CN106056212B - A kind of artificial neural networks core - Google Patents
A kind of artificial neural networks core Download PDFInfo
- Publication number
- CN106056212B CN106056212B CN201610354206.3A CN201610354206A CN106056212B CN 106056212 B CN106056212 B CN 106056212B CN 201610354206 A CN201610354206 A CN 201610354206A CN 106056212 B CN106056212 B CN 106056212B
- Authority
- CN
- China
- Prior art keywords
- neuron
- address
- module
- router
- interface
- 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.)
- Active
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
Abstract
The present invention relates to a kind of artificial neural networks cores, including：One router-module, at least a neuron computing module, at least a nuclear control device, the neuron computing module and the nuclear control device correspond.The router-module receives and parses through outer input data, and by after parsing address information and aixs cylinder value information be sent into corresponding neuron computing module；And the neuron calculated result that the neuron computing module exports is sent to destination address, and send this frame data processing complement mark to the corresponding nuclear control device of the neuron computing module.Neuron calculated result is sent to the router-module for carrying out neuron calculating by the neuron computing module.The nuclear control device enters next neuron period after receiving described frame data processing complement mark.The networking of multiple artificial neural networks cores may be implemented using artificial neural networks core provided by the invention.
Description
Technical field
The invention belongs to artificial neural networks field, in particular to a kind of neuron computing unit.
Background technique
Artificial neural network is to use for reference computation model made of biological brain cynapse-neuronal structure develops, can be concurrently
Large-scale complex operation is carried out, and has the characteristics that nonlinearity, adaptivity.With biological brain neural network structure class
Seemingly, artificial neural network may be defined as by basic structures such as neuron computing unit, aixs cylinder unit, dendron unit, cynapse units
Composition.Neuron computing unit is most basic computing unit, can be carried out simple mathematical operation；Aixs cylinder unit is responsible for output mind
Through first calculated result, a neuron has an aixs cylinder；Dendron unit is the input that neuron calculates, and a neuron can have more
A dendron；Cynapse unit indicates the weight of a upper neuron axon and the connection of next neuron dendron；Between neuron and neuron
By aixs cylinder, cynapse, dendricity layering connection to constitute neural network.
Artificial neural network is coupled by multilayer neuron computing unit, passes through a large amount of distributed simple operation units
Interaction realizes that parallel complex nonlinear calculates, and has powerful information processing capability.In research artificial neural networks
When model, the common equipment such as general purpose computer that calculates often will appear the disadvantages of calculating efficiency is low, and time-consuming, and energy consumption is high.Closely
The characteristics of Nian Lai, part research institution is directed to artificial neural network, develops general neural computing chip, such as IBM
TrueNorth has tentatively achieved the purpose that carry out artificial neural network operation in chip.However, current majority neural network meters
Chip is calculated using by computing unit（Neuron）And connection unit（Cynapse matrix）Isolated mentality of designing, due to using
The layout type that fixed scale couples entirely, to limit the connection number of neuron in monokaryon, while in non-full connection nerve net
In the application of network, the waste for not coupling cynapse point storage resource will also result in.
Summary of the invention
In view of this, it is necessory to provide a kind of neural computing chip that flexible networking may be implemented.
A kind of artificial neural networks core, including：One router-module, at least a neuron computing module, at least one
Nuclear control device, the neuron computing module and the nuclear control device correspond；The router-module, for receiving and solving
Analyse outer input data, and by after parsing address information and aixs cylinder value information be sent into corresponding neuron computing module；It is described
Neuron computing module is sent to the router-module for carrying out neuron calculating, and by neuron calculated result；It is described
Router-module is also used to the neuron calculated result that the neuron computing module exports being sent to destination address, and to
The corresponding nuclear control device of the neuron computing module sends this frame data processing complement mark；The nuclear control device, for connecing
Enter next neuron period after receiving described frame data processing complement mark.
Compared with prior art, multiple artificial minds may be implemented using artificial neural networks core provided by the invention
Networking through network query function core realizes the flexible expansion and the effective use of storage resource for calculating core, enhances and calculate core
Adaptability avoids additional resource overhead.
Detailed description of the invention
Fig. 1 is the neuron computing unit schematic diagram that first embodiment of the invention provides.
Fig. 2 is the neuron computing module schematic diagram that second embodiment of the invention provides.
Fig. 3 is the artificial neural networks core schematic diagram that third embodiment of the invention provides.
Fig. 4 is the artificial neural networks core schematic diagram that fourth embodiment of the invention provides.
Fig. 5 is the internuclear transmitting data frame format of artificial neural networks core that fourth embodiment of the invention provides.
Fig. 6 is that artificial neural networks core router data provided by the invention receives and dispatches flow chart.
Fig. 7 is artificial neural networks nucleus neuron computing module calculated result transmission flow figure provided by the invention.
Fig. 8 is the artificial neural networks core schematic diagram that fifth embodiment of the invention provides.
Fig. 9 is artificial neural networks core networking schematic diagram provided by the invention.
Main element symbol description
| Artificial neural networks core | 10、20、30、40 |
| Router-module | 100、400 |
| Path control deivce | 110、410 |
| Upper transmitting and receiving caching and interface | 120 |
| Left transmitting and receiving caching and interface | 130 |
| Lower transmitting and receiving caching and interface | 140 |
| Right transmitting and receiving caching and interface | 150 |
| Route information table | 160 |
| Neuron-router interface | 170 |
| Peripheral sensory neuron-router interface | 471 |
| Nervus opticus member-router interface | 472 |
| Third nerve member-router interface | 473 |
| Fourth nerve member-router interface | 474 |
| Neuron computing module | 200 |
| Peripheral sensory neuron computing module | 200a |
| Nervus opticus member computing module | 200b |
| Third nerve member computing module | 200c |
| Fourth nerve member computing module | 200d |
| Neuron computing unit | 210 |
| Decoder module | 211 |
| Address weight module | 212 |
| Multiplier | 213 |
| Accumulator | 214 |
| Send caching and interface | 220 |
| Function information table | 230 |
| Nuclear control device | 300 |
| First nuclear control device | 300a |
| Second nuclear control device | 300b |
| Third nuclear control device | 300c |
| 4th nuclear control device | 300d |
The present invention that the following detailed description will be further explained with reference to the above drawings.
Specific embodiment
Below in conjunction with the accompanying drawings and the specific embodiments, artificial neural networks core provided by the invention is made further
It is described in detail.
Referring to Figure 1, first embodiment of the invention provides a kind of neuron computing unit 210, including decoder module 211,
Address weight module 212, multiplier 213, accumulator 214.
The decoder module 211 is for receiving neural network information and parsing.The neural network information includes address letter
Breath and aixs cylinder value information, address information therein is sent into the address weight module 212 by the decoder module 211, by aixs cylinder
Value information is sent into the multiplier 213.
Storage address weight is to list in the address weight module 212, realizes to being input to the address weight module 212
Address information matched.If input address is matched with stored address information in address weight module 212, address power
Molality block 212 exports corresponding weighted value to multiplier 213, and the weighted value of output is the numerical value changed in a certain range；If
Input address and stored address information mismatch, then address weight module 212 exports zero to multiplier 213.In addition, institute
Address effective marker state can also be arranged according to matching judgment result by stating address weight module 212.Specifically, if input address
It being matched with address information stored in address weight module 212, the address weight module 212 issues address effective marker,
If stored address information mismatches in input address and address weight module 212, the address weight module 212 does not change
Address effective marker state.
The multiplier 213 receives the weighted value that the address weight module 212 exports and the decoder module 211 is defeated
The weighted value is multiplied by aixs cylinder value out with aixs cylinder value, and product is sent into accumulator 214.
The calculated result that multiplier 213 exports is added up and is exported by the realization of accumulator 214.
It is integrated with calculating that the neuron computing unit 210 that first embodiment of the invention provides realizes addressing, by adopting
With addressing with calculate integrated design, breach layout type that fixed scale couples entirely to the limit of neuron computing unit number
System, simultaneously because taking the connecting mode of address matching, avoids the wasting of resources caused by useless connection, to realize flexibility
Expansion and the effective use of storage resource, enhance neural computing efficiency.
Second embodiment of the invention further provides for a kind of neuron computing module, for establishing connecting relation between neuron
And carry out neuron calculating.Fig. 2 is referred to, the neuron computing module 200 includes：Multiple neuron computing units 210, hair
Send caching and interface 220, function information table 230.
The multiple neuron computing unit 210 is received from external address information and aixs cylinder value information, and to described
Address information is judged, if the address information received is matched with the address information that itself is stored, to the letter received
Breath carries out neural computing, and calculated result and address effective marker are sent to the transmission caching and interface 220, if connecing
The address information received and the address information of itself storage mismatch, then do not carry out neural network meter to the information received
It calculates.
Each neuron computing unit further comprises in multiple neuron computing units described in the present embodiment：Decode mould
Block 211, address weight module 212, multiplier 213, accumulator 214.Neuron computing unit and the present invention in the present embodiment
210 structure of neuron computing unit and function that first embodiment provides are essentially identical, the mould of address weight described in the present embodiment
Block 212 is also used to match when input address with stored address information in address weight module 212, then address weight module
212 export corresponding weighted value to multiplier 213, and issue address effective marker to transmission caching and interface 220；If input
Address and stored address information mismatch, then address weight module 212 exports zero to multiplier 213, and unchanged
Location effective marker state.
The transmission caching and interface 220 are used to calculate the multiple neuron in external global clock gclk triggering
The calculated result and address effective marker of unit 210 are latched, and there is criterion in the address for resetting each neuron computing unit 210
Will, and specified neuron calculated result is exported to the function information table 230 and obtains transformation result.
The neuron calculated result is converted to corresponding functional value by the function information table 230.
Single neuron computing unit will be addressed and be calculated in the neuron computing module that second embodiment of the invention provides
It is combined into one, neuron addressing can be carried out simultaneously and calculates, by using addressing and calculate integrated design, breach fixed gauge
Limitation of the layout type that mould couples entirely to neuron computing unit number, simultaneously because taking the connection side of address matching
Formula avoids the wasting of resources caused by useless connection, to realize flexible expansion and the effective use of storage resource, enhances
Neural computing efficiency.
Third embodiment of the invention provides a kind of artificial neural networks core, including：Neuron computing module, nuclear control
Device.The neuron computing module enters different nerves for carrying out neuron calculating under the control of the nuclear control device
The transmission flow of first calculated result.The neuron computing module can be existing various neuron computing modules, this implementation
It is illustrated by taking the neuron computing module 200 that second embodiment of the invention provides as an example in example.Fig. 3 specifically is referred to, manually
Neural computing core 10 includes neuron computing module 200, nuclear control device 300.
The nuclear control device 300 is artificial by issuing neuron period N_peri and comparing pulse N_data_comp control
Neural computing core 10 enters the transmission flow of different neuron calculated results.When entering the new neuron period, pass through
The comparison of pulse N_data_comp triggering neuron calculated result, output are compared in output, and wait data processed result mark R_
N_data_done or N_data_null is effectively to enter the new neuron period.The nuclear control device 300 is in global clock
The lower sending cycle into new round neuron calculated result of gclk triggering.
The neuron computing module 200 is used to establish connecting relation and progress neuron calculating between neuron, with this hair
The neuron computing module structure and function that bright second embodiment provides are essentially identical, and transmission described in the present embodiment is cached and connect
Mouth 220 locks the calculated result of each neuron computing unit 210 and address effective marker when global clock gclk is triggered
It deposits, and resets the address effective marker of each neuron computing unit 210, hereafter pulse N_ is compared in the sending of nuclear control device 300
Data_comp reads serial number and Current neural after neuron computing module 200 receives the relatively pulse N_data_comp
The neuron calculated result is then sent into function letter if effective by the identical neuron address effective marker of first period N_peri
Breath table 230 simultaneously obtains transformation result, if invalid, then issues data invalid mark N_data_null to nuclear control device 300, drives
Kinetonucleus controller 300 enters next neuron period.The transmission caching and interface 220, which are worked as, judges that address effective marker is effective
When, the transformation result of function information table 230 is exported by SPI interface, while resetting the tired of corresponding neuron computing unit 210
Add device 214, when judge address effective marker for it is invalid when, accumulator 214 is not handled, is simultaneously emitted by data invalid mark
N_data_null, driving nuclear control device 300 enter next neuron period.
The artificial neural networks core 10 that third embodiment of the invention provides is set by using addressing is integrated with calculating
Thinking is counted, the connection number of single neuron computing unit also can be changed in neuron computing unit number in dynamically configurable core.
Fourth embodiment of the invention provides a kind of artificial neural networks core, including router-module, neuron calculate
Module, nuclear control device.Neuron calculated result is sent to institute for carrying out neuron calculating by the neuron computing module
State router-module.The neuron computing module can be existing various neuron computing modules, with this in the present embodiment
It is illustrated for the neuron computing module 200 that invention second embodiment provides.Specifically refer to Fig. 4, artificial neural network
Calculating core 20 includes router-module 100, neuron computing module 200, nuclear control device 300.
The present embodiment and 3rd embodiment difference is to increase router-module 100, now to the router-module 100 into
Row is described in detail.
The router-module 100 is for receiving and parsing through outer input data, and by the address information and axis after parsing
Prominent value information is sent into corresponding neuron computing module 200；And the neuron meter for exporting the neuron computing module 200
It calculates result and is sent to destination address, and send this frame data to nuclear control device 300 and handle complement mark.
The router-module 100 further comprises：Path control deivce 110, internuclear interaction caching and interface, routing iinformation
Table 160, neuron-router interface 170.Wherein the internuclear interaction caching and interface further comprise right four hairs in upper lower-left
Send reception caching and interface, i.e., it is upper to transmit and receive caching and interface 120, left transmitting and receiving caching and interface 130, lower transmitting and receiving
Caching and interface 140, right transmitting and receiving caching and interface 150.The number of caching and its data-interface is transmitted and received in practical application
Mesh can do variation appropriate according to specific requirements.The router-module 100 is interacted with external data using serial peripheral
Interface (Serial Peripheral Interface, SPI).The right four transmitting and receivings caching in upper lower-left is internuclear interaction caching
Area.Neuron-the router interface 170 is used to receive the neuron calculated result of the output of neuron computing module 200, this reality
It applies in example using SPI interface, is only capable of one frame calculated result of storage.
When the router-module 100 receives data, external interface（Internuclear interaction caching including neighboring router module
And interface and local neuron interface）Caching non-empty mark is sent to the router-module 100, prompts local router module
100 carry out data receiver.External interface works in slave pattern at this time, and local interface works in holotype.If local router mould
It is less than to receive caching for corresponding direction in block 100, then starts a frame data to receive according to first in, first out rule, and to path control deivce
110 issue reception caching non-empty mark, and data frame in 110 pairs of receptions cachings of path control deivce is prompted to handle；If local road
Receiving caching by corresponding direction in device module 100 has expired, then without data receiver, outside sends the pending datas such as caching and sends.
When the router-module 100 sends data, path control deivce 110, which is first read, sends buffer status.If sending slow
Deposit less than, then data frame to be sent be written according to first in, first out rule and send caching, at the same send out to external interface send it is slow
Non-empty mark is deposited, external receive is waited；If sending caching has expired, path control deivce 110 skips the transmission of this frame data, under reading
One frame data simultaneously parse.
The path control deivce 110 is the core for routing data parsing and transmission control.It is cached according to upper transmitting and receiving
And interface 120, left transmitting and receiving caching and interface 130, lower transmitting and receiving caching and interface 140, right transmitting and receiving are cached and are connect
Mouth 150, the sequence of neuron-router interface 170 are successively read whether each reception caching is empty.If non-empty, then one is read
Frame data are parsed；If it is empty, then it skips to next direction and receives caching.
Fig. 5 is referred to, which is the internuclear transmitting data frame format of artificial neural networks core, i.e., in internuclear interaction caching
One frame data format of storage, the data frame include：Neuron computing module address field, axis in target core address field, core
Prominent address field, neuron output field.Specifically in the present embodiment, the target core address field further comprises：Left and right
Direction target core address field, up and down direction target core address field.Wherein, left and right directions target core address field, upper and lower
It is respectively 8 bit signed numbers to target core address field, highest order is 0 to represent to the left or upwards, and highest order is 1 to represent to the right
Or downwards；Neuron computing module address field is 4 bit unsigned numbers in core, represents neuron computing module in some core；
Aixs cylinder address field is 8 bit unsigned numbers, carries out address matching for neuron computing unit in neuron computing module；Mind
It is divided into 8 bit unsigned numbers through first output field.The data frame format that the neuron-router interface 170 stores only includes
Neuron output field in Fig. 5 data frame.Single artificial neural networks core issues the network position that data frame can reach
It sets, neuron computing module number and single neuron computing unit can couple in core the aixs cylinder address number upper limit is by the core
Between transmitting data frame determine.It should be noted that the present embodiment has been merely given as a kind of specific form of data frames, practical application
The sequence of each field and bit number can make the appropriate adjustments in middle data frame.
The path control deivce 110 carries out data parsing after reading one frame data of internuclear interaction caching.If being sent to this
The data frame on ground, i.e. the left and right directions target core address, up and down direction target core address of data frame are 0, then remove data frame
In target core address field, while neuron computing module address and aixs cylinder address and neuron are exported into two words in analytic kernel
Section is sent into target nerve member computing module, is then removed in reception caching and is resolved data frame；If not being sent to local
Data frame, then according to first left and right, it is rear above and below sequence by target core address decrement, then detect target send caching whether be it is full,
If not dealing with while skipping to the next direction reception caching of reading then completely, if not full, then by treated, data frame feeding is sent out
It send caching medium to be sent, while removing in reception caching and being resolved data frame.
After the path control deivce 110 reads reception one frame data of caching of neuron-router interface 170, with nuclear control
The Current neural member period N_peri that device 300 exports makees the input of route information table 160, extracts and stores in route information table 160
Different neuronal target addresses, and by the neuron of address and reading export carry out framing, while detect target send caching
Whether be it is full, if it is full do not deal with then while skipping to read next direction and receive caching, if not full then will treated counts
It is medium to be sent that transmission caching is sent into according to frame, while being removed in the reception caching of neuron-router interface 170 and being resolved data frame
And send this frame data to nuclear control device 300 and handle complement mark R_N_data_done, driving nuclear control device 300 enters next
The neuron period.Neuron-the router interface 170 issues data effective marker N_data_ in neuron computing module 200
After en, start if reception caching is less than and receive a frame data, keeps reception to wait if receiving caching and having expired.Neuron calculates
Module 200 empties data effective marker N_data_en after being successfully transmitted a frame data to neuron-router interface 170.
What is stored in the route information table 160 is the destination address of different neurons, is inputted as the Current neural member period,
Output is corresponding destination address.
20 operational process of artificial neural networks core includes router data transmitting-receiving and neuron computing module meter
It calculates result and sends two parts, refer to Fig. 6, the router data transmitting-receiving comprises the steps of：
S101, the Neural Network Data packet to be received such as router-module 100；
S102, router-module 100 receive Neural Network Data packet and parse；
S103, router-module 100 judge whether the Neural Network Data packet received is sent to local, if so then execute
S105, if otherwise executing S104；
S104 will be sent into transmission by the sequence above and below behind first left and right after target core address decrement in the Neural Network Data packet
Caching returns to S101；
S105, by the Neural Network Data packet aixs cylinder address and aixs cylinder value be sent into neuron computing module 200, the mind
Each neuron computing unit is passed to through network packet；
S106, each neuron computing unit judge whether the aixs cylinder address matches with itself storage address, if then
S107 is executed, if otherwise returning to S101；
S107, neuron computing unit exports weighted value corresponding with the aixs cylinder address, by the weighted value and the axis
Prominent value is sent into accumulator after being multiplied, and returns to S101.
Fig. 7 is referred to, 200 calculated result of the neuron computing module transmission comprises the steps of：
S201 waits global clock triggering to come into force；
S202, judges whether global clock triggering comes into force, if so then execute S203, if otherwise returning to S201；
Each neuron computing module accumulator result is latching to corresponding transmission caching and interface by S203；
S204, judges whether each transmission caching and interface receive the comparison pulse that corresponding nuclear control device issues, if
S205 is then executed, if otherwise returning to S204；
S205, reading address value neuron computing module identical with the Current neural member period is sent in caching and interface
In data；
S206 judges whether the data are effective, if so then execute S207, if otherwise executing S215；
S207 by the data input function look-up table and exports result；
S208 judges whether the output result is effective, if so then execute S209, if otherwise executing S215；
The output result is sent to caching and router is waited to receive by S209, resets corresponding neuron accumulator；
S210 waits router idle, S211 is executed if the router free time, if router is not idle to return to S210；
S211, router-module start a data cached reception of neuron computing module；
S212 judges that neuron computing module issues whether information handles completion, if so then execute S213, if otherwise returning
S212；
S213, router-module complete pulse to the processing of nuclear control device photos and sending messages；
S214, nuclear control device judges whether the last one neuron period, if then returning to S201, if otherwise executing
S216；
S215, to nuclear control device photos and sending messages data invalid pulse；
S216, nuclear control device drive the neuron period to substitute, and pulse is compared in sending, return to S204.
The artificial neural networks core 20 that fourth embodiment of the invention provides calculates mould using single-router single neuron
The working method of block realizes the networking of multiple artificial neural networks cores 20 by router-module 100.
Fifth embodiment of the invention provides a kind of artificial neural networks core, including a router-module, multiple nerves
First computing module, multiple nuclear control devices, the multiple neuron computing module and the multiple nuclear control device correspond.This reality
It applies example to be with 3rd embodiment difference, 3rd embodiment is configuration of the single channel by single neuron computing module, and the present embodiment
It is single channel by the configuration of multi-neuron computing module, i.e., single router-module can be established with multiple neuron computing modules to be joined
It connects.
The present embodiment is specifically illustrated so that single channel is by four neuron computing modules as an example, calculates mould through member in practical application
The quantity of block can do variation appropriate according to specific requirements.Refer to Fig. 8, a kind of artificial neural networks core 30, packet
Include router-module 400, peripheral sensory neuron computing module 200a, the first nuclear control device 300a, nervus opticus member computing module
200b, the second nuclear control device 300b, third nerve member computing module 200c, third nuclear control device 300c, fourth nerve member calculate
Module 200d, the 4th nuclear control device 300d.
The router-module 400 is distinguished with the router-module 100 in 3rd embodiment to be, router-module 400
It include multiple neuron-router interfaces, i.e. peripheral sensory neuron-router interface 471, nervus opticus member-router interface
472, third nerve member-router interface 473, fourth nerve member-router interface 474.Every group of nuclear control device, neuron calculate
Module, neuron-router interface respectively with path control deivce 410 by single channel in embodiment three by single neuron computing module
Configuration mode establishes connection.
A router-module corresponds to multiple minds in the artificial neural networks core 30 that fourth embodiment of the invention provides
Through first computing module, if multiple neuron computing module has the function of same or similar, cluster calculation may be implemented, can drop
Low router number of nodes improves transmission bandwidth.The flexible expansion and the effective use of storage resource for calculating core are realized, with enhancing
The adaptability for calculating core, avoids additional resource overhead.
Fig. 9 is referred to, the present invention further provides a kind of artificial neural networks systems, including multiple artificial neural network
Network calculates core 40, and multiple artificial neural networks core 40 realizes being coupled to each other for multiple directions by router-module 100.
Only gived in embodiment include 9 artificial neural networks core 40 situation, and each artificial neural networks core
40 include 4 transmitting and receivings caching and interface, realizes being coupled to each other for 4 directions up and down.It is appreciated that practical application
The transmitting and receiving caching and interface for including in the number of middle artificial neural networks core and each artificial neural networks core
Number can change according to actual application scenarios.
Present embodiments provide a kind of artificial neural network by forming multiple 40 networkings of artificial neural networks core
Network computing system realizes the network connection of more artificial neural networks cores 40.
In addition, those skilled in the art can also do other variations in spirit of that invention, certainly, these are according to the present invention
The variation that spirit is done, all should be comprising within scope of the present invention.
Claims (11)
1. a kind of artificial neural networks system, which is characterized in that including：One router-module, an at least neuron calculate
Module, at least a nuclear control device, the neuron computing module and the nuclear control device correspond；
The router-module, for receiving and parsing through outer input data, and by after parsing address information and aixs cylinder value believe
Breath is sent into corresponding neuron computing module；
The neuron computing module is sent to the router for carrying out neuron calculating, and by neuron calculated result
Module；
The router-module, the neuron calculated result for being also used to export the neuron computing module is with being sent to target
Location, and this frame data processing complement mark is sent to the corresponding nuclear control device of the neuron computing module；
The nuclear control device, for entering next neuron period after receiving described frame data processing complement mark.
2. artificial neural networks system as described in claim 1, which is characterized in that the artificial neural networks core system
System includes a router-module, N number of neuron computing module, N number of nuclear control device, N number of neuron computing module and the N
A nuclear control device corresponds, and N is the integer greater than 1.
3. artificial neural networks system as described in claim 1, which is characterized in that
The nuclear control device is also used to issue the neuron period to itself corresponding neuron computing module and compares pulse；
The neuron computing module is also used to after receiving the comparison pulse that itself corresponding nuclear control device issues, reads
Address value neuron identical with Current neural member period address effective marker, if effectively then by the neuron calculated result
It is converted, if in vain then to itself corresponding nuclear control device sending data invalid mark；
The neuron computing module, be also used to judge conversion after neuron calculated result it is whether effective, if effectively then will
The neuron calculated result after changing is sent to the router-module, if then issuing in vain to itself corresponding nuclear control device
Data invalid mark；
The nuclear control device is also used to laggard in the data invalid mark for receiving itself corresponding neuron computing module transmission
Enter next neuron period.
4. artificial neural networks system as claimed in claim 3, which is characterized in that the neuron computing module includes
Multiple neuron computing units send caching and interface, function information table；
The multiple neuron computing unit receives and parses through address information and aixs cylinder value letter that the router-module is sent
Breath, if the address information received is matched with the address information that neuron computing unit itself stores, which calculates single
Member carries out neural computing to the information received, and calculated result and address effective marker are sent to described send and delayed
It deposits and interface, if the address information received and the address information of neuron computing unit itself storage mismatch, not to this
The information received carries out neural computing；
Transmission caching and interface, in the triggering of external global clock by the calculating of the multiple neuron computing unit
As a result and address effective marker is latched, and resets the address effective marker of the multiple neuron computing unit；And
After receiving the comparison pulse that itself corresponding nuclear control device issues, serial number neuron identical with the Current neural member period is read
Address effective marker, if effective, which is sent into the function information table by the transmission caching and interface
And transformation result is obtained, if invalid, the transmission caching and interface issue data invalid mark to itself corresponding nuclear control device
Will；
The neuron calculated result is converted to corresponding functional value by the function information table；
Transmission caching and interface, are also used to judge whether the transformation result effective, if effective, transmissions cache and
Interface exports the transformation result, and if invalid, the transmission caching and interface issue data to itself corresponding nuclear control device
Invalid flag.
5. artificial neural networks system as claimed in claim 4, which is characterized in that the router-module further wraps
It includes：Neuron-router interface, path control deivce, route information table, internuclear interaction caching and interface；
Neuron-the router interface, for receiving the neuron calculated result of the neuron computing module output；
The path control deivce, the neuron calculated result exported for reading the neuron-router interface, and will
The Current neural member period of nuclear control device output corresponding with the neuron computing module is sent to the route information table；
The route information table, for storing the destination address of different neurons, current mind that the path control deivce is inputted
It is corresponding destination address through first periodic conversion；
The path control deivce, the destination address for being also used to export the route information table and the neuron calculated result group
The internuclear interaction caching and interface are sent at the first data frame, and by first data frame；
The internuclear interaction caching and interface, for sending first data frame.
6. artificial neural networks system as claimed in claim 5, which is characterized in that first data frame includes：
Target core address field stores target artificial neural network computing system address；
Neuron computing module address field in core stores target nerve member in target artificial neural network computing system and calculates mould
Block address；
Aixs cylinder address field stores target nerve member computing unit address in target nerve member computing module；
Neuron output field stores neuron calculated result.
7. artificial neural networks system as claimed in claim 6, which is characterized in that
The internuclear interaction caching and interface, are also used to receive extrinsic neural network packet；
The path control deivce is also used to parse the extrinsic neural network packet, if the extrinsic neural network packet
To be sent to local data packet, then by the extrinsic neural network packet aixs cylinder address and aixs cylinder value be sent into target nerve
First computing module, if the extrinsic neural network packet is not to be sent to local data packet, to the extrinsic neural network
Target core address in data packet is updated, and the updated extrinsic neural network packet is sent into the internuclear interaction
Caching and interface.
8. artificial neural networks system as claimed in claim 7, which is characterized in that the neuron-router interface
Using Serial Peripheral Interface (SPI), slave pattern is worked in when the Serial Peripheral Interface (SPI) is as sender, works in master when as recipient
Mode.
9. artificial neural networks system as claimed in claim 8, which is characterized in that the internuclear interaction caching and interface
Using Serial Peripheral Interface (SPI), slave pattern is worked in when the Serial Peripheral Interface (SPI) is as sender, works in master when as recipient
Mode.
10. a kind of data receiver method of artificial neural networks system as claimed in claim 9, which is characterized in that packet
Include following steps：
S101, the Neural Network Data packet to be received such as described router-module；
S102, the router-module receive Neural Network Data packet and parse；
S103, whether the Neural Network Data packet that the router mould judgement receives is sent to local, if so then execute S105, if
Otherwise S104 is executed；
S104 will be sent into transmission caching by the sequence above and below behind first left and right after target core address decrement in the Neural Network Data packet,
Return to S101；
S105, by the Neural Network Data packet aixs cylinder address and aixs cylinder value be sent into neuron computing module, the neural network
Data packet is passed to each neuron computing unit；
S106, each neuron computing unit judge whether the aixs cylinder address matches with itself storage address, if so then execute
S107, if otherwise returning to S101；
S107, neuron computing unit exports weighted value corresponding with the aixs cylinder address, by the weighted value and the aixs cylinder value
It is sent into accumulator after multiplication, returns to S101.
11. a kind of data transmission method for uplink of artificial neural networks system as claimed in claim 9, which is characterized in that packet
Include following steps：
S201 waits global clock triggering to come into force；
S202, judges whether global clock triggering comes into force, if so then execute S203, if otherwise returning to S201；
Each neuron computing module accumulator result is latching to corresponding transmission caching and interface by S203；
S204, judges whether each transmission caching and interface receive the comparison pulse that corresponding nuclear control device issues, if then holding
Row S205, if otherwise returning to S204；
S205 reads address value neuron computing module identical with the Current neural member period and sends in caching and in interface
Data；
S206 judges whether the data are effective, if so then execute S207, if otherwise executing S215；
S207 by the data input function look-up table and exports result；
S208 judges whether the output result is effective, if so then execute S209, if otherwise executing S215；
The output result is sent to caching and router is waited to receive by S209, resets corresponding neuron accumulator；
S210 waits router idle, S211 is executed if the router free time, if router is not idle to return to S210；
S211, router-module start a data cached reception of neuron computing module；
S212 judges that neuron computing module issues whether information handles completion, if so then execute S213, if otherwise returning
S212；
S213, router-module complete pulse to the processing of nuclear control device photos and sending messages；
S214, nuclear control device judges whether the last one neuron period, if then returning to S201, if otherwise executing S216；
S215, to nuclear control device photos and sending messages data invalid pulse；
S216, nuclear control device drive the neuron period to substitute, and pulse is compared in sending, return to S204.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610354206.3A CN106056212B (en) | 2016-05-25 | 2016-05-25 | A kind of artificial neural networks core |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610354206.3A CN106056212B (en) | 2016-05-25 | 2016-05-25 | A kind of artificial neural networks core |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106056212A CN106056212A (en) | 2016-10-26 |
| CN106056212B true CN106056212B (en) | 2018-11-23 |
Family
ID=57175911
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610354206.3A Active CN106056212B (en) | 2016-05-25 | 2016-05-25 | A kind of artificial neural networks core |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106056212B (en) |
Families Citing this family (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108256641A (en) * | 2016-12-28 | 2018-07-06 | 上海磁宇信息科技有限公司 | For the cellular array internal network communication method of cellular array computing system |
| WO2018137412A1 (en) * | 2017-01-25 | 2018-08-02 | 清华大学 | Neural network information reception method, sending method, system, apparatus and readable storage medium |
| CN106971228B (en) * | 2017-02-17 | 2020-04-07 | 北京灵汐科技有限公司 | Method and system for sending neuron information |
| CN106971229B (en) * | 2017-02-17 | 2020-04-21 | 清华大学 | Neural network computing core information processing method and system |
| CN110413551B (en) | 2018-04-28 | 2021-12-10 | 上海寒武纪信息科技有限公司 | Information processing apparatus, method and device |
| WO2019001418A1 (en) | 2017-06-26 | 2019-01-03 | 上海寒武纪信息科技有限公司 | Data sharing system and data sharing method therefor |
| CN109426553A (en) | 2017-08-21 | 2019-03-05 | 上海寒武纪信息科技有限公司 | Task cutting device and method, Task Processing Unit and method, multi-core processor |
| CN109214616B (en) * | 2017-06-29 | 2023-04-07 | 上海寒武纪信息科技有限公司 | Information processing device, system and method |
| CN109214502B (en) | 2017-07-03 | 2021-02-26 | 清华大学 | Neural network weight discretization method and system |
| CN108334942B (en) * | 2017-12-22 | 2020-08-04 | 清华大学 | Data processing method, device, chip and storage medium of neural network |
| CN108345936B (en) * | 2018-01-31 | 2020-12-04 | 清华大学 | Neuromorphic chip, system and method based on internal state mark |
| CN110399977A (en) * | 2018-04-25 | 2019-11-01 | 华为技术有限公司 | Pond arithmetic unit |
| CN110188872B (en) * | 2019-06-05 | 2021-04-13 | 北京灵汐科技有限公司 | Heterogeneous cooperative system and communication method thereof |
| CN112163673B (en) * | 2020-09-28 | 2023-04-07 | 复旦大学 | Population routing method for large-scale brain-like computing network |
| CN112242963B (en) * | 2020-10-14 | 2022-06-24 | 广东工业大学 | Rapid high-concurrency neural pulse data packet distribution and transmission method and system |
| CN112650705A (en) * | 2020-12-31 | 2021-04-13 | 清华大学 | Routing control method and artificial intelligence processor |
| CN112817898A (en) * | 2021-02-08 | 2021-05-18 | 清华大学 | Data transmission method, processor, chip and electronic equipment |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP0664516A2 (en) * | 1994-01-19 | 1995-07-26 | Nippon Telegraph And Telephone Corporation | Neural network with reduced calculation amount |
| CN101681449A (en) * | 2007-06-15 | 2010-03-24 | 佳能株式会社 | Calculation processing apparatus and method |
-
2016
- 2016-05-25 CN CN201610354206.3A patent/CN106056212B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP0664516A2 (en) * | 1994-01-19 | 1995-07-26 | Nippon Telegraph And Telephone Corporation | Neural network with reduced calculation amount |
| US5630024A (en) * | 1994-01-19 | 1997-05-13 | Nippon Telegraph And Telephone Corporation | Method and apparatus for processing using neural network with reduced calculation amount |
| CN101681449A (en) * | 2007-06-15 | 2010-03-24 | 佳能株式会社 | Calculation processing apparatus and method |
Non-Patent Citations (1)
| Title |
|---|
| 数字神经元芯片的设计与应用;罗莉;《计算机研究与发展》;19980930;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106056212A (en) | 2016-10-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106056212B (en) | A kind of artificial neural networks core | |
| CN106056211B (en) | Neuron computing unit, neuron computing module and artificial neural networks core | |
| CN105989539B (en) | A kind of financial transaction market obtain system and acquisition methods | |
| CN107766148A (en) | A kind of isomeric group and task processing method and device | |
| CN106951926A (en) | The deep learning systems approach and device of a kind of mixed architecture | |
| CN107710237A (en) | Deep neural network divides on server | |
| CN106294233B (en) | A kind of transfer control method and device of direct memory access | |
| CN106815329A (en) | A kind of data cached update method and device | |
| CN107430704A (en) | Neural network algorithm is realized in nerve synapse substrate based on the metadata associated with neural network algorithm | |
| CN107688853A (en) | A kind of device and method for being used to perform neural network computing | |
| CN104461943B (en) | Method for reading data, device and system | |
| CN112468401B (en) | Network-on-chip routing communication method for brain-like processor and network-on-chip | |
| CN109246027B (en) | Network maintenance method and device and terminal equipment | |
| Goh et al. | Nonlocal evolution of weighted scale-free networks | |
| CN108334942A (en) | Data processing method, device, chip and the storage medium of neural network | |
| CN107241305A (en) | A kind of network protocol analysis system and its analysis method based on polycaryon processor | |
| CN110188872A (en) | A kind of isomery cooperative system and its communication means | |
| CN108365996B (en) | FPGA + ARM architecture-based on-chip network simulation platform | |
| CN105430060B (en) | Constraint publish-subscribe model space flight measurement and control data exchange system | |
| CN104717272A (en) | Event stream processing system and method thereof | |
| CN109327511A (en) | A kind of data request method and server based on http protocol | |
| CN108595379A (en) | A kind of parallelization convolution algorithm method and system based on multi-level buffer | |
| CN113312177A (en) | Wireless edge computing system and optimization method based on federal learning | |
| CN110086731A (en) | A kind of cloud framework lower network data stabilization acquisition method | |
| CN113642700A (en) | Cross-platform multi-modal public opinion analysis method based on federal learning and edge calculation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |