS transistor. The first end of the second PMOS transistor is used to connect to the power supply, and the second end of the second PMOS transistor is used to connect to the first end of the fourth PMOS transistor. The first end of the third PMOS transistor is used to connect to the power supply, the second end of the third PMOS transistor is used to connect to the first end of the fifth PMOS transistor, and the gates of the second PMOS transistor and the third PMOS transistor are connected to the ENCS signal terminal. The gate of the fourth PMOS transistor is connected to the gate of the fifth PMOS transistor, and the second end of the fourth PMOS transistor is connected to the first end of the twentieth NMOS transistor. The second end of the fifth PMOS transistor is connected to the first end of the twentieth NMOS transistor, and the second end of the fifth PMOS transistor is used as the output end of the first voltage follower readout unit. The gates of the fourth PMOS transistor and the fifth PMOS transistor are connected to the first end of the twentieth NMOS transistor. The gate of the twentieth NMOS transistor is the input terminal of the first segmented bit line, and the second terminal of the twentieth NMOS transistor is grounded; The gate of the 21st NMOS transistor is the input terminal of the first reference voltage, and the second terminal of the 21st NMOS transistor is grounded.

6. The in-memory computing circuit of claim 5, wherein, The nth segment bit line and the nth segment source line also include an nth first segment transistor and an nth second segment transistor, wherein: The gate of the nth first segment transistor and the gate of the nth second segment transistor transmit the in-memory calculation signal; the separation and connection of the nth segment bit line and the nth segment source line are controlled by the nth first segment transistor and the nth second segment transistor; The in-memory computing circuit further includes a pre-charge transistor. The first terminal of the pre-charge transistor is used to input a PL signal, the second terminal is used to connect to a power supply, and the third terminal is used to connect to the bit line. The bit line is precharged to a high potential by the precharge transistor controlled by the PL signal; When the bit line is not precharged, the nth first segment transistor and the nth second segment transistor are both turned on under the control of the memory calculation signal, connecting the nth segment bit line and the nth segment source line; When the bit line is precharged, the PL signal disconnects the power supply and the bit line. At least one of the nth first segment transistor and the nth second segment transistor is turned off under the control of the memory calculation signal, thus separating the nth segment bit line and the nth segment source line. The in-memory computing circuit further includes an nth control transistor disposed between the nth voltage follower read unit and the nth first segment transistor, wherein: The gate of the nth control transistor is connected to the output terminal of the nth input data detection signal, and the gate of the nth control transistor is used to input the nth input data detection signal; The first terminal of the nth control transistor is connected to the output terminal of the nth voltage follower readout unit; The second terminal of the nth control transistor is connected to the CL calculation line of the memory cell array, wherein the CL calculation line is used to transmit the CL signal.

7. The in-memory computing circuit of claim 6, wherein, When n=1: The in-memory computing circuit further includes a first control transistor disposed between the first voltage follower read unit and the first segmented transistor, wherein: The gate of the first control transistor is connected to the output terminal of the first input data detection signal, and the gate of the first control transistor is used to input the first input data detection signal; The first terminal of the first control transistor is connected to the output terminal of the first voltage follower read unit; The second terminal of the first control transistor is connected to the CL calculation line of the memory cell array, wherein the CL calculation line is used to transmit the CL signal.

8. The in-memory computing circuit according to claim 7, characterized in that, The register accumulator module includes thirteen input terminals and twelve output terminals. The thirteen input terminals are as follows: First accumulated data input terminal, second accumulated data input terminal, third accumulated data input terminal, fourth accumulated data input terminal, fifth accumulated data input terminal, sixth accumulated data input terminal, seventh accumulated data input terminal, eighth accumulated data input terminal, CL signal input terminal, ENCS signal terminal, R input terminal, S input terminal and DESUM signal input terminal; The in-memory calculation result is a 12-bit binary number, and the twelve output terminals are used to output one bit of the in-memory calculation result. Wherein, the first accumulated data input terminal is used to input the nth first bit data; the second accumulated data input terminal is used to input the nth second bit data; the third accumulated data input terminal is used to input the nth third bit data; the fourth accumulated data input terminal is used to input the nth fourth bit data; the fifth accumulated data input terminal is used to input the nth fifth bit data; the sixth accumulated data input terminal is used to input the nth sixth bit data; the seventh accumulated data input terminal is used to input the nth seventh bit data; and the eighth accumulated data input terminal is used to input the nth eighth bit data. The CL signal input terminal is connected to the CL calculation line, and the CL signal is input to the CL signal input terminal; The DESUM signal input terminal is connected to the output terminal of the fourth two-input OR gate, and the DESUM signal input terminal is used to input the DESUM signal; The register accumulator module includes eight data selectors, twelve output registers, and twelve full adders, wherein: The second terminal of each of the eight data selectors is grounded, and the third terminal of each of the eight data selectors is connected to the SC signal terminal. When the SC signal is 1, the first terminals of the eight data selectors are respectively used to input the nth first bit data, the nth second bit data, the nth third bit data, the nth fourth bit data, the nth fifth bit data, the nth sixth bit data, the nth seventh bit data and the nth eighth bit data; When the SC signal is 0, the first terminals of the eight data selectors are used to input 0 respectively; The fourth terminal of each of the eight data selectors is connected to the full adder; The twelve full adders include the first full adder to the twelfth full adder, and the eight data selectors include the first data selector to the eighth data selector, wherein: The fourth terminal of the p-th data selector is connected to the B input terminal of the p-th full adder, where p is an integer greater than or equal to 1 and less than or equal to 8; The B input terminal of the ninth full adder is grounded to the B input terminal of the twelfth full adder; The twelve output registers include a first output register to a twelfth output register, wherein the first output register is used to output the most significant bit of the in-memory calculation result, and the twelfth output register is used to output the least significant bit of the in-memory calculation result.

9. The in-memory computing circuit according to claim 8, characterized in that, The register accumulation module further includes a first three-input AND gate and a second NOT gate, wherein: The first input terminal of the first three-input AND gate is the input terminal of the CL signal, the second input terminal of the first three-input AND gate is the input terminal of the DESUM signal, the third input terminal of the first three-input AND gate is the input terminal of the ENCS signal, and the output terminal of the first three-input AND gate is used to output the fourth AND result of the CL signal, the DESUM signal, and the ENCS signal, and the fourth AND result is used as the value of the SC signal; The input of the second NOT gate is connected to the output of the first three-input AND gate. The output of the second NOT gate is used to output the second NOT result of the fourth AND result and use the second NOT result as the value of the SCB signal. The SCB signal is connected to the clock input of the first output register to the clock input of the twelfth output register. The R input terminal is connected to the first output register to the twelfth output register, respectively, and the S input terminal is connected to the first output register to the twelfth output register, respectively; Wherein, the precharge transistor is a PMOS, the control transistor and the segment transistor are NMOS, the first terminal of the precharge transistor is the gate, the second terminal of the precharge transistor is either the source or the drain, and the third terminal of the precharge transistor is either the source or the drain.

Citation Information

Patent Citations

  • In-memory computing architecture for realizing acceleration and acceleration method thereof

    CN111026700A