CN113642706A - Neuron network unit, convolution operation module and convolution neural network - Google Patents
Neuron network unit, convolution operation module and convolution neural network Download PDFInfo
- Publication number
- CN113642706A CN113642706A CN202110913994.6A CN202110913994A CN113642706A CN 113642706 A CN113642706 A CN 113642706A CN 202110913994 A CN202110913994 A CN 202110913994A CN 113642706 A CN113642706 A CN 113642706A
- Authority
- CN
- China
- Prior art keywords
- transistor
- isolation branch
- transmission transistor
- reading
- reverse
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 210000002569 neuron Anatomy 0.000 title claims abstract description 13
- 238000013528 artificial neural network Methods 0.000 title claims description 17
- 238000002955 isolation Methods 0.000 claims abstract description 46
- 230000005540 biological transmission Effects 0.000 claims abstract description 35
- 230000003068 static effect Effects 0.000 claims abstract description 13
- 239000003990 capacitor Substances 0.000 claims description 15
- 238000013139 quantization Methods 0.000 claims description 15
- 210000004027 cell Anatomy 0.000 claims description 6
- 238000013527 convolutional neural network Methods 0.000 claims description 2
- 238000007599 discharging Methods 0.000 claims description 2
- 238000010586 diagram Methods 0.000 description 6
- 238000004364 calculation method Methods 0.000 description 2
- 238000005034 decoration Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 239000002699 waste material Substances 0.000 description 2
- 230000001133 acceleration Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Neurology (AREA)
- Logic Circuits (AREA)
Abstract
The invention provides a neuron network unit, which comprises a static random storage unit, a forward read isolation branch and a reverse read isolation branch; the static random access memory unit comprises a first transmission transistor and a second transmission transistor which are electrically connected in series, and two opposite interlocked first inverters and second inverters which are connected between the first transmission transistor and the second transmission transistor in parallel, wherein the forward direction reading isolation branch circuit is connected between the first transmission transistor and the two opposite interlocked inverters and is used for converting an externally input digital voltage into an analog current to be output according to a control signal stored in the static random access memory unit; the reverse readout isolation branch is connected between the second transmission transistor and the two opposite interlocked inverters and used for converting an externally input digital voltage into an analog current according to a control signal stored in the static random access memory unit and outputting the analog current.
Description
Technical Field
The invention relates to the field of integrated circuit design, in particular to a neuron network unit, a convolution operation module and a convolution neural network.
Background
With the development of the big data era, artificial intelligence has become a very important subject field, and a neural network special chip is an important hardware tool for a computing system to efficiently complete neural network computation. Traditional computing architectures employ a von neumann architecture with separate computation and storage, and memory bandwidth and memory power consumption in von neumann architectures have begun to dominate computing bandwidth and energy under big data trends. A significant portion of this power consumption is spent on memory and data handling by the compute unit. The memory calculation which takes the memory as the leading part reduces huge time and power consumption expense brought by data transportation to a great extent through the combination of the neural network algorithm and the storage hardware architecture.
The conventional convolutional neural network memory (CIM SRAM) uses the word line and bit line architecture of the conventional SRAM, so that when data is input, the data received by each row of word lines is the same set of input data. Although different convolution kernels can be used for performing the multiply-add operation when the convolution operation is performed on the same group of data, based on the current simple convolution neural network algorithm, the maximum number of 16-20 groups of convolution kernels are used for the data of the same group of convolution windows, so that for the SRAM array, the resource waste is caused by the fact that only 16-20 columns of bit lines are used for performing the operation.
Disclosure of Invention
The technical problem to be solved by the invention is to provide a neuron network unit, a convolution operation module and a convolution neural network, so that the resource utilization rate of an array unit is improved, and the power consumption is reduced.
In order to solve the above problems, the present invention provides a neural network unit, which includes a static random access memory unit, a forward sense isolation branch, and a reverse sense isolation branch; the static random access memory unit comprises a first transmission transistor and a second transmission transistor which are electrically connected in series, and two opposite interlocked first inverters and second inverters which are connected between the first transmission transistor and the second transmission transistor in parallel, wherein the forward direction reading isolation branch circuit is connected between the first transmission transistor and the two opposite interlocked inverters and is used for converting an externally input digital voltage into an analog current to be output according to a control signal stored in the static random access memory unit; the reverse readout isolation branch is connected between the second transmission transistor and the two opposite interlocked inverters and used for converting an externally input digital voltage into an analog current according to a control signal stored in the static random access memory unit and outputting the analog current.
Optionally, the forward read isolation branch comprises a first readout transistor and a second readout transistor connected in series; the grid electrode of the first reading transistor is connected between the first transmission transistor and the two oppositely interlocked inverters and is used as the control end of the forward reading isolation branch circuit, and the drain/source electrode of the first reading transistor is connected with the working voltage; the grid electrode of the second reading transistor is used as an external input end, and the source/drain electrode of the second reading transistor is used as the output end of the forward reading isolation branch.
Optionally, the reverse read isolation branch includes a third readout transistor and a fourth readout transistor connected in series; the grid electrode of the third reading transistor is connected between the second transmission transistor and the two oppositely interlocked inverters and is used as the control end of the reverse reading isolation branch circuit, and the drain/source electrode of the third reading transistor is grounded; the grid electrode of the fourth reading transistor is used as an external input end, and the source/drain electrode of the fourth reading transistor is used as the output end of the reverse reading isolation branch circuit.
The invention also provides a convolution operation module, which comprises a convolution operation sub-block and a pulse frequency quantization unit; the convolution operation sub-block comprises an array formed by N multiplied by N neuron network units, wherein N is a positive integer, and the array comprises: the first transmission transistor and the second transmission transistor of the neuron network unit are connected with a write word line of the array, a drain/source electrode of the first transmission transistor and a source/drain electrode of the second transmission transistor are respectively connected with a bit line and a reverse bit line, the external input end of the forward isolation branch is connected with a row word line, the external input end of the reverse isolation branch is connected with a column word line, the output end of the forward isolation branch is connected with a row current output, and the output end of the reverse isolation branch is connected with a column current output; the pulse frequency quantization unit is connected with the output end of the convolution operation sub-block, converts the current output in the convolution operation sub-block into a pulse signal to carry out pulse frequency counting, and further converts the pulse signal into multi-bit digital output.
Optionally, the pulse frequency quantization unit includes an amplifier, a comparator, and a counter connected in series: the amplifier is connected with an integrating capacitor in parallel and is connected with the comparator in series to serve as an input end of the pulse frequency quantization unit; the output end of the comparator is further connected to two ends of the integrating capacitor through a discharge control switch, and the discharge control switch is used for discharging the integrating capacitor when the comparator outputs a high level; the integrating capacitor is charged by analog current input and then amplified by an amplifier to cause pulse output of a comparator, so that the current value is equivalently converted into pulse frequency; the counter is connected with the comparator in series and counts the number of pulses to form multi-bit digital output.
The invention also provides a convolution neural network which comprises an array formed by the convolution operation modules with the number of N multiplied by N, wherein N is a positive integer.
According to the invention, each row of word lines receives two groups of different data and respectively carries out convolution operation, more groups of convolution kernels can be used for the data of the same group of convolution windows, and resource waste is avoided.
Drawings
Fig. 1 is a schematic circuit structure diagram of a neural network unit according to an embodiment of the present invention.
Fig. 2 is a schematic circuit diagram of a convolution operation sub-block in a convolution operation module according to an embodiment of the present invention.
Fig. 3 is a schematic circuit diagram of a pulse quantization unit in a convolution operation module according to an embodiment of the present invention.
Detailed Description
The following describes in detail specific embodiments of the sram cell, the convolution operation module, and the convolution neural network according to the present invention with reference to the accompanying drawings.
In the following description, the source and drain connections of the transistor are described alternatively, and it should be noted that in other embodiments, the source and drain connections of the transistor may be equivalently replaced without affecting the actual electrical function.
Fig. 1 is a schematic circuit diagram of a neural network unit according to an embodiment of the present invention, which includes a sram cell M1, a forward sense isolation branch B1, and a reverse sense isolation branch B2. The static random access memory cell M1 is a 6T structure including a first pass transistor T1 and a second pass transistor T2 electrically connected in series, and two opposing interlocked first and second inverters D1 and D2 connected in parallel between the first and second pass transistors T1 and T2.
In this particular embodiment, the forward sense isolation branch B1 includes a first sense transistor R1 and a second sense transistor R2 connected in series; the gate of the first readout transistor R1 is connected between the first transfer transistor T1 and the two opposite interlocked inverters as the control terminal of the forward read isolation branch, and the drain of the first readout transistor R1 is connected to the operating voltage VDD; the gate of the second sense transistor R2 serves as the external input and the source of the second sense transistor R2 serves as the output of the forward sense isolation branch. The forward sense isolation branch B1 is connected between the first pass transistor and two opposite interlocked inverters, and is used for converting an externally input digital voltage into an analog current output according to a control signal stored in the SRAM cell M1.
In this embodiment, the reverse read isolation branch B2 includes a third read transistor R3 and a fourth read transistor R4 connected in series. The gate of the third readout transistor R3 is connected between the second transmission transistor T2 and the two opposite interlocked inverters, and serves as a control terminal of the reverse read isolation branch, and the drain of the third readout transistor R3 is grounded. The gate of the fourth pass transistor R4 serves as the external input and the source of the fourth sense transistor R4 serves as the output of the reverse sense isolation branch. The reverse sense isolation branch is connected between the second pass transistor T2 and two opposite interlocked inverters, and is used for converting an externally input digital voltage into an analog current output according to a control signal stored in the sram cell M1.
Next, a specific embodiment of the convolution operation module of the present invention will be given. The module comprises a convolution operation sub-block and a pulse frequency quantization unit. Fig. 2 is a schematic circuit diagram of a convolution operation sub-block in a convolution operation module according to an embodiment of the present invention. The convolution operation sub-block comprises an array formed by N multiplied by N neuron network units. In the present embodiment, 3 × 3 neuron network units C11 to C33 are described as an example. The first transmission transistor T1 and the second transmission transistor T2 of the neuron network unit are connected with a write word line WWL of the array, the drain electrode of the first transmission transistor T1 and the source electrode of the second transmission transistor are respectively connected with a bit line BL and a reverse bit line BLB, the external input end of the forward isolation branch is connected with a row word line LWL, the external input end of the reverse isolation branch is connected with a column word line CWL, the output end of the forward isolation branch is connected with a row current output Lo, and the output end of the reverse isolation branch is connected with a column current output Co. The array can realize convolution operation of a 4 x 8 input matrix by a 3 x 3 convolution kernel in three periods.
Fig. 3 is a schematic circuit diagram of a pulse quantization unit in a convolution operation module according to an embodiment of the present invention. The pulse frequency quantization unit is connected with the output end of the convolution operation sub-block, converts the current output in the convolution operation sub-block into a pulse signal to carry out pulse frequency counting, and further converts the pulse signal into multi-bit digital output. As one embodiment of the foregoing, as shown in fig. 2, the pulse frequency quantization unit includes an amplifier 21, a comparator 22, and a counter 23 connected in series. The amplifier 21 is connected in parallel with an integrating capacitor C and in series with a comparator 22 as an input of the pulse frequency quantization unit. The output end of the comparator 22 is further connected to two ends of the integrating capacitor C through a discharge control switch K, where the discharge control switch K is used to discharge the integrating capacitor C when the comparator 22 outputs a high level 1; the integrating capacitor C is charged by an analog current input and amplified by the amplifier 21 to generate a pulse output of the comparator 22, thereby equivalently converting the current value into a pulse frequency. Specifically, the inverted output of the comparator 22 is a discharge signal of the integrating capacitor C, the output of the comparator 22 is inverted by 0 → 1 each time the integrating capacitor C is charged to the inverting voltage of the comparator 22 by the analog current input, and when the output of the comparator 22 is 1, the integrating capacitor C is discharged, the output of the comparator 22 is 1 → 0, and further the comparator 22 forms a pulse output, and the frequency of the pulse formation is faster as the analog input current is larger. The counter 23 is connected in series with the comparator 22 to count the number of pulses and form a multi-bit digital output.
The convolution operation sub-block and the convolution operation module composed of the pulse frequency quantization unit and composed of the convolution operation sub-block and the convolution operation module composed of the convolution operation sub-block of fig. 2 and fig. 3 form an N × N array, namely a convolution neural network, and the acceleration calculation of the convolution neural network can be carried out.
The foregoing is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, various modifications and decorations can be made without departing from the principle of the present invention, and these modifications and decorations should also be regarded as the protection scope of the present invention.
Claims (6)
1. A neural network unit comprises a static random access memory unit, a forward read isolation branch and a reverse read isolation branch;
the static random access memory cell comprises a first transmission transistor and a second transmission transistor which are electrically connected in series, and two opposite interlocked first and second inverters which are connected between the first transmission transistor and the second transmission transistor in parallel, and is characterized in that;
the forward direction reading isolation branch circuit is connected between the first transmission transistor and the two oppositely interlocked phase inverters and used for converting digital voltage input from outside into analog current to be output according to the control signal stored in the static random access memory unit.
The reverse readout isolation branch is connected between the second transmission transistor and the two opposite interlocked inverters and used for converting an externally input digital voltage into an analog current according to a control signal stored in the static random access memory unit and outputting the analog current.
2. The neuron network unit of claim 1, wherein the forward read isolation branch comprises a first sense transistor and a second sense transistor in series; the grid electrode of the first reading transistor is connected between the first transmission transistor and the two oppositely interlocked inverters and is used as the control end of the forward reading isolation branch circuit, and the drain/source electrode of the first reading transistor is connected with the working voltage; the grid electrode of the second reading transistor is used as an external input end, and the source/drain electrode of the second reading transistor is used as the output end of the forward reading isolation branch.
3. The neuron network unit of claim 1, wherein the reverse read isolation branch comprises a third read transistor and a fourth read transistor in series; the grid electrode of the third reading transistor is connected between the second transmission transistor and the two oppositely interlocked inverters and is used as the control end of the reverse reading isolation branch circuit, and the drain/source electrode of the third reading transistor is grounded; the grid electrode of the fourth reading transistor is used as an external input end, and the source/drain electrode of the fourth reading transistor is used as the output end of the reverse reading isolation branch circuit.
4. A convolution operation module is characterized by comprising a convolution operation sub-block and a pulse frequency quantization unit;
the sub-block of convolution operations comprises an array of nxn neural network elements of claim 1, N being a positive integer, wherein:
the first transmission transistor and the second transmission transistor of the neuron network unit are connected with a write word line of the array, a drain/source electrode of the first transmission transistor and a source/drain electrode of the second transmission transistor are respectively connected with a bit line and a reverse bit line, the external input end of the forward isolation branch is connected with a row word line, the external input end of the reverse isolation branch is connected with a column word line, the output end of the forward isolation branch is connected with a row current output, and the output end of the reverse isolation branch is connected with a column current output;
the pulse frequency quantization unit is connected with the output end of the convolution operation sub-block, converts the current output in the convolution operation sub-block into a pulse signal to carry out pulse frequency counting, and further converts the pulse signal into multi-bit digital output.
5. The convolution operation module of claim 4, wherein the pulse frequency quantization unit comprises an amplifier, a comparator, and a counter connected in series:
the amplifier is connected with an integrating capacitor in parallel and is connected with the comparator in series to serve as an input end of the pulse frequency quantization unit;
the output end of the comparator is further connected to two ends of the integrating capacitor through a discharge control switch, and the discharge control switch is used for discharging the integrating capacitor when the comparator outputs a high level;
the integrating capacitor is charged by analog current input and then amplified by an amplifier to cause pulse output of a comparator, so that the current value is equivalently converted into pulse frequency;
the counter is connected with the comparator in series and counts the number of pulses to form multi-bit digital output.
6. A convolutional neural network comprising an array of N x N convolutional arithmetic modules of claim 4, where N is a positive integer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110913994.6A CN113642706A (en) | 2021-08-10 | 2021-08-10 | Neuron network unit, convolution operation module and convolution neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110913994.6A CN113642706A (en) | 2021-08-10 | 2021-08-10 | Neuron network unit, convolution operation module and convolution neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113642706A true CN113642706A (en) | 2021-11-12 |
Family
ID=78420503
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110913994.6A Pending CN113642706A (en) | 2021-08-10 | 2021-08-10 | Neuron network unit, convolution operation module and convolution neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113642706A (en) |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200125935A1 (en) * | 2017-06-21 | 2020-04-23 | Semiconductor Energy Laboratory Co., Ltd. | Semiconductor device having neural network |
| CN111126579A (en) * | 2019-11-05 | 2020-05-08 | 复旦大学 | Memory computing device suitable for binary convolution neural network computing |
| CN111652363A (en) * | 2020-06-08 | 2020-09-11 | 中国科学院微电子研究所 | Storage and calculation integrated circuit |
| CN112232502A (en) * | 2020-12-17 | 2021-01-15 | 中科院微电子研究所南京智能技术研究院 | Same or memory unit and memory array device |
| CN112435700A (en) * | 2019-08-26 | 2021-03-02 | 意法半导体国际有限公司 | In-memory compute, high density array |
| CN112562756A (en) * | 2020-12-15 | 2021-03-26 | 中国科学院上海微系统与信息技术研究所 | Radiation-resistant static random access memory cell and memory |
| CN112687308A (en) * | 2020-12-29 | 2021-04-20 | 中国科学院上海微系统与信息技术研究所 | Low-power consumption static random access memory unit and memory |
-
2021
- 2021-08-10 CN CN202110913994.6A patent/CN113642706A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200125935A1 (en) * | 2017-06-21 | 2020-04-23 | Semiconductor Energy Laboratory Co., Ltd. | Semiconductor device having neural network |
| CN112435700A (en) * | 2019-08-26 | 2021-03-02 | 意法半导体国际有限公司 | In-memory compute, high density array |
| CN111126579A (en) * | 2019-11-05 | 2020-05-08 | 复旦大学 | Memory computing device suitable for binary convolution neural network computing |
| CN111652363A (en) * | 2020-06-08 | 2020-09-11 | 中国科学院微电子研究所 | Storage and calculation integrated circuit |
| CN112562756A (en) * | 2020-12-15 | 2021-03-26 | 中国科学院上海微系统与信息技术研究所 | Radiation-resistant static random access memory cell and memory |
| CN112232502A (en) * | 2020-12-17 | 2021-01-15 | 中科院微电子研究所南京智能技术研究院 | Same or memory unit and memory array device |
| CN112687308A (en) * | 2020-12-29 | 2021-04-20 | 中国科学院上海微系统与信息技术研究所 | Low-power consumption static random access memory unit and memory |
Non-Patent Citations (2)
| Title |
|---|
| SYED AHMED AAMIR: ""An Accelerated LIF Neuronal Network Array for a Large-Scale Mixed-Signal Neuromorphic Architecture"", 《IEEE》, 27 June 2018 (2018-06-27) * |
| 朱航涛: ""基于神经元晶体管和忆阻器的Hopfield神经网络及其在联想记忆中的应用"", 《西南大学学报（自然科学版）》, vol. 40, no. 2, 31 December 2018 (2018-12-31) * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11948659B2 (en) | Sub-cell, mac array and bit-width reconfigurable mixed-signal in-memory computing module | |
| CN111816231B (en) | Memory computing device with double-6T SRAM structure | |
| CN114546335B (en) | Memory computing device for multi-bit input and multi-bit weight multiplication accumulation | |
| CN111816232B (en) | In-memory computing array device based on 4-pipe storage structure | |
| CN112558919B (en) | Memory computing bit unit and memory computing device | |
| US11574173B2 (en) | Power efficient near memory analog multiply-and-accumulate (MAC) | |
| CN113627601A (en) | Subunit, MAC array and analog-digital mixed memory computing module with reconfigurable bit width | |
| CN115039177A (en) | Low power consumption in-memory compute bit cell | |
| CN114937470B (en) | Fixed point full-precision memory computing circuit based on multi-bit SRAM unit | |
| US11693560B2 (en) | SRAM-based cell for in-memory computing and hybrid computations/storage memory architecture | |
| CN110176264A (en) | A kind of high-low-position consolidation circuit structure calculated interior based on memory | |
| CN117271436B (en) | SRAM-based current mirror complementary in-memory calculation macro circuit and chip | |
| Ha et al. | A 36.2 dB high SNR and PVT/leakage-robust eDRAM computing-in-memory macro with segmented BL and reference cell array | |
| CN113936717A (en) | Storage and calculation integrated circuit for multiplexing weight | |
| CN115390789A (en) | Magnetic tunnel junction calculation unit-based analog domain full-precision memory calculation circuit and method | |
| CN115588446A (en) | Memory operation circuit, memory calculation circuit and chip thereof | |
| CN114038492B (en) | Multiphase sampling memory internal computing circuit | |
| CN118034644A (en) | EDRAM-based high-density high-reliability in-memory computing circuit | |
| CN117130978A (en) | Charge domain in-memory computing circuit based on sparse tracking ADC and computing method thereof | |
| Zhang et al. | In-memory multibit multiplication based on bitline shifting | |
| CN115691613B (en) | Charge type memory internal calculation implementation method based on memristor and unit structure thereof | |
| CN114895869B (en) | Multi-bit memory computing device with symbols | |
| CN116204490A (en) | 7T memory circuit and multiply-accumulate operation circuit based on low-voltage technology | |
| CN113642706A (en) | Neuron network unit, convolution operation module and convolution neural network | |
| CN115910152A (en) | Charge domain memory calculation circuit and calculation circuit with positive and negative number operation function |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |