CN107832841A - The power consumption optimization method and circuit of a kind of neural network chip - Google Patents
The power consumption optimization method and circuit of a kind of neural network chip Download PDFInfo
- Publication number
- CN107832841A CN107832841A CN201711121900.1A CN201711121900A CN107832841A CN 107832841 A CN107832841 A CN 107832841A CN 201711121900 A CN201711121900 A CN 201711121900A CN 107832841 A CN107832841 A CN 107832841A
- Authority
- CN
- China
- Prior art keywords
- convolution
- unit
- power domain
- control unit
- matrix
- 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.)
- Granted
Links
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
- G06N3/06—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
-
- 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
- G06N3/08—Learning methods
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
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)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Neurology (AREA)
- Power Sources (AREA)
- Design And Manufacture Of Integrated Circuits (AREA)
Abstract
The present invention provides a kind of power consumption optimization method and circuit of neural network chip, a power domain is separately provided for each convolutional calculation Internet, a power domain is separately provided for each convolution algorithm unit, the data block treated in convolution matrix is all connected with a gate controlled clock unit by behavior unit, the power supply of each row of data block;The n row data blocks that convolution matrix is treated by matrix resolution unit are analyzed, analysis result controls convolutional calculation Internet power domain switch control unit, convolution unit power domain switch control unit and convolution unit clock switch control unit by power consumption control unit, so as to control each convolutional calculation Internet power domain, each convolution power domain or the on or off of each gate controlled clock unit.Single neuron processing unit and whole neural net layer are made multi-level power domain by the present invention, can dynamically be turned off according to demand respectively, so as to effectively reduce the power consumption consumed during convolutional neural networks circuit computing.
Description
Technical field
The present invention relates to chip technology field, the power consumption optimization method and circuit of more particularly to a kind of neural network chip.
Background technology
With the rise of AI industry, artificial intelligence special chip is also quickly developing.And current manual is intelligent
One big problem of chip is precisely due to the complexity of deep learning neutral net, causes computing circuit very huge, this is resulted in
Chip cost is high and chip power-consumption is high.If can be from the characteristic of deep learning, further reduction deep learning be artificial
The cost and power consumption of intelligent chip are all significantly.
The information in this world is complicated, but the information of human brain processing is sparse.We directly can not possibly locate firmly
The numerous and jumbled input of sense organ is managed, it is necessary to which the process of an information extraction, this process are referred to as abstract in human brain.
Why DeepLearning propagates its belief on a large scale, exactly because it simulates being abstracted for human brain processing information to a certain extent
Journey.In terms of Neuscience, neuroscientist also found the sparse activity of neuron.2001, Attwell et al. was based on
On the observational learning of cerebral energy consumption, thus it is speculated that neuron coding work mode has openness and distributivity.2003
The neuron that Lennie et al. estimation brains are activated simultaneously only has 1~4%, further demonstrates that the openness of neuron work.
In terms of signal, i.e., neuron is simultaneously only to the small part selective response of input signal, and a large amount of signals are by screen deliberately
Cover, can so improve the precision of study, faster and better extract sparse features.It is openness be exactly herein map it is past
Matrix has many 0 elements.Just because of this, openness during neural network computing is the big characteristic of one, and the present invention is just
It is the openness matrix operation occurred for convolutional neural networks in calculating process, it is proposed that targetedly optimised power consumption side
Method.This method can effectively reduce the power consumption consumed during convolutional neural networks circuit computing.
The content of the invention
The technical problem to be solved in the present invention, it is the power consumption optimization method and circuit that a kind of neural network chip is provided,
To effectively reduce the power consumption consumed during convolutional neural networks circuit computing.
What the inventive method was realized in:A kind of power consumption optimization method of neural network chip, the neural network chip
Including plural layer convolutional calculation Internet, each convolutional calculation Internet includes a plurality of convolution algorithm units, each convolution fortune
The computing that unit is responsible for treating a full line data block corresponding to the convolution kernel height of convolution matrix is calculated, this treats that convolution matrix includes n
Row data block is simultaneously stored in corresponding hidden layer matrix memory cell, and the power consumption optimization method is:
Step S1, a power domain is separately provided for each convolutional calculation Internet, is convolutional calculation Internet power domain, and
Connect a convolution calculating network layer power domain switch control unit;
A power domain is separately provided for each convolution algorithm unit, is convolution power domain, and connects a convolution unit power supply
Domain switch control unit;
The data block treated in convolution matrix is all connected with a gated clock list by behavior unit, the power supply of each row of data block
Member, each gate controlled clock unit reconnect convolution unit clock switch control unit;
Step S2, the n row data blocks that convolution matrix is treated by matrix resolution unit are analyzed, and analysis result passes through one
Power consumption control unit controls the convolutional calculation Internet power domain switch control unit, the convolution unit power supply domain switch control
Unit processed and convolution unit clock switch control unit, so as to control each convolutional calculation Internet power domain, each convolution power supply
Domain or the on or off of each gate controlled clock unit.
The n rows data block that the matrix resolution unit treats convolution matrix carries out analysis process and is:
(1) treat convolution matrix according to the size of convolution kernel to be progressively scanned, judge each data in a full line one by one
Whether block is complete zero, if some data block is complete zero, clock can be closed by marking the data block;
(2) complete a full line data block judgement after, then to all data blocks in a full line whether generally complete zero
Do and once judge, if explanation can integrally close the clock of a full line, be then noted for the computing one full line data block
Convolution algorithm unit can close convolution power domain;
(3) whether the last data block for judging entirely to treat convolution matrix again is complete zero, if it is marks the whole convolution
Calculating network layer power domain can close.
What circuit of the present invention was realized in:A kind of optimised power consumption circuit of neural network chip, the neural network chip
Including plural layer convolutional calculation Internet, each convolutional calculation Internet includes a plurality of convolution algorithm units, each convolution fortune
The computing that unit is responsible for treating a full line data block corresponding to the convolution kernel height of convolution matrix is calculated, this treats that convolution matrix includes n
Row data block is simultaneously stored in corresponding hidden layer matrix memory cell;
The optimised power consumption circuit includes corresponding the power domain set control electricity with a plurality of convolutional calculation Internets
Road, each power domain control circuit include matrix resolution unit, power consumption control unit, convolutional calculation Internet power supply domain switch control
Unit processed, convolution unit power domain switch control unit, convolution unit clock switch control unit, convolutional calculation Internet power supply
Domain, n convolution power domain and n gate controlled clock unit;
The matrix resolution unit connects the corresponding hidden layer matrix memory cell and the power consumption control list respectively
Member;The power consumption control unit connects convolutional calculation Internet power domain switch control unit respectively, convolution unit power domain is opened
Close control unit and convolution unit clock switch control unit;The convolutional calculation Internet power domain switch control unit connection
The convolutional calculation Internet power domain;The convolution unit power domain switch control unit connects n convolution power domain respectively;
The convolution unit clock switch control unit connects n gate controlled clock unit respectively, and n gate controlled clock unit corresponds to respectively to be connected
Connect n row data blocks.
Further, the matrix resolution unit in each power domain control circuit is combined into a matrix resolution unit,
Power consumption control unit in each power domain control circuit is combined into a power consumption control unit.
The invention has the advantages that:Single neuron processing unit and whole neural net layer are made multilayer by the present invention
Secondary power domain, can dynamically it turn off according to demand respectively, so as to effectively reduce convolutional neural networks circuit computing mistake
The power consumption consumed in journey.
Brief description of the drawings
The present invention is further illustrated in conjunction with the embodiments with reference to the accompanying drawings.
Fig. 1 is schematic block circuit diagram of the present invention.
Embodiment
The power consumption optimization method of the neural network chip of the present invention, the neural network chip include plural layer convolutional calculation net
Network layers, each convolutional calculation Internet include a plurality of convolution algorithm units, and each convolution algorithm unit is responsible for treating convolution square
The computing of a full line data block corresponding to the convolution kernel height of battle array, this treats that convolution matrix includes n rows data block and is stored in correspondingly
Hidden layer matrix memory cell in, the power consumption optimization method is:
Step S1, a power domain is separately provided for each convolutional calculation Internet, is convolutional calculation Internet power domain, and
Connect a convolution calculating network layer power domain switch control unit;
A power domain is separately provided for each convolution algorithm unit, is convolution power domain, and connects a convolution unit power supply
Domain switch control unit;
The data block treated in convolution matrix is all connected with a gated clock list by behavior unit, the power supply of each row of data block
Member, each gate controlled clock unit reconnect convolution unit clock switch control unit;
Step S2, the n row data blocks that convolution matrix is treated by matrix resolution unit are analyzed, and analysis result passes through one
Power consumption control unit controls the convolutional calculation Internet power domain switch control unit, the convolution unit power supply domain switch control
Unit processed and convolution unit clock switch control unit, so as to control each convolutional calculation Internet power domain, each convolution power supply
Domain or the on or off of each gate controlled clock unit.
Wherein, the matrix resolution unit treats the n rows data block of convolution matrix and carries out analysis process and be:
(1) treat convolution matrix according to the size of convolution kernel to be progressively scanned, judge each data in a full line one by one
Whether block is complete zero, if some data block is complete zero, clock can be closed by marking the data block;
(2) complete a full line data block judgement after, then to all data blocks in a full line whether generally complete zero
Do and once judge, if explanation can integrally close the clock of a full line, be then noted for the computing one full line data block
Convolution algorithm unit can close convolution power domain;
(3) whether the last data block for judging entirely to treat convolution matrix again is complete zero, if it is marks the whole convolution
Calculating network layer power domain can close.
Refer to shown in Fig. 1, there is provided neural network chip of the invention includes plural layer convolutional calculation Internet, Mei Gejuan
Product calculating network layer includes a plurality of convolution algorithm units, and the convolution kernel that each convolution algorithm unit is responsible for treating convolution matrix is high
The computing of a full line data block corresponding to degree, this treats that convolution matrix includes n rows data block and is stored in corresponding hiding layer matrix
In memory cell;The neural network chip includes synapse input block, input layer convolution algorithm unit, convolution kernel;Often
One convolution calculating network layer is additionally provided with an activation primitive arithmetic element, a pond processing unit and one hides layer matrix storage list
Member;
The optimised power consumption circuit includes corresponding the power domain set control electricity with a plurality of convolutional calculation Internets
Road, each power domain control circuit include matrix resolution unit, power consumption control unit, convolutional calculation Internet power supply domain switch control
Unit processed, convolution unit power domain switch control unit, convolution unit clock switch control unit, convolutional calculation Internet power supply
Domain, n convolution power domain and n gate controlled clock unit;In a specific embodiment, each power domain control circuit bag
A matrix resolution unit and a power consumption control unit are included, but the invention is not restricted to this, can also be:Each power domain control
Matrix resolution unit in circuit is combined into a matrix resolution unit, the power consumption control list in each power domain control circuit
Member is combined into a power consumption control unit.
The matrix resolution unit connects the corresponding hidden layer matrix memory cell and the power consumption control list respectively
Member;The power consumption control unit connects convolutional calculation Internet power domain switch control unit respectively, convolution unit power domain is opened
Close control unit and convolution unit clock switch control unit;The convolutional calculation Internet power domain switch control unit connection
The convolutional calculation Internet power domain;The convolution unit power domain switch control unit connects n convolution power domain respectively;
The convolution unit clock switch control unit connects n gate controlled clock unit respectively, and n gate controlled clock unit corresponds to respectively to be connected
Connect n row data blocks.
Wherein,
The synapse input block is responsible for the value of synapse collection being sent to the input layer convolution algorithm
Unit;
The data that the input layer convolution algorithm unit is responsible for inputting synapse according to convolution kernel are carried out at convolution
Reason, and after convolution is completed, convolution value is sent to the activation primitive arithmetic element;
The function arithmetic unit living carries out activation primitive computing to convolution value, due to the characteristic of activation primitive, now produces
Raw data matrix has been openness matrix;
The pond processing unit is responsible for hiding to that will treat that convolution matrix is sent to after the matrix progress pond processing after activation
Layer matrix memory cell is stored;The different matrixes that different convolution kernels are dealt are stored to different addresses, such as in figure
Matrix A and matrix B;
The convolutional calculation Internet is responsible for the convolution algorithm of corresponding neutral net hidden layer, is parallel processing knot
Structure, i.e., each convolution algorithm unit are responsible for treating the computing of a full line corresponding to the convolution kernel height of convolution matrix;
Convolution kernel has size, such as 4x4 convolution kernel, in 320x180 image, if to sweep a full line, that
Convolution kernel 1 is just responsible for a full line as 320x4 in image, and convolution kernel 2 is just responsible in image the 5th row to eighth row, this
A full line corresponding to the convolution kernel height of sample, the present invention in treat convolution matrix n row data blocks row be and convolution kernel height
A corresponding full line;
The matrix resolution unit is responsible for analyzing the matrix of hidden layer successively, and convolution algorithm is being carried out to the matrix
When, determine whether to close a certain data block clock, if the power domain of convolution algorithm unit corresponding to closing, or whether
Close the power domain of whole neural net layer convolution circuit;
The power consumption control unit is responsible for when starting to carry out convolution algorithm to matrix to convolution unit power supply domain switch control
Unit, convolution unit clock switch control unit and neutral net flood power domain switch control unit processed are controlled, to every
The clock and power domain and flood neutral net power domain of individual convolution algorithm unit are finely controlled;
The convolution algorithm unit starts to carry out convolution algorithm to matrix after matrix resolution unit completes matrix analysis,
In calculating process, clock and power supply of the power consumption control unit according to the output control that analysis marks to each convolution algorithm unit
The shut-off in domain and the shut-off of flood neutral net power domain, reach the effect for saving power consumption;
Accordingly, the optimization process of optimised power consumption circuit of the invention is:
1st, the value that synapse gathers is sent to input layer convolution algorithm unit by synapse input block;Input layer
The data that convolution algorithm unit is responsible for inputting synapse according to convolution kernel carry out process of convolution;After completing convolution, convolution
Value is sent to the first convolution calculating network layer (i.e. the first hidden layer), that is, is sent to the first activation primitive arithmetic element and enters line activating letter
Number computing;It is hidden to that will treat that convolution matrix is sent to first after the matrix progress pond processing after activation by the first pond processing unit again
Layer matrix memory cell is hidden to be stored.
2 and then be responsible for analyzing the matrix of the first convolution calculating network layer successively by matrix resolution unit, still with
Exemplified by 4x4 convolution kernel, concrete analysis process is:
Since the matrix upper left corner, judge whether first 4x4 data block is complete zero, if complete zero, then marks convolution
First data block in the computing matrix of arithmetic element 1 can close clock, and then move right a data, then takes one
4x4 data blocks determine whether complete zero, if complete zero, then mark second number in the computing matrix of convolution algorithm unit 1
Clock can be closed according to block, by that analogy, last 4x4 data block of the row is determined always, has judged convolution algorithm list
After whether first 1 last data block can close clock, then whether can be with to data block corresponding to whole convolution algorithm units 1
Overall closing clock, which is done, once to be judged, if all data blocks can integrally close clock, then convolution during the computing matrix
Arithmetic element 1 can close power domain;Then again to judge that the method for convolution algorithm unit 1 is sentenced to convolution algorithm unit 2
It is disconnected, until convolution algorithm unit n completes to judge and whether mark, the last data block for judging whole matrix again are complete zero, such as
Fruit can close for the power domain of complete zero whole convolutional calculation Internet.Then annotation results are sent to power consumption control unit, and
By power consumption control unit by controlling convolutional calculation Internet power domain switch control unit, convolution unit power supply domain switch to control
Unit and convolution unit clock switch control unit, so as to control each convolutional calculation Internet power domain, each convolution power domain
Or the on or off of each gate controlled clock unit, so as to reach the purpose of optimization power consumption.
3rd, after the convolution algorithm for completing this layer, continue to start next layer of neural network computing, i.e. the first convolutional calculation net
After network layers complete convolution, convolution results are sent to the second convolution calculating network layer, that is, are sent to the second activation primitive arithmetic element, so
The second pond processing unit is being sent to afterwards, is continuing to start the computing of the second layer, the computing mode of each hidden layer can be with first
Layer is consistent, full articulamentum to the last, completes result and judges.
As can be seen here, single neuron processing unit and whole neural net layer are made multi-level power supply by the present invention
Domain, can dynamic turns off according to demand respectively, consumed so as to effectively reduce during convolutional neural networks circuit computing
Power consumption.
Although the foregoing describing the embodiment of the present invention, those familiar with the art should manage
Solution, the specific embodiment described by us are merely exemplary, rather than for the restriction to the scope of the present invention, are familiar with this
The equivalent modification and change that the technical staff in field is made in the spirit according to the present invention, should all cover the present invention's
In scope of the claimed protection.
Claims (4)
1. a kind of power consumption optimization method of neural network chip, the neural network chip includes plural layer convolutional calculation Internet,
Each convolutional calculation Internet includes a plurality of convolution algorithm units, and each convolution algorithm unit is responsible for treating the volume of convolution matrix
The computing of a full line data block corresponding to product core height, this treats that convolution matrix includes n rows data block and is stored in corresponding hide
In layer matrix memory cell, n is the individual data tile height for entirely treating convolution matrix height divided by convolution kernel processing, and its feature exists
In:The power consumption optimization method is:
Step S1, a power domain is separately provided for each convolutional calculation Internet, is convolutional calculation Internet power domain, and connect
One convolution calculating network layer power domain switch control unit;
A power domain is separately provided for each convolution algorithm unit, is convolution power domain, and connects a convolution unit power domain and opens
Close control unit;
The data block treated in convolution matrix is all connected with a gate controlled clock unit by behavior unit, the power supply of each row of data block, often
Individual gate controlled clock unit reconnects convolution unit clock switch control unit;
Step S2, the n row data blocks that convolution matrix is treated by matrix resolution unit are analyzed, and analysis result passes through power consumption control
Unit processed controls the convolutional calculation Internet power domain switch control unit, the convolution unit power domain switch control unit
And convolution unit clock switch control unit, so as to control each convolutional calculation Internet power domain, each convolution power domain or each
The on or off of gate controlled clock unit.
A kind of 2. optimised power consumption circuit of neural network chip according to claim 1, it is characterised in that:
The n rows data block that the matrix resolution unit treats convolution matrix carries out analysis process and is:
(1) treat convolution matrix according to the size of convolution kernel to be progressively scanned, judge that each data block is in a full line one by one
No is complete zero, if some data block is complete zero, clock can be closed by marking the data block;
(2) after the judgement of a full line data block is completed, then whether generally complete 0 one is done to all data blocks in a full line
Secondary judgement, if explanation can integrally close the clock of a full line, then it is noted for the volume of the computing one full line data block
Product arithmetic element can close convolution power domain;
(3) whether the last data block for judging entirely to treat convolution matrix again is complete zero, if it is marks the whole convolutional calculation
Internet power domain can close.
3. a kind of optimised power consumption circuit of neural network chip, the neural network chip includes plural layer convolutional calculation Internet,
Each convolutional calculation Internet includes a plurality of convolution algorithm units, and each convolution algorithm unit is responsible for treating the volume of convolution matrix
The computing of a full line data block corresponding to product core height, this treats that convolution matrix includes n rows data block and is stored in corresponding hide
In layer matrix memory cell, it is characterised in that:
The optimised power consumption circuit includes corresponding the power domain control circuit set with a plurality of convolutional calculation Internets, often
Individual power domain control circuit includes matrix resolution unit, power consumption control unit, convolutional calculation Internet power supply domain switch control list
Member, convolution unit power domain switch control unit, convolution unit clock switch control unit, convolutional calculation Internet power domain, n
Individual convolution power domain and n gate controlled clock unit;
The matrix resolution unit connects the corresponding hidden layer matrix memory cell and the power consumption control unit respectively;Institute
State power consumption control unit and connect convolutional calculation Internet power domain switch control unit, the control of convolution unit power supply domain switch respectively
Unit and convolution unit clock switch control unit;The convolutional calculation Internet power domain switch control unit connects the volume
Product calculating network layer power domain;The convolution unit power domain switch control unit connects n convolution power domain respectively;The volume
Product unit clock switch control unit connects n gate controlled clock unit respectively, and each gate controlled clock unit is connected respectively one
Row data block.
A kind of 4. optimised power consumption circuit of neural network chip according to claim 3, it is characterised in that:Each electricity
Matrix resolution unit in source domain control circuit is combined into a matrix resolution unit, the work(in each power domain control circuit
Consumption control unit is combined into a power consumption control unit.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711121900.1A CN107832841B (en) | 2017-11-14 | 2017-11-14 | Power consumption optimization method and circuit of neural network chip |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711121900.1A CN107832841B (en) | 2017-11-14 | 2017-11-14 | Power consumption optimization method and circuit of neural network chip |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107832841A true CN107832841A (en) | 2018-03-23 |
CN107832841B CN107832841B (en) | 2020-05-05 |
Family
ID=61655395
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711121900.1A Active CN107832841B (en) | 2017-11-14 | 2017-11-14 | Power consumption optimization method and circuit of neural network chip |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107832841B (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108647774A (en) * | 2018-04-23 | 2018-10-12 | 福州瑞芯微电子股份有限公司 | A kind of neural network method and circuit of optimization sparsity matrix operation |
CN108712630A (en) * | 2018-04-19 | 2018-10-26 | 安凯(广州)微电子技术有限公司 | A kind of internet camera system and its implementation based on deep learning |
CN109948775A (en) * | 2019-02-21 | 2019-06-28 | 山东师范大学 | The configurable neural convolutional network chip system of one kind and its configuration method |
CN111199273A (en) * | 2019-12-31 | 2020-05-26 | 深圳云天励飞技术有限公司 | Convolution calculation method, device, equipment and storage medium |
WO2023246077A1 (en) * | 2022-06-20 | 2023-12-28 | 东南大学 | Digital integrated circuit optimization method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150262057A1 (en) * | 2012-03-29 | 2015-09-17 | International Business Machines Corporation | Synaptic, dendritic, somatic, and axonal plasticity in a network of neural cores using a plastic multi-stage crossbar switching |
CN106127302A (en) * | 2016-06-23 | 2016-11-16 | 杭州华为数字技术有限公司 | Process the circuit of data, image processing system, the method and apparatus of process data |
WO2017006512A1 (en) * | 2015-07-08 | 2017-01-12 | 株式会社デンソー | Arithmetic processing device |
-
2017
- 2017-11-14 CN CN201711121900.1A patent/CN107832841B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150262057A1 (en) * | 2012-03-29 | 2015-09-17 | International Business Machines Corporation | Synaptic, dendritic, somatic, and axonal plasticity in a network of neural cores using a plastic multi-stage crossbar switching |
WO2017006512A1 (en) * | 2015-07-08 | 2017-01-12 | 株式会社デンソー | Arithmetic processing device |
CN106127302A (en) * | 2016-06-23 | 2016-11-16 | 杭州华为数字技术有限公司 | Process the circuit of data, image processing system, the method and apparatus of process data |
Non-Patent Citations (5)
Title |
---|
COLIN LIN: "Design space exploration for sparse matrix-matrix multiplication on FPGAs", 《INTERNATIONAL JOURNAL OF CIRCUIT THEORY AND APPLICATIONS》 * |
DE LA ROCHE SAINT ANDRE: "Tunnel Effect in CNNs:Image Reconstruction From Max-Switch Locations", 《IEEE SIGNAL PROCESSING LETTERS》 * |
伏龙: "SPI接口在神经网络加速器芯片中的设计与应用", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
方睿,等: "卷积神经网络的FPGA并行加速方案设计", 《计算机工程与应用》 * |
陆志坚: "基于FPGA的卷积神经网络并行结构研究", 《中国博士学位论文全文数据库信息科技辑》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108712630A (en) * | 2018-04-19 | 2018-10-26 | 安凯(广州)微电子技术有限公司 | A kind of internet camera system and its implementation based on deep learning |
CN108647774A (en) * | 2018-04-23 | 2018-10-12 | 福州瑞芯微电子股份有限公司 | A kind of neural network method and circuit of optimization sparsity matrix operation |
CN108647774B (en) * | 2018-04-23 | 2020-11-20 | 瑞芯微电子股份有限公司 | Neural network method and circuit for optimizing sparsity matrix operation |
CN109948775A (en) * | 2019-02-21 | 2019-06-28 | 山东师范大学 | The configurable neural convolutional network chip system of one kind and its configuration method |
CN111199273A (en) * | 2019-12-31 | 2020-05-26 | 深圳云天励飞技术有限公司 | Convolution calculation method, device, equipment and storage medium |
CN111199273B (en) * | 2019-12-31 | 2024-03-26 | 深圳云天励飞技术有限公司 | Convolution calculation method, device, equipment and storage medium |
WO2023246077A1 (en) * | 2022-06-20 | 2023-12-28 | 东南大学 | Digital integrated circuit optimization method |
Also Published As
Publication number | Publication date |
---|---|
CN107832841B (en) | 2020-05-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107832841A (en) | The power consumption optimization method and circuit of a kind of neural network chip | |
Mozafari et al. | Bio-inspired digit recognition using reward-modulated spike-timing-dependent plasticity in deep convolutional networks | |
Indiveri et al. | Memory and information processing in neuromorphic systems | |
US8515885B2 (en) | Neuromorphic and synaptronic spiking neural network with synaptic weights learned using simulation | |
CN109460817A (en) | A kind of convolutional neural networks on piece learning system based on nonvolatile storage | |
CN105184366B (en) | A kind of time-multiplexed general neural network processor | |
Mendoza et al. | Interval type-2 fuzzy logic and modular neural networks for face recognition applications | |
JP6328787B2 (en) | Neuromorphological graph compression method, system, and computer program product using multiple associative memories | |
Yakopcic et al. | Energy efficient perceptron pattern recognition using segmented memristor crossbar arrays | |
CN106570516A (en) | Obstacle recognition method using convolution neural network | |
Zhao et al. | A memristor-based spiking neural network with high scalability and learning efficiency | |
CN109886422A (en) | Model configuration method, device, electronic equipment and read/write memory medium | |
CN105913119B (en) | The heterogeneous polynuclear heart class brain chip and its application method of ranks interconnection | |
CN108416327A (en) | A kind of object detection method, device, computer equipment and readable storage medium storing program for executing | |
CN106874688A (en) | Intelligent lead compound based on convolutional neural networks finds method | |
CN101971166A (en) | Neuromorphic circuit | |
CN107369108A (en) | Multilayer artificial neural networks and its control method | |
CN107273972A (en) | It is a kind of based on resistive device and to adapt to excite the neuromorphic system and implementation method of neuron | |
CN106980830A (en) | One kind is based on depth convolutional network from affiliation recognition methods and device | |
Ma et al. | A memristive neural network model with associative memory for modeling affections | |
CN109214048A (en) | Utilize mixing CMOS- memristor fuzzy logic gate circuit and its design method | |
Liu et al. | Numerical P systems with Boolean condition | |
Huang et al. | Differential evolution-based convolutional neural networks: An automatic architecture design method for intrusion detection in industrial control systems | |
Yang et al. | A novel parallel merge neural network with streams of spiking neural network and artificial neural network | |
Fritscher et al. | Mitigating the effects of RRAM process variation on the accuracy of artificial neural networks |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CP01 | Change in the name or title of a patent holder |
Address after: 350000 building, No. 89, software Avenue, Gulou District, Fujian, Fuzhou 18, China Patentee after: Ruixin Microelectronics Co., Ltd Address before: 350000 building, No. 89, software Avenue, Gulou District, Fujian, Fuzhou 18, China Patentee before: Fuzhou Rockchips Electronics Co.,Ltd. |
|
CP01 | Change in the name or title of a patent holder |