RRAM (Resistance Random Access Memory) memory calculation circuit device oriented to edge artificial intelligence reasoning
By employing inverter pre-amplification technology and self-reflecting array to generate reference current in the RRAM in-memory computing circuit, the problem of high-speed MAC operation in existing RRAM in-memory computing circuits is solved, achieving high energy efficiency and low latency computing performance, and supporting edge artificial intelligence inference tasks.
Patent Information
- Application Number
- CN202511047660.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-29
- Publication Date
- 2025-11-18
AI Technical Summary
Existing RRAM in-memory computing circuits struggle to achieve high-speed multiply-accumulate MAC operations with high energy efficiency.
By employing inverter-based pre-amplification technology and using a self-reflecting array to generate a reference current, a 2T2R computing unit and a high-speed readout circuit module are designed, including a sensing resistor, an inverter amplifier circuit, a comparator, and a conversion module, to realize the digital conversion of analog current signals.
Achieve fast MAC computation with high energy efficiency, support edge AI inference tasks, reduce computation latency and improve computation accuracy.
Smart Images

Figure CN120977349A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of RRAM in-memory computing circuit design technology, and more specifically, to an RRAM in-memory computing circuit device for edge artificial intelligence inference. Background Technology
[0002] Resistance-based random-access memory (RRAM) devices are generally made of special metal oxide materials. Compared with traditional memory devices such as static random-access memory (SRAM) and dynamic random-access memory (DRAM), they have many advantages, such as non-volatility, high density, and low power consumption, which can greatly improve the computational density and energy efficiency of in-memory computing circuits. In the design of in-memory computing circuit systems based on novel memory devices, RRAM devices have many advantages over other memory devices. For example, their material structure is compatible with common complementary metal-oxide-semiconductor (CMOS) processes; the low current when operating in a high-impedance state achieves extremely low power consumption, etc. However, RRAM devices provided by current mainstream processes still have many defects that need to be solved by circuit design, such as the set / reset operating voltage being much higher than the power supply voltage of advanced nodes; and the number of stable read / write operations being limited (<100k).
[0003] The variable resistance characteristic of RRAM allows for the construction of various computation primitives, among which the most widely used basic operator is the multiply-and-accumulate (MAC) operation. Based on this, composite operations such as vector-matrix multiplication can be implemented, ultimately leading to the realization of pulsed convolutional neural networks. However, existing RRAM in-memory computing circuits struggle to achieve high-speed multiply-and-accumulate (MAC) operations with high energy efficiency. Summary of the Invention
[0004] In view of this, the present invention provides an RRAM in-memory computing circuit device for edge artificial intelligence inference, which aims to solve the technical problem that existing RRAM in-memory computing circuits are difficult to achieve high-speed multiply-accumulate MAC operations with high energy efficiency.
[0005] One aspect of the present invention provides an RRAM in-memory computing circuit device for edge AI inference, comprising: a power control and input drive module for inputting input data and weight data for AI inference tasks in edge scenarios; an RRAM computing array comprising k×j RRAM computing units, wherein each RRAM computing unit is configured to consist of at least one transistor and at least one RRAM device, the RRAM computing array being used to perform weight writing, weight reading, and multiply-accumulate operations based on the input data and weight data to generate analog current signals; and a high-speed readout circuit module configured as a successive approximation current-sensitive amplifier based on inverter amplification, for converting the j-column analog current signals generated by the RRAM computing array into 3-bit digital signals for output.
[0006] According to an embodiment of the present invention, the RRAM computing unit is configured to consist of two transistors and two RRAM devices; wherein, one transistor and one RRAM device in the RRAM computing unit together form a pair of sub-units, and the sub-units are connected to each other through a shared sensing line.
[0007] According to an embodiment of the present invention, the device further includes: a digital control bus; wherein the power control and input drive module is configured to input input data and weight data for artificial intelligence inference tasks in edge scenarios via the digital control bus; and each RRAM computing unit in the RRAM computing array is configured to be scheduled via the digital control bus.
[0008] According to an embodiment of the present invention, the device further includes: a peripheral control and drive circuit, used to convert the control signal generated by the power control and input drive module into a voltage domain signal required by the RRAM computing array through a level conversion module.
[0009] According to an embodiment of the present invention, a successive approximation current-sensitive amplifier based on inverter amplification includes: a sensing resistor for converting an input analog current signal into a voltage signal; an inverter-based amplification circuit for amplifying the analog current signal using the high-gain region (VM) of the inverter; a comparator for comparing the analog current signal with a reference current and gradually approximating the digital representation of the analog current signal through successive approximation logic; and a conversion module for converting the analog current signal into a 3-bit digital signal output.
[0010] According to an embodiment of the present invention, the sensing resistor is configured as an N-type metal-oxide-semiconductor field-effect transistor.
[0011] According to an embodiment of the present invention, the comparator is configured as a dual current source structure.
[0012] According to an embodiment of the present invention, the conversion module includes: a reference current selection logic module, used to dynamically select a suitable reference current according to the current step of the successive approximation logic; and an output register, used to store binary bits generated by the comparator and output the final 3-bit digital signal.
[0013] According to an embodiment of the present invention, the input data is 2 bits.
[0014] According to an embodiment of the present invention, the weight data is 2 bits.
[0015] Compared with the prior art, the RRAM in-memory computing circuit device for edge artificial intelligence inference provided in the embodiments of the present invention has at least the following beneficial effects:
[0016] The RRAM in-memory computing circuit device for edge AI inference provided in this embodiment of the invention adopts inverter-based pre-amplification technology and uses a self-reflecting array to generate reference current. Therefore, it not only has high energy efficiency, but also can perform fast MAC calculation and achieves extremely low computing latency under the control of a high-frequency clock. Attached Figure Description
[0017] The above and other objects, features and advantages of the present invention will become more apparent from the following description of embodiments of the invention with reference to the accompanying drawings, in which:
[0018] Figure 1 This schematically illustrates a structural diagram of an RRAM in-memory computing circuit device for edge artificial intelligence inference according to an embodiment of the present invention;
[0019] Figure 2 The schematic diagram illustrates the principle of implementing MAC operation using a 2T2R-RRAM computing unit according to an embodiment of the present invention;
[0020] Figure 3 The schematic diagram illustrates the layout of an 18x8-2T2R array according to an embodiment of the present invention;
[0021] Figure 4 A schematic diagram of the peripheral control and drive circuit according to an embodiment of the present invention is shown.
[0022] Figure 5 The schematic diagram illustrates a successive approximation current-sensitive amplifier based on inverter amplification according to an embodiment of the present invention. Detailed Implementation
[0023] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the invention. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the invention for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.
[0024] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.
[0025] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.
[0026] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).
[0027] In the embodiments of this invention, the collection, updating, analysis, processing, use, transmission, provision, disclosure, and storage of data (e.g., including but not limited to user personal information) comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures have been taken to prevent unauthorized access to user personal information data and to maintain the security of user personal information and network security.
[0028] Resistance-based random-access memory (RRAM) devices are generally made of special metal oxide materials. Compared with traditional memory devices such as static random-access memory (SRAM) and dynamic random-access memory (DRAM), they have many advantages, such as non-volatility, high density, and low power consumption, which can greatly improve the computational density and energy efficiency of in-memory computing circuits. In the design of in-memory computing circuit systems based on novel memory devices, RRAM devices have many advantages over other memory devices. For example, their material structure is compatible with common complementary metal-oxide-semiconductor (CMOS) processes; the low current when operating in a high-impedance state achieves extremely low power consumption, etc. However, RRAM devices provided by current mainstream processes still have many defects that need to be solved by circuit design, such as the set / reset operating voltage being much higher than the power supply voltage of advanced nodes; and the number of stable read / write operations being limited (<100k).
[0029] The variable resistance characteristic of RRAM can be used to build various computation primitives, among which the most widely used basic operator is the multiply-and-accumulate (MAC) operation. Based on this, composite operations such as vector-matrix multiplication can be implemented, ultimately leading to the realization of a pulsed convolutional neural network. Current research on RRAM-based MAC computation circuits mainly includes the following two key technologies:
[0030] I. Innovation in Array Architecture. Currently, there are four commonly used array architectures: crossbar, selector-crossbar, 1T1R, and its variant xTyR. In the crossbar architecture, RRAM cells in the same row / column share the input / output voltage, thus causing mutual interference between different RRAM cells. For example, when reading or writing a target cell, other cells in the same row / column may also have a path, leading to problems such as Sneak-Path and IR-Drop. The selector-crossbar architecture is an improvement on the crossbar architecture. Its main idea is to add a bipolar selector device to each RRAM cell, thereby alleviating the leakage and crosstalk problems of the original crossbar array. The 1T1R and xTyR architectures use a transistor to replace the selector device in the 1T1R architecture, achieving more stable computation. The RRAM cells in these architectures can be fabricated in CMOS-BEOL.
[0031] Second, improvements to peripheral circuits. Currently, two common understandings in peripheral circuit research are: (1) If the multi-bit calculation results are read directly, the analog-to-digital converter (ADC) often accounts for most of the calculation power consumption, while the power consumption of RRAM cells and other peripheral circuits can be ignored; (2) RRAM devices fabricated with advanced processes still have the problem of writing resistance accuracy. In addition, the quantization noise introduced by the analog calculation paradigm often results in large errors in multi-bit calculation results (compared to traditional digital calculation).
[0032] In summary, existing RRAM in-memory computing circuits struggle to achieve high-speed multiply-accumulate MAC operations with high energy efficiency.
[0033] To address the above issues, potential research directions include: (1) simplifying the readout circuit as much as possible (ADC-Less) and reducing the power consumption of the ADC, such as ultra-low power readout circuits based on structures like the Sense Amplifier (SA); (2) implementing readout circuits that do not require an ADC (ADC-Free), which often requires a special algorithm framework; and (3) improving the readout circuit to reduce the error rate of the MAC result, which requires the designed ADC to be able to adapt to the write resistance offset of the RRAM device and have low quantization noise.
[0034] Based on this, embodiments of the present invention provide an RRAM in-memory computing circuit device for edge AI inference. It adopts inverter-based pre-amplification technology and uses a self-reflecting array to generate a reference current to solve the technical problem that existing RRAM in-memory computing circuits are difficult to achieve high-speed multiply-accumulate MAC operations with high energy efficiency, thus providing higher-performance hardware devices for edge AI inference applications.
[0035] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.
[0036] Figure 1 The schematic diagram illustrates a structural diagram of an RRAM in-memory computing circuit device for edge artificial intelligence inference according to an embodiment of the present invention.
[0037] like Figure 1 As shown, the RRAM in-memory computing circuit device for edge AI inference in this embodiment may include, for example, a power control and input drive module, an RRAM computing array, and a high-speed readout circuit module.
[0038] The power control and input drive module is used to input input data and weight data for artificial intelligence inference tasks in edge scenarios.
[0039] The RRAM computing array comprises k×j RRAM computing units, each configured to consist of at least one transistor and at least one RRAM device. The RRAM computing array is used to perform weight writing, weight reading, and multiply-accumulate operations based on input data and weight data to generate analog current signals.
[0040] The high-speed readout circuit module is configured as a successive approximation current-sensitive amplifier (SAR-mode Inverter-based Amplifier CSA, SARIA-CSA) to convert the j-column analog current signals generated by the RRAM computing array into 3-bit digital signal outputs.
[0041] In this embodiment, the entire device is built based on a 2T2R 2bIN-2bW-3bOUT in-memory computing architecture, specifically:
[0042] 2T2R refers to the cell structure layout of RRAM computing units.
[0043] According to an embodiment of the present invention, the RRAM computing unit is configured to consist of two transistors (2 Transistors, 2T) and two RRAM devices (2 Resistive RAM, 2R).
[0044] In this RRAM computing unit, a transistor and an RRAM device together form a pair of sub-units. The sub-units are connected by a shared sense line (SL), which means that 2T2R is formed by merging two 1T1R sub-units (each sub-unit contains one transistor and one RRAM device) through a shared sense line.
[0045] In addition, in this embodiment, the RRAM computing array can be configured with 18×8 RRAM computing units.
[0046] 2bIN indicates that the input data is 2 bits, 2bW indicates that the weight data is 2 bits, and 3bOUT indicates that the final output is a 3-bit digital signal.
[0047] Figure 2 The schematic diagram illustrates the principle of implementing MAC operation using a 2T2R-RRAM computing unit according to an embodiment of the present invention.
[0048] like Figure 2 As shown, the 2T2R-RRAM calculation unit of this embodiment consists of two 1T1R-Cells combined into SL. When performing calculations, an equal-difference voltage is first applied between BL0 / SL / BL1, then WL0 and WL1 are pulled high, and the current on SL is waited for to stabilize. Finally, the current value is read as the calculation result. For example, assuming that the resistance values of R0 and R1 have been updated to 33k (LRS) and 100k (HRS) respectively through SET / RESET operations, and then equal-difference voltages of 600mV / 400mV / 200mV are applied to BL0 / SL / BL1 respectively and WL0 and WL1 are pulled high, according to KCL, a 4uA current with an outward direction can be obtained on the SL branch.
[0049] Several 2T2R-Cells are connected in rows and columns to form an array. Each row shares BL control, and each column shares WL and SL control, enabling MAC calculations in a specified format. Based on the above analysis, each 2T2R-Cell has three output currents: {-1, 0, +1}. Figure 3 As shown.
[0050] Figure 3 The schematic diagram illustrates the layout of an 18x8-2T2R array according to an embodiment of the present invention.
[0051] like Figure 3 As shown, in this embodiment, the 2T2R-RRAM computing unit, in its 18x8 array, each column SL has a corresponding output current of {-18,⋯,0,⋯,+18}, totaling 2×18+1=37 types.
[0052] Assuming that the output current of the 2T2R-Cell is in units of ΔuA under given parameter configuration, and if the quantization unit of the output current of the 18x8 array is taken as 2ΔuA, then the MAC output precision of this array can be considered to be floor(log2(37 / 2)) = 4 bits, ultimately outputting a signed 3-bit calculation result. In summary, this 18x8-2T2R array can achieve MAC calculation in 2bIN-2bW-3bOUT mode.
[0053] According to an embodiment of the present invention, the RRAM in-memory computing circuit device for edge artificial intelligence inference may further include a digital control bus.
[0054] The power control and input drive module is configured to input the input data and weight data for artificial intelligence inference tasks in edge scenarios via a digital control bus.
[0055] Each RRAM computing unit in the RRAM computing array is configured to be scheduled via a digital control bus.
[0056] In this embodiment, both the 2-bit input data and the weight data are sent to the architecture for processing via a digital control bus.
[0057] Furthermore, each RRAM computing unit is scheduled via a shared digital control bus, allowing it to independently perform weight write, weight read, and MAC calculation operations. For example, while one computing unit is performing a weight write operation, another computing unit can be controlled to perform a weight read operation, and the remaining computing units can perform MAC calculations, thereby maximizing the parallelism of the entire MAC calculation cluster.
[0058] When performing a weight write operation, the original 2-bit weight data is pre-encoded and sent to the designated computing unit via the digital control bus. Then, it is input to the corresponding level conversion and drive array, and then raised to the operating voltage domain provided by the RRAM power supply. Finally, the voltage values on BL (Bit Line) and SL (Sense Line) are changed, and the resistance value of the target RRAM cell is updated.
[0059] When performing weight readout or MAC calculation operations, the readout signal or the original 2-bit calculation data is also input into the calculation unit in a similar manner. After receiving the handshake prompt, it is necessary to continue sending the reference array enable signal and configure the corresponding 3-bit current-mode sensitive amplifier (Current SA, CSA) through the digital control bus for conversion, and finally output 8 columns of 3-bit readout / calculation results in parallel.
[0060] According to an embodiment of the present invention, the RRAM in-memory computing circuit device for edge artificial intelligence inference may further include: peripheral control and drive circuitry.
[0061] The peripheral control and drive circuit is used to convert the control signals generated by the power control and input drive module into voltage domain signals required by the RRAM computing array through a level conversion module, specifically as follows: Figure 4 As shown.
[0062] Figure 4 A schematic diagram of the peripheral control and drive circuit according to an embodiment of the present invention is shown.
[0063] like Figure 4 As shown, the peripheral control and drive circuit of this embodiment first boosts the digital control signal to the operating voltage domain of the RRAM through a level conversion module, and then connects it to the BL / SL power supply module of the original array for control selection. Among them, BL has 6 sets of operating voltages VBL0→5; SL has 3 sets of operating voltages VSL0→2; WL has only one operating voltage. Therefore, after boosting, WL0 / 1 is directly output to control the two NMOS switches of 2T2R respectively.
[0064] According to embodiments of the present invention, the inverter-based successive approximation current-sensitive amplifier SARIA-CSA may, for example, include: a sensing resistor, an inverter-based amplifier circuit, a comparator, and a conversion module, specifically as follows: Figure 5 As shown.
[0065] Figure 5 The schematic diagram illustrates a successive approximation current-sensitive amplifier based on inverter amplification according to an embodiment of the present invention.
[0066] like Figure 5 As shown, in the inverter-based successive approximation current-sensitive amplifier of this embodiment, the sensing resistor is used to convert the input analog current signal into a voltage signal, and can be implemented with a resistor much smaller than the input impedance.
[0067] In this embodiment, the sensing resistor is replaced by an N-Metal-Oxide-Semiconductor (NMOS) instead of a fixed resistor. This can greatly reduce the layout area and also act as a switch to receive control signals and realize input switching.
[0068] Inverter-based amplifiers utilize the high-gain region (VM) of inverters to amplify analog current signals, and are key to achieving high energy efficiency in readout circuits.
[0069] A comparator is used to compare an analog current signal with a reference current and to approximate the digital representation of the analog current signal step by step through successive approximation logic.
[0070] In this embodiment, the comparator is a dynamic comparator with a double-taillatch structure, which features high sensitivity and low feedback noise.
[0071] The conversion module is used to convert analog current signals into 3-bit digital signals for output. Specifically, it may include a reference current selection logic module and an output register.
[0072] The reference current selection logic module is used to dynamically select a suitable reference current based on the current step of the successive approximation logic.
[0073] The output register is used to store the binary bits generated by the comparator and output the final 3-bit digital signal.
[0074] In this embodiment, the converter module receives the clock and enable signal of the digital control bus, completes the successive approximation logic based on the 3-bit intermediate quantization result temporarily stored in the output register, and generates relevant SA control signals to drive each component to work, and finally outputs DOUT[2:0] to the digital control bus.
[0075] The RRAM in-memory computing circuit device for edge AI inference provided in this invention employs inverter-based pre-amplification technology and uses a self-reflecting array to generate the reference current. Therefore, it not only has high energy efficiency but also enables fast MAC calculation, achieving extremely low computation latency under high-frequency clock control. This circuit device can support AI inference tasks in various edge scenarios. For example, it can be equipped with a 2-bit signed weighted spiking neural network model, independently completing classification and recognition tasks based on the handwritten digit set (modified national institute of standard and technology, MNIST) with high accuracy.
[0076] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features described in the various embodiments of the present invention can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention can be combined and / or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or pairings fall within the scope of this invention.
[0077] The embodiments of the present invention have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of the invention. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the invention, and all such substitutions and modifications should fall within the scope of the invention.
Claims
1. An RRAM in-memory computing circuit device for edge artificial intelligence inference, characterized in that, The device includes: The power control and input drive module is used to input input data and weight data for artificial intelligence inference tasks in edge scenarios; An RRAM computing array includes k×j RRAM computing units, wherein each RRAM computing unit is configured to consist of at least one transistor and at least one RRAM device. The RRAM computing array is used to perform weight writing, weight reading and multiply-accumulate operations based on the input data and the weight data to generate an analog current signal. The high-speed readout circuit module is configured as a successive approximation current-sensitive amplifier based on inverter amplification, used to convert the j-column analog current signals generated by the RRAM computing array into 3-bit digital signal outputs.
2. The apparatus according to claim 1, characterized in that, The RRAM computing unit is configured to consist of two transistors and two RRAM devices; In this RRAM computing unit, a transistor and an RRAM device together form a pair of sub-units, and the sub-units are connected by a shared sensing line.
3. The apparatus according to claim 1, characterized in that, The device further includes: a digital control bus; The power control and input drive module is configured to input input data and weight data for artificial intelligence inference tasks in edge scenarios via the digital control bus. Each RRAM computing unit in the RRAM computing array is configured to be scheduled via the digital control bus.
4. The apparatus according to claim 3, characterized in that, The device further includes: The peripheral control and drive circuit is used to convert the control signals generated by the power control and input drive module into voltage domain signals required by the RRAM computing array through the level conversion module.
5. The apparatus according to claim 1, characterized in that, The inverter-based successive approximation current-sensitive amplifier includes: A sensing resistor is used to convert an input analog current signal into a voltage signal. An inverter-based amplifier circuit is used to amplify the analog current signal by utilizing the high-gain region (VM) of the inverter. A comparator is used to compare the analog current signal with a reference current and to gradually approximate the digital representation of the analog current signal through successive approximation logic; The conversion module is used to convert the analog current signal into a 3-bit digital signal output.
6. The apparatus according to claim 5, characterized in that, The sensing resistor is configured as an N-type metal-oxide-semiconductor field-effect transistor.
7. The apparatus according to claim 5, characterized in that, The comparator is configured as a dual current source structure.
8. The apparatus according to claim 5, characterized in that, The conversion module includes: The reference current selection logic module is used to dynamically select a suitable reference current based on the current step of the successive approximation logic. The output register is used to store the binary bits generated by the comparator and output the final 3-bit digital signal.
9. The apparatus according to any one of claims 1 to 8, characterized in that, The input data is 2-bit data.
10. The apparatus according to any one of claims 1 to 8, characterized in that, The weight data is 2-bit data.
Citation Information
Cited By
Expert model reasoning system and reasoning method based on RRAM memory internal calculation
CN121352033A
