Circuit for achieving low-power-consumption work of analog storage and calculation integrated SRAM

By designing the 11T unit and dynamic aXNOR structure in SRAM, combining the in-cell current isolation mechanism and block sleep strategy, the problem of increasing power consumption in the analog current MAC calculation mode is solved, and low power consumption and high efficiency storage and computing are achieved. SRAM.

CN118939103APending Publication Date: 2024-11-12SHANGHAI INST OF MICROSYSTEM & INFORMATION TECH CHINESE ACAD OF SCI +1
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410891649.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-04
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

In SRAM in analog current MAC calculation mode, as the number of cells and array range increases, the total current on the current accumulation bus is too large, resulting in an increase in overall power consumption.

Method used

A 11T SRAM cell is designed, combining the storage capacity of 6T SRAM and the 5T dynamic aXNOR structure, and adopts the in-cell galvanic isolation mechanism and programmable block sleep strategy to achieve low-power analog memory and computing integrated SRAM. This unit turns off the dynamic analog multiplication accumulation operation in memory mode, and controls the multiplication operation through the characteristic data input port and the operation enable signal port to reduce power consumption.

Benefits of technology

It effectively reduces the power consumption of SRAM integrated memory and computing, improves the energy efficiency ratio, achieves higher chip efficiency, and further reduces static power consumption when not participating in the calculation through the block sleep strategy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118939103A_ABST
    Figure CN118939103A_ABST
Patent Text Reader

Abstract

The invention provides a circuit for realizing low-power-consumption work of an analog storage and calculation integrated SRAM (Static Random Access Memory). The circuit comprises an SRAM storage unit; the dynamic simulation product accumulation operation unit is connected with a port for storing digits of the SRAM storage unit and comprises a characteristic data input port and an operation enable signal input port; the feature data input port is used for receiving binary feature data from external input or pre-stage processing; the operation enable signal input port controls whether the single dynamic bit multiplication operation is executed or not; and the dynamic simulation product accumulation operation unit is configured to be closed in a storage mode, perform multiplication operation on the binary characteristic data and a storage digit obtained from the SRAM storage unit in a calculation mode, and output a result. Through a programmable block dormancy strategy of a storage and calculation integrated 11T SRAM unit and an SRAM array adopting an in-unit current isolation mechanism, the simulation storage and calculation integrated SRAM with low power consumption is realized. The invention designs a storage and calculation integrated SRAM unit which is used for reducing the power consumption when the storage and calculation integrated SRAM performs in-storage calculation work, realizing programmable SRAM block dormancy, achieving a higher energy efficiency ratio and further improving the chip efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of integrated circuit design, and in particular to a circuit for realizing low-power operation of an analog storage-computation integrated SRAM. Background Art

[0002] With the development of integrated storage and computing SRAM technology, the technology of completing neural network algorithms by implementing large-scale parallel current calculations inside the SRAM array has become increasingly mature. Compared with the voltage mode in-memory calculation, the integrated storage and computing SRAM technology that adopts the analog current MAC calculation mode does not require additional adder circuits, which can save a large amount of auxiliary calculation circuit area. At the same time, compared with the in-memory calculation SRAM in charge sharing mode and capacitor coupling mode, it is not necessary to add a capacitor structure in each unit for charge accumulation and coupling, so it has a smaller unit and array area, and a higher area efficiency ratio. However, in the SRAM based on the existing analog current MAC calculation mode, there is still a key problem: when the number of units participating in the calculation at the same time and the array range are large, the total current on the current accumulation bus will be very large, resulting in an increase in overall power consumption. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a low-power analog storage and computing integrated SRAM.

[0004] In order to solve the above problems, the present invention provides a circuit for realizing low-power operation of an analog storage and calculation integrated SRAM, comprising: an SRAM storage unit; a dynamic analog multiplication and accumulation operation unit, which is connected to a port for storing digits of the SRAM storage unit and comprises a feature data input port and an operation enable signal input port; the feature data input port is used to receive binary feature data from external input or previous stage processing; the operation enable signal input port is a dynamic signal interface for controlling whether the single dynamic bit multiplication operation is executed; the dynamic analog multiplication and accumulation operation unit is configured to be closed in storage mode, and to multiply the binary feature data with the storage digit obtained from the SRAM storage unit in calculation mode and output the result.

[0005] Optionally, the SRAM storage cell includes a first inverter composed of a first transistor T1 and a second transistor T2, a second inverter composed of a third transistor T3 and a fourth transistor T4, the two inverters are opposed and interlocked to form charge storage terminals Q and QB for storing digits, and the fifth transistor T5 and the sixth transistor T6 are connected to the bit lines BL and BLB respectively. The dynamic analog multiplication and accumulation operation unit includes a seventh transistor T7 and an eighth transistor T8 connected in series with source / drain, the suspended source / drain is respectively connected to the binary characteristic data X and XB, and the gate is respectively connected to the charge storage terminal Q and QB of the 6T SRAM; the source / drain of the ninth transistor T9 is connected between the seventh transistor T7 and the eighth transistor T8 connected in series, and the gate is connected to the dynamic signal interface; the output current of the ninth transistor T9 enters the current mirror formed by the tenth transistor T10 and the eleventh transistor T11 opposite to each other, and the other end of the current mirror is externally connected to the storage bit line RBL; in the storage mode, the dynamic signal of the above structure outputs a low level, the ninth transistor T9 is turned off, and the dynamic analog multiplication and accumulation operation unit is turned off; in the calculation mode, the dynamic signal outputs a high level, allowing the binary characteristic data to be multiplied with the digit stored in the SRAM storage unit and the result is output to the storage bit line RBL.

[0006] The present invention realizes low-power analog storage-computation integrated SRAM through a storage-computation integrated 11T SRAM unit and a programmable block sleep strategy of an SRAM array using an intra-unit current isolation mechanism. The present invention designs a storage-computation integrated SRAM unit to reduce the power consumption of the storage-computation integrated SRAM when performing in-memory computing work, and realizes programmable SRAM block sleep, achieving a higher energy efficiency ratio, thereby improving chip efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Attached Figure 1 The figure shows a 11TSRAM circuit with in-memory computing capability according to a specific embodiment of the present invention.

[0008] Attached Figure 2 Shown is a signal combination and multiplication result of an 11T SRAM cell according to a specific embodiment of the present invention.

[0009] Attached Figure 3 FIG. 1 is a schematic diagram of a block sleep mode in an SRAM array according to a specific embodiment of the present invention.

[0010] Attached Figure 4 Shown is a schematic diagram of a block calculation mode in an SRAM array given in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0011] The specific implementation of the circuit for realizing low-power operation of analog storage-computation integrated SRAM provided by the present invention is described in detail below in conjunction with the accompanying drawings.

[0012] The present invention aims to propose an SRAM cell that realizes low-power computing of integrated storage and computing SRAM and a programmable block sleep strategy based on the cell. The 11T SRAM cell design is based on dynamic aXNOR computing, combining the storage capacity of 6T SRAM and the computing power of 5T dynamic aXNOR structure, effectively executing single-bit level multiplication tasks, and in order to cope with the high power consumption problem caused by large-scale parallel computing, a current isolation mechanism within the cell is specially introduced to reduce the static power consumption and computing power consumption of the entire array.

[0013] In order to solve the above problems, as shown in the attached Figure 1 As shown, this specific embodiment provides an 11T SRAM circuit with in-memory computing capability, which mainly includes a 6T SRAM and a dynamic analog multiplication and accumulation operation unit, and the dynamic analog multiplication and accumulation operation unit is a 5T dynamic aXNOR structure in this specific embodiment. The 11T SRAM unit is composed of 11T, including a standard 6T SRAM storage unit and a dynamically working 5T aXNOR logic structure.

[0014] The standard 6T SRAM part completes the storage and reading of memory data. The structure follows the working principle of the standard SRAM unit and provides normal and reliable read and write operations in the memory mode. It is mainly used to efficiently and safely store and retrieve important weight parameters W in the binary neural network. In this specific embodiment, the standard 6T SRAM includes a first inverter composed of a first transistor T1 and a second transistor T2, and a second inverter composed of a third transistor T3 and a fourth transistor T4. The two inverters are opposed and interlocked to form charge storage terminals Q and QB for storing digits. The fifth transistor T5 and the sixth transistor T6 are connected to the bit lines BL and BLB respectively.

[0015] The dynamic calculation 5T aXNOR structure adds a calculation function to the 11T SRAM unit proposed in this specific embodiment. This structure includes two independent input ports: a feature data input port Feature Port and an operation enable signal input port COM_EN Port. The feature data input port Feature Port is used to receive binary feature data X and XB from external input or pre-stage processing; the operation enable signal input port COM_EN Port is a dynamic signal (CEN) interface that controls whether the single dynamic bit multiplication operation is executed.

[0016] In this specific embodiment, the dynamic calculation 5T aXNOR structure includes a seventh transistor T7 and an eighth transistor T8 with source / drain connected in series, the suspended source / drain is connected to the binary feature data X and XB respectively, and the gate is connected to the charge storage terminal Q and QB of the 6T SRAM respectively. The source / drain of the ninth transistor T9 is connected between the seventh transistor T7 and the eighth transistor T8 connected in series, and the gate is connected to the dynamic signal (CEN) interface. The output current of the ninth transistor T9 enters the current mirror composed of the tenth transistor T10 and the eleventh transistor T11. The other end of the current mirror is connected to the storage bit line signal RBL.

[0017] Under normal working conditions, the 11T SRAM cell can flexibly realize dual-mode operation thanks to its unique decoupling design architecture of storage word lines and computing word lines: on the one hand, it is a standard SRAM storage mode, and on the other hand, it is a computing mode. In the traditional SRAM storage mode, the 11T SRAM cell maintains the basic storage and reading characteristics of the 6T SRAM cell, and is responsible for storing key weight information W in neural networks or other data-intensive applications, performing efficient data access in binary form, and ensuring that these weight data can respond to read requests from the host computer or associated computing components at any time, thereby ensuring high-speed and reliable weight information access and update capabilities.

[0018] In the computing mode, the storage array composed of 11T SRAM cells realizes the ability of large-scale parallel computing. It realizes the dynamic analog multiplication and accumulation (MAC) operation of the external input signal X and the internal storage weight data W through the integrated aXNOR computing transistor structure. When receiving the real-time changing external input signal X, the 11T SRAM cell will immediately perform parallel computing operations with the weight W stored in it, integrating the original storage function with the computing function, and directly generating the cell current output corresponding to the input-weight pair (X, W) on the computing word line.

[0019] Before conducting an in-depth analysis of the MAC operation current generation, the recoding logic for the input signal of the Feature port is first explained. The proposed analog multiplication design redefines the pair of complementary signals {X, XB} received by the Feature port in a binary sense: when {X, XB} takes the value of {1, 0}, the Feature signal is a positive attribute, that is, the Feature value is +1; conversely, when {X, XB} takes the value of {0, 1}, the Feature value is defined as -1. Similarly, for the Weight storage value stored in the DAM SRAM cell, a similar binary redefinition strategy is also adopted, where the complementary signal pair {Q, QB} takes the value of {1, 0}, the corresponding Weight value is determined to be +1; when {Q, QB} takes the value of {0, 1}, the Weight value is defined as -1.

[0020] In the SRAM storage mode, the dynamic aXNOR enable signal in the 11T SRAM unit is transmitted through the COM_EN port. At this time, the signal (CEN) value is 0, causing the MOS tube T9 controlled by CEN to be in the cut-off state and unable to form an effective current path, thereby ensuring that the DAM SRAM unit will not perform any analog MAC operation in this state.

[0021] When switching to calculation mode, the CEN signal value is set to 1, which turns on the MOS tube T9 to allow the X signal input from the Feature input port to perform analog multiplication with the Weight signal Q stored in the DAM SRAM unit. Figure 2 The signal combinations and multiplication results of the 11T SRAM cell shown in the figure have four different signal combination configurations, each of which corresponds to a unique multiplication result:

[0022] (1) When the Feature signal pair {X, XB} takes the value of {1, 0}, corresponding to +1, and the Weight signal pair {Q, QB} also takes the value of {1, 0}, corresponding to +1, the result of analog multiplication is +1, and a mirror current IC proportional to the analog multiplication current IM is generated on RBL;

[0023] (2) If the Feature signal pair {X, XB} takes the value of {0, 1}, corresponding to -1, and the Weight signal pair {Q, QB} also takes the value of {0, 1}, corresponding to -1, although both are negative values, the product of the two is still +1, and a mirror current IC of the analog multiplication current IM will also be formed on RBL.

[0024] (3) When the value of the Feature signal pair {X, XB} is +1 and the value of the Weight signal pair {Q, QB} is -1, the result of the analog multiplication becomes -1. At this time, theoretically, no analog multiplication result current should be generated, so the corresponding mirror current will not be observed on RBL.

[0025] (4) Finally, if the Feature signal pair {X, XB} takes the value of -1 and the Weight signal pair {Q, QB} takes the value of +1, the result of the analog multiplication is -1 again. Similarly, in this case, no analog current reflecting the multiplication result will be generated on RBL. Figure 2 As shown, the truth table of the 11T SRAM cell exhaustively records all possible combinations of the input signal X and the weight data W, as well as the cell current output value corresponding to each combination.

[0026] The aXNOR part of the computing structure of the 11T SRAM unit adopts an "intra-unit current isolation mechanism" that can significantly reduce the system's computing power consumption. On the one hand, the programmable "intra-unit current isolation mechanism" allows the SRAM array to complete flexible block dormancy and reduce the static power consumption of the SRAM CORE blocks that do not participate in the calculation; on the other hand, the high-speed dynamic intra-unit current isolation mechanism can effectively reduce the computing power consumption of the SRAM block in the computing mode, and realize the low-power operation of the SRAM CORE core.

[0027] like Figure 3 The figure shows a schematic diagram of the block sleep mode in the SRAM array. In the block sleep mode, when the sleep instruction is executed in block 1 in the SRAM array, the COM_EN port of the entire block 1 will set the CEN signal to a low level state, closing the intra-cell calculation path of all SRAM cells in the entire block. The node voltage VK in the cell aXNOR structure is pulled down, so that the left branch in the voltage gradient quantization circuit cannot form a current path, and there is only a small static current IST of the right branch, thereby realizing low-power sleep of the block.

[0028] like Figure 4 The following is a schematic diagram of the block calculation mode in the SRAM array. In the calculation mode, Figure 4As shown, when the SRAM array performs calculations, the COM_EN control port in the area corresponding to the row where the executed data is located will be triggered, and CEN will be set to a high level state. The current path used for the calculation function in the SRAM cell in the executed block is turned on, and the node voltage VK in the aXNOR logic structure of each SRAM cell will be pulled up to a preset higher level, thereby activating the current mirror in the cell connected to it. The current mirror is responsible for regulating the unit current intensity generated during the analog calculation process. At the same time, the left branch of the voltage gradient quantization circuit forms a continuous current path, which can generate a corresponding gradient voltage output based on the state difference of the data in the cell. In the actual calculation mode, the SRAM array can synchronously open 16 independent rows for calculation. Assuming that the output calculation results of all SRAM cells in each row in this calculation cycle are +1, the left branch of the voltage gradient quantization circuit will accumulate the basic calculation current provided by the 16 cells, and finally synthesize a total current output of 160uA. In order to save energy and prevent unnecessary power consumption accumulation, during the calculation cycle, the system will dynamically control the COM_EN port to cancel the high-level state of the CEN signal and restore it to a low level, thereby cutting off the left branch current path that performs the calculation function in the voltage gradient quantization circuit and controlling the calculation time to 1 / 10 of the calculation cycle. Therefore, the dynamic enable signal can reduce the array calculation energy to 10%.

[0029] The above is only a preferred embodiment of the present invention. It should be pointed out that ordinary technicians in this technical field can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A circuit for realizing low power consumption operation of analog storage and computing integrated SRAM, characterized in that: include: SRAM storage unit; A dynamic simulation multiplication and accumulation operation unit is connected to a port for storing digits of the SRAM storage unit and includes a characteristic data input port and an operation enable signal input port; The feature data input port is used to receive binary feature data from external input or previous stage processing; the operation enable signal input port is a dynamic signal interface that controls whether the single dynamic bit multiplication operation is executed; The dynamic simulation multiplication and accumulation unit is configured to be closed in the storage mode, and to multiply the binary feature data with the storage digit obtained from the SRAM storage unit in the calculation mode and output the result.

2. The circuit according to claim 1, characterized in that The SRAM storage cell includes a first inverter composed of a first transistor T1 and a second transistor T2, and a second inverter composed of a third transistor T3 and a fourth transistor T4. The two inverters are opposed and interlocked to form charge storage terminals Q and QB for storing digits. The fifth transistor T5 and the sixth transistor T6 are connected to the bit lines BL and BLB respectively.

3. The circuit according to claim 2, characterized in that The dynamic analog multiplication and accumulation operation unit includes a seventh transistor T7 and an eighth transistor T8 with source / drain connected in series, the suspended source / drain are respectively connected to the binary characteristic data X and XB, and the gate is respectively connected to the charge storage terminal Q and QB of the 6T SRAM; The source / drain of the ninth transistor T9 is connected between the seventh transistor T7 and the eighth transistor T8 connected in series, and the gate is connected to the dynamic signal interface; The output current of the ninth transistor T9 enters a current mirror formed by the tenth transistor T10 and the eleventh transistor T11 in opposition, and the other end of the current mirror is externally connected to the storage bit line RBL; In the storage mode, the dynamic signal of the above structure outputs a low level, the ninth transistor T9 is turned off, and the dynamic analog multiplication and accumulation operation unit is turned off; in the calculation mode, the dynamic signal outputs a high level, allowing the binary feature data to be multiplied by the digits stored in the SRAM storage unit and the result is output to the storage bit line RBL.

Citation Information

Cited By

  • Calculation unit for simulating computer chip and control method thereof

    CN119337894A