CN106056212A - Artificial neural network calculating core - Google Patents
Artificial neural network calculating core Download PDFInfo
- Publication number
- CN106056212A CN106056212A CN201610354206.3A CN201610354206A CN106056212A CN 106056212 A CN106056212 A CN 106056212A CN 201610354206 A CN201610354206 A CN 201610354206A CN 106056212 A CN106056212 A CN 106056212A
- 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.)
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Neurology (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
Abstract
The invention relates to an artificial neural network calculating core. The artificial neural network calculating core comprises a router module, at least one neuron calculation module and at least one core controller; and the neuron calculation modules correspond to the core controllers one by one. The router module receives and parses external input data, and sends parsed address information and axonal value information to one corresponding neuron calculation module; the router module sends a neuron calculation result output by the corresponding neuron calculation module to a target address, and sends a local-frame data processing completion flag to one core controller corresponding to the corresponding neuron calculation module. Each neuron calculation module is used in a neuron calculation and to send the neuron calculation result to the router module. Each core controller enters a next neuron period after receiving the local-frame data processing completion flag. Networking of the multiple artificial neural network calculating cores can be achieved by utilizing the artificial neural network calculating core provided in the invention.
Description
Technical field
The invention belongs to artificial neural networks field, particularly to a kind of neuron computing unit.
Background technology
Artificial neural network is to use for reference the computation model that biological brain synapse-neuronal structure develops, can be concurrently
Carry out large-scale complex computing, and there is the feature of 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, synapse unit
Composition.Neuron computing unit is most basic computing unit, can carry out simple mathematical operation；Aixs cylinder unit is responsible for output god
Through unit's result of calculation, a neuron has an aixs cylinder；Dendron unit is the input that neuron calculates, and a neuron can have many
Individual dendron；Synapse unit represents the weight that a neuron axon and next neuron dendron couple；Between neuron and neuron
By aixs cylinder, synapse, dendricity layering couple thus constitute neutral net.
Artificial neural network is formed, by a large amount of distributed simple operation unit by the connection of multilamellar neuron computing unit
Interaction realizes parallel complex nonlinear and calculates, and has powerful information processing capability.At research artificial neural networks
During model, it is low that conventional calculating equipment such as general purpose computer often occurs calculating usefulness, the longest, the shortcomings such as energy consumption is high.Closely
Nian Lai, part research institution develops general neural computing chip, such as IBM for the feature of artificial neural network
TrueNorth, has tentatively reached to carry out in chip the purpose of artificial neural network computing.But, most neutral net meters
Calculate chip and use mentality of designing computing unit (neuron) and connection unit (synapse matrix) separated, owing to have employed
The layout type that fixing scale couples entirely, thus limit the connection number of neuron in monokaryon, simultaneously at non-full connection nerve net
In the application of network, will also result in the waste not coupling synapse point position storage resource.
Summary of the invention
In view of this, a kind of neural computing chip that can realize flexible networking of necessary offer.
A kind of artificial neural networks core, including a: router-module, at least one neuron computing module, at least one
Nuclear control device, described neuron computing module and described nuclear control device one_to_one corresponding；Described router-module, is used for receiving and solving
Analysis outer input data, and address information and the aixs cylinder value information after resolving send into corresponding neuron computing module；Described
Neuron computing module, is used for carrying out neuron calculating, and sends neuron result of calculation to described router-module；Described
Router-module, the neuron result of calculation being additionally operable to export described neuron computing module sends to destination address, and to
The nuclear control device that this neuron computing module is corresponding sends these frame data and processes complement mark；Described nuclear control device, for connecing
Receive after described frame data process complement mark and enter next neuron cycle.
Compared with prior art, the artificial neural networks core utilizing the present invention to provide can realize multiple artificial god
Networking through network calculations core, it is achieved calculate the flexible of core and expand and effective utilization of storage resource, enhance and calculate core
Adaptability, it is to avoid extra resource overhead.
Accompanying drawing explanation
The neuron computing unit schematic diagram that Fig. 1 provides for first embodiment of the invention.
The neuron computing module schematic diagram that Fig. 2 provides for second embodiment of the invention.
The artificial neural networks core schematic diagram that Fig. 3 provides for third embodiment of the invention.
The artificial neural networks core schematic diagram that Fig. 4 provides for fourth embodiment of the invention.
The artificial neural networks core internuclear transmitting data frame form that Fig. 5 provides for fourth embodiment of the invention.
The artificial neural networks core router data transmitting-receiving flow chart that Fig. 6 provides for the present invention.
The artificial neural networks nucleus neuron computing module result of calculation transmission flow figure that Fig. 7 provides for the present invention.
The artificial neural networks core schematic diagram that Fig. 8 provides for fifth embodiment of the invention.
The artificial neural networks core networking schematic diagram that Fig. 9 provides for the present invention.
Main element symbol description
| Artificial neural networks core | 10、20、30、40 |
| Router-module | 100、400 |
| Path control deivce | 110、410 |
| Upper transmission receives caching and interface | 120 |
| Left transmission receives caching and interface | 130 |
| Lower transmission receives caching and interface | 140 |
| Right transmission receives caching and interface | 150 |
| Route information table | 160 |
| Neuron-router interface | 170 |
| Protoneuron-router interface | 471 |
| Nervus opticus unit-router interface | 472 |
| Third nerve unit-router interface | 473 |
| Fourth nerve unit-router interface | 474 |
| Neuron computing module | 200 |
| Protoneuron computing module | 200a |
| Nervus opticus unit computing module | 200b |
| Third nerve unit computing module | 200c |
| Fourth nerve unit 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 |
| 3rd nuclear control device | 300c |
| 4th nuclear control device | 300d |
Following detailed description of the invention will further illustrate the present invention in conjunction with above-mentioned accompanying drawing.
Detailed description of the invention
Below in conjunction with the accompanying drawings and the specific embodiments, the artificial neural networks core provided the present invention is made further
Describe in detail.
Referring to Fig. 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.
Described decoder module 211 is used for receiving neutral net information and resolving.Described neutral net information includes that address is believed
Breath and aixs cylinder value information, address information therein is sent into described address weight module 212 by described decoder module 211, by aixs cylinder
Value information sends into described multiplier 213.
Described address weight module 212 stores address weight to list, it is achieved to being input to this address weight module 212
Address information mate.If input address is mated with the address information stored in address weight module 212, then address power
Molality block 212 exports corresponding weighted value to multiplier 213, and the weighted value of output is the numerical value of the interior change of certain limit；If
Input address is not mated with the address information stored, then address weight module 212 exports null value to multiplier 213.Additionally, institute
State address weight module 212 and address effective marker state can also be set according to matching judgment result.If specifically, input address
Mating with the address information stored in address weight module 212, described address weight module 212 sends address effective marker,
If input address is not mated with the address information stored in address weight module 212, described address weight module 212 does not changes
Address effective marker state.
Described multiplier 213 receives the weighted value of described address weight module 212 output and described decoder module 211 is defeated
The aixs cylinder value gone out, is multiplied described weighted value with aixs cylinder value, and product is sent into accumulator 214.
The result of calculation that described accumulator 214 realizes being exported by multiplier 213 carries out cumulative and exports.
It is integrated, by adopting with calculating that the neuron computing unit 210 that first embodiment of the invention provides achieves addressing
With addressing with calculate integrated design, breach layout type that fixing scale couples the entirely limit to neuron computing unit number
System, simultaneously because take the connecting mode of address coupling, it is to avoid the wasting of resources that useless connection causes, thus realizes flexibility
Expand and effective utilization of storage resource, enhance neural computing efficiency.
Second embodiment of the invention further provides for a kind of neuron computing module, is used for setting up connecting relation between neuron
And carry out neuron calculating.Referring to Fig. 2, described neuron computing module 200 includes: multiple neuron computing units 210, send out
Send caching and interface 220, function information table 230.
The plurality of neuron computing unit 210 receives from outside address information and aixs cylinder value information, and to described
Address information judges, if the address information received is mated with the address information self stored, then and the letter this received
Breath carries out neural computing, and result of calculation and address effective marker is sent to described transmission caching and interface 220, if connecing
The address information that the address information received stores with self is not mated, and this information received is not carried out neutral net meter
Calculate.
Described in the present embodiment, in multiple neuron computing units, each neuron computing unit farther includes: decoding mould
Block 211, address weight module 212, multiplier 213, accumulator 214.Neuron computing unit in the present embodiment and the present invention
Neuron computing unit 210 structure that first embodiment provides is essentially identical with function, the mould of address weight described in the present embodiment
Block 212 is additionally operable to when input address is mated with the address information stored in address weight module 212, then address weight module
212 export corresponding weighted value to multiplier 213, and send address effective marker to transmission caching and interface 220；If input
Address is not mated with the address information stored, then address weight module 212 exports null value to multiplier 213, and unchanged
Location effective marker state.
Described caching and the interface 220 of sending is for calculating the plurality of neuron when outside global clock gclk triggers
Result of calculation and the address effective marker of unit 210 latch, and there is criterion the address resetting each neuron computing unit 210
Will, and the neuron result of calculation specified output to described function information table 230 and is obtained transformation result.
Described neuron result of calculation is converted to the functional value of correspondence by described function information table 230.
In the neuron computing module that second embodiment of the invention provides, single neuron computing unit will addressing and calculating
Unite two into one, neuron addressing can be carried out simultaneously and calculate, by using addressing and calculating integrated design, breach fixed gauge
The layout type that mould the couples full restriction to neuron computing unit number, simultaneously because take the connection side of address coupling
Formula, it is to avoid the wasting of resources that useless connection causes, thus realize flexible expansion and effective utilization of storage resource, enhance
Neural computing efficiency.
Third embodiment of the invention provides a kind of artificial neural networks core, including: neuron computing module, nuclear control
Device.Described neuron computing module is used for carrying out neuron calculating, and enters different neural under the control of described nuclear control device
The transmission flow of unit's result of calculation.Described neuron computing module can be existing various neuron computing module, this enforcement
Example illustrates as a example by the neuron computing module 200 that second embodiment of the invention provides.Specifically refer to Fig. 3, manually
Neural computing core 10 includes neuron computing module 200, nuclear control device 300.
Described nuclear control device 300 is by sending neuron cycle N_peri and comparing pulse N_data_comp control manually
Neural computing core 10 enters the transmission flow of different neuron result of calculation.When entering the new neuron cycle, pass through
Output is compared pulse N_data_comp and is triggered the comparison of neuron result of calculation, output, and waits data processed result mark R_
N_data_done or N_data_null is effectively to enter the new neuron cycle.Described nuclear control device 300 is at global clock
Gclk triggers the lower transmission cycle entering new round neuron result of calculation.
Described neuron computing module 200 is used for setting up between neuron connecting relation and carries out neuron calculating, with this
The neuron computing module structure that bright second embodiment provides is essentially identical with function, sends caching and connect described in the present embodiment
Result of calculation and the address effective marker of each neuron computing unit 210 are locked by mouth 220 when global clock gclk triggers
Depositing, and reset the address effective marker of each neuron computing unit 210, the most described nuclear control device 300 sends and compares pulse N_
Data_comp, neuron computing module 200 receives described after comparing pulse N_data_comp, reads sequence number and Current neural
The neuron address effective marker that unit cycle N_peri is identical, if effectively, then sends into function letter by this neuron result of calculation
Breath table 230 also obtains transformation result, if invalid, then send data invalid mark N_data_null to nuclear control device 300, drives
Kinetonucleus controller 300 enters next neuron cycle.Described caching and the interface 220 of sending is when judging that address effective marker is effective
Time, the transformation result of function information table 230 is exported by SPI interface, resets the tired of corresponding neuron computing unit 210 simultaneously
Add device 214, when judging that address effective marker is invalid, accumulator 214 is not processed, be simultaneously emitted by data invalid mark
N_data_null, drives nuclear control device 300 to enter next neuron cycle.
The artificial neural networks core 10 that third embodiment of the invention provides is by using addressing and calculating integrated setting
Meter thinking, can dynamically configure neuron computing unit number in core, it is possible to change the connection number of single neuron computing unit.
Fourth embodiment of the invention provides a kind of artificial neural networks core, calculates including router-module, neuron
Module, nuclear control device.Described neuron computing module is used for carrying out neuron calculating, and sends neuron result of calculation to institute
State router-module.Described neuron computing module can be existing various neuron computing module, with this in the present embodiment
Illustrate as a example by inventing the neuron computing module 200 that the second embodiment provides.Specifically refer to Fig. 4, artificial neural network
Calculate core 20 and include router-module 100, neuron computing module 200, nuclear control device 300.
The present embodiment and the difference of the 3rd embodiment are to add router-module 100, now enter this router-module 100
Row describes in detail.
Described router-module 100 is used for receiving and parsing through outer input data, and the address information after resolving and axle
Prominent value information sends into corresponding neuron computing module 200；And the neuron meter that described neuron computing module 200 exported
Calculate result to send to destination address, and send these frame data process complement mark to nuclear control device 300.
Described router-module 100 farther includes: path control deivce 110, internuclear mutual caching and interface, routing iinformation
Table 160, neuron-router interface 170.Wherein said internuclear mutual caching and interface farther include upper left bottom right four and send out
Send to receive and cache and interface, i.e. above send reception and cache and interface 120, left transmission reception caching and interface 130, lower transmission reception
Caching and interface 140, right transmission reception caching and interface 150.Actual application sends and receives caching and the number of data-interface thereof
Mesh can do suitable change according to real needs.Described router-module 100 uses serial peripheral alternately with external data
Interface (Serial Peripheral Interface, SPI).Upper left bottom right four sends and receives caching is internuclear mutual caching
District.Described neuron-router interface 170 is for receiving the neuron result of calculation of neuron computing module 200 output, this reality
Execute employing SPI interface in example, be only capable of storing a frame result of calculation.
When described router-module 100 receives data, external interface (includes the internuclear mutual caching of neighboring router module
And interface and local neuron interface) send caching non-NULL mark to this router-module 100, point out local router module
100 carry out data receiver.Now external interface works in from pattern, and local interface works in holotype.If local router mould
In block 100 correspondence direction receive caching less than, then start frame data to receive, and to path control deivce according to FIFO rule
110 send reception caching non-NULL mark, and prompting path control deivce 110 processes receiving Frame in caching；If local road
Being received caching by correspondence direction in device module 100 the fullest, the most do not carry out data receiver, the outside pending datas such as caching that send send.
When described router-module 100 sends data, path control deivce 110 first reads transmission buffer status.If sending slow
Deposit less than, then Frame to be sent is sent caching according to the write of FIFO rule, simultaneously sends transmission to external interface slow
Deposit non-NULL mark, wait external reception；If it is the fullest to send caching, then path control deivce 110 skips the transmission of these frame data, under reading
One frame data also resolve.
Described path control deivce 110 is the core that route data resolves and transmission controls.It receives caching according to upper transmission
And interface 120, left transmission receives caching and interface 130, and lower transmission receives caching and interface 140, and right transmission receives caching and connect
Mouth 150, the order of neuron-router interface 170 is successively read whether each reception caching is empty.If non-NULL, then read one
Frame data resolve；If it is empty, then skip to next direction and receive caching.
Referring to Fig. 5, this figure is artificial neural networks core internuclear transmitting data frame form, in the most internuclear mutual caching
One frame data format of storage, this Frame includes: neuron computing module address field, axle in target core address field, core
Prominent address field, neuron output field.The most in the present embodiment, described target core address field farther includes: left and right
Direction target core address field, above-below direction target core address field.Wherein, left and right directions target core address field, upper and lower
Be respectively 8 bit signed numbers to target core address field, highest order be 0 representative to the left or upwards, highest order be 1 representative to the right
Or downwards；In core, neuron computing module address field is 4 bit unsigned numbers, represents neuron computing module in certain core；
Aixs cylinder address field is 8 bit unsigned numbers, and in neuron computing module, neuron computing unit carries out address coupling；God
Through unit, output field is divided into 8 bit unsigned numbers.The data frame format of described neuron-router interface 170 storage only includes
Neuron output field in Fig. 5 Frame.Single artificial neural networks core sends the network position that Frame can arrive
Put, neuron computing module number and single neuron computing unit can couple in core the aixs cylinder address number upper limit is by this core
Between transmitting data frame determine.It should be noted that the present embodiment has been merely given as a kind of concrete form of data frames, actual application
In middle Frame, order and the bit number of each field can make the appropriate adjustments.
Described path control deivce 110 reads the laggard row data parsing of internuclear mutual caching one frame data.If sending to this
The Frame on ground, i.e. the left and right directions target core address of Frame, above-below direction target core address is 0, then removes Frame
In target core address field, neuron computing module address aixs cylinder address and neuron are exported two words in analytic kernel simultaneously
Section is sent in target nerve unit computing module, then removes and receives resolved Frame in caching；If not sending to this locality
Frame, then according to first left and right, the most upper and lower order is by target core address decrement, and then detection target sends whether caching is full,
If full then do not deal with and skip to read next direction simultaneously and receive caching, if not full the most then will process after Frame send into and send out
Send caching medium to be sent, concurrently disinfect resolved Frame in reception caching.
After described path control deivce 110 reads neuron-router interface 170 reception caching one frame data, with nuclear control
The Current neural unit cycle N_peri of device 300 output makees the input of route information table 160, extracts in route information table 160 and stores
Different neuronal target addresses, and the neuron output of address and reading is carried out framing, detection target sends caching simultaneously
Whether it is full, if full the most not dealing with skips to read next direction reception caching simultaneously, if not the number after the most then processing
Send into transmission caching according to frame medium to be sent, concurrently disinfect neuron-router interface 170 and receive resolved Frame in caching
And send these frame data process complement mark R_N_data_done to nuclear control device 300, drive nuclear control device 300 to enter next
The neuron cycle.Described neuron-router interface 170 sends data effective marker N_data_ at neuron computing module 200
After en, if receive caching less than; start reception one frame data, if it is full to receive caching, keeps receiving and wait.Neuron calculates
Module 200 empties data effective marker N_data_en after neurad unit-router interface 170 is successfully transmitted frame data.
In described route information table 160, storage is the destination address of different neuron, inputs as the Current neural unit cycle,
It is output as the destination address of correspondence.
Described artificial neural networks core 20 operational process includes router data transmitting-receiving and neuron computing module meter
Calculating result and send two parts, refer to Fig. 6, the transmitting-receiving of described router data comprises the steps of
S101, the Neural Network Data bag to be received such as router-module 100；
S102, router-module 100 receives Neural Network Data bag and resolves；
S103, router-module 100 judges whether the Neural Network Data bag received mails to this locality, if then performing S105,
If otherwise performing S104；
S104, order upper and lower after pressing about first feeding will send caching after target core address decrement in this Neural Network Data bag,
Return S101；
S105, sends the aixs cylinder address in this Neural Network Data bag and aixs cylinder value into neuron computing module 200, this nerve net
Network packet is passed to each neuron computing unit；
S106, each neuron computing unit judges whether described aixs cylinder address mates, if then performing with self storage address
S107, if otherwise returning S101；
S107, the weighted value that the output of neuron computing unit is corresponding with described aixs cylinder address, by this weighted value and described aixs cylinder value
Send into accumulator after being multiplied, return S101.
Referring to Fig. 7, described neuron computing module 200 result of calculation sends and comprises the steps of
S201, waits that global clock triggers and comes into force；
S202, it is judged that whether global clock triggering comes into force, if then performing S203, if otherwise returning S201；
S203, is latching to transmission caching and the interface of correspondence by each neuron computing module accumulator result；
S204, it is judged that whether each transmission caching and interface receive the comparison pulse that the nuclear control device of correspondence sends, if then holding
Row S205, if otherwise returning S204；
S205, reads the address value neuron computing module identical with the Current neural unit cycle and sends in caching and in interface
Data；
S206, it is judged that these data are the most effective, if then performing S207, if otherwise performing S215；
S207, by this data input function look-up table and export result；
S208, it is judged that this output result is the most effective, if then performing S209, if otherwise performing S215；
S209, delivers to this output result caching and waits that router receives, reset corresponding neuron accumulator；
S210, waits that router is idle, if the router free time, performs S211, if router is the most idle, return S210；
S211, router-module starts a data cached reception of neuron computing module；
S212, it is judged that neuron computing module sends whether information has processed, if then performing S213, if otherwise returning
S212；
S213, router-module has processed pulse to nuclear control device photos and sending messages；
S214, nuclear control device judges whether last neuron cycle, if then returning S201, if otherwise performing S216；
S215, to nuclear control device photos and sending messages data invalid pulse；
S216, nuclear control device drives the neuron cycle to substitute, sends and compare pulse, returns S204.
The artificial neural networks core 20 that fourth embodiment of the invention provides uses single-router mononeuron to calculate mould
The working method of block, realizes the networking of multiple artificial neural networks core 20 by router-module 100.
Fifth embodiment of the invention provides a kind of artificial neural networks core, including a router-module, multiple nerve
Unit computing module, multiple nuclear control device, the plurality of neuron computing module and the plurality of nuclear control device one_to_one corresponding.This reality
Executing example to be with the 3rd embodiment difference, the 3rd embodiment is the single channel configuration by mononeuron computing module, and the present embodiment
Connection can be set up with multiple neuron computing modules by the configuration of multi-neuron computing module, the most single router-module for single channel
Connect.
The present embodiment specifically illustrates by as a example by four neuron computing modules by single channel, calculates mould through unit in actual application
The quantity of block can do suitable change according to real needs.Refer to Fig. 8, a kind of artificial neural networks core 30, bag
Include router-module 400, protoneuron computing module 200a, the first nuclear control device 300a, nervus opticus unit computing module
200b, the second nuclear control device 300b, third nerve unit computing module 200c, the 3rd nuclear control device 300c, fourth nerve unit calculate
Module 200d, the 4th nuclear control device 300d.
Described router-module 400 is distinguished with the router-module 100 in the 3rd embodiment and is, router-module 400
Include multiple neuron-router interface, i.e. protoneuron-router interface 471, nervus opticus unit-router interface
472, third nerve unit-router interface 473, fourth nerve unit-router interface 474.Often group nuclear control device, neuron calculate
Module, neuron-router interface respectively with path control deivce 410 by single channel in embodiment three by mononeuron computing module
Configuration mode is set up and is coupled.
The corresponding multiple god of router-module in the artificial neural networks core 30 that fourth embodiment of the invention provides
Through unit's computing module, if the plurality of neuron computing module has same or like function, it is possible to achieve cluster calculation, can drop
Low router nodes, raising transmission bandwidth.Realize calculating the flexible of core to expand and effective utilization of storage resource, to strengthen
Calculate the adaptability of core, it is to avoid extra resource overhead.
Refer to Fig. 9, the present invention further provides a kind of artificial neural networks system, including multiple ANN
Network calculates core 40, and the plurality of artificial neural networks core 40 realizes being coupled to each other of multiple directions by router-module 100.
Embodiment only gives the situation including 9 artificial neural networks cores 40, and each artificial neural networks core
40 include 4 sends and receives caching and interface, it is achieved being coupled to each other of 4 directions up and down.It is appreciated that actual application
The transmission comprised in the number of middle artificial neural networks core and each artificial neural networks core receives caching and interface
Number can change according to actual application scenarios.
Present embodiments provide a kind of ANN by multiple artificial neural networks core 40 networkings being formed
Network calculates system, it is achieved that the network of many artificial neural networks core 40 connects.
It addition, those skilled in the art can also do other change in spirit of the present invention, certainly, these are according to the present invention
The change that spirit is done, within all should being included in scope of the present invention.
Claims (11)
1. an artificial neural networks core, it is characterised in that including a: router-module, at least one neuron calculate mould
Block, at least one nuclear control device, described neuron computing module and described nuclear control device one_to_one corresponding；
Described router-module, is used for receiving and parsing through outer input data, and the address information after resolving and aixs cylinder value letter
Breath sends into corresponding neuron computing module；
Described neuron computing module, is used for carrying out neuron calculating, and sends neuron result of calculation to described router
Module；
Described router-module, the neuron result of calculation being additionally operable to be exported by described neuron computing module sends to target ground
Location, and send these frame data process complement mark to the nuclear control device that this neuron computing module is corresponding；
Described nuclear control device, for entering next neuron cycle after receiving described frame data process complement mark.
2. artificial neural networks core as claimed in claim 1, it is characterised in that this artificial neural networks core includes
One router-module, N number of neuron computing module, N number of nuclear control device, described N number of neuron computing module and described N number of core control
Device one_to_one corresponding processed, N is the integer more than 1.
3. artificial neural networks core as claimed in claim 1, it is characterised in that
Described nuclear control device, is additionally operable to the neuron computing module to self correspondence and sends the neuron cycle and compare pulse；
Described neuron computing module, is additionally operable to after the comparison pulse that the nuclear control device receiving self correspondence sends, and reads
The neuron address effective marker that address value is identical with the Current neural unit cycle, if effectively then by this neuron result of calculation
Change, if invalid, send data invalid mark to the nuclear control device of self correspondence；
Described neuron computing module, is additionally operable to the neuron result of calculation after judging conversion the most effective, if the most then will
Neuron result of calculation after this changes sends to described router-module, if invalid, sends to the nuclear control device of self correspondence
Data invalid mark；
Described nuclear control device, the data invalid mark being additionally operable to send at the neuron computing module receiving self correspondence is laggard
Enter next neuron cycle.
4. artificial neural networks core as claimed in claim 3, it is characterised in that described neuron computing module includes many
Individual neuron computing unit, transmission caching and interface, function information table；
The plurality of neuron computing unit, receives and parses through address information and aixs cylinder value letter that described router-module sends
Breath, if the address information of the address information received and neuron computing unit self storage is mated, then this neuron calculates single
Unit carries out neural computing to this information received, and result of calculation and the transmission of address effective marker is delayed to described transmission
Deposit and interface, if the address information of the address information received and neuron computing unit self storage is not mated, the most not to this
The information received carries out neural computing；
Described transmission caches and interface, is used for the calculating of the plurality of neuron computing unit when outside global clock triggers
Result and address effective marker latch, and reset the address effective marker of the plurality of neuron computing unit；And
After receiving the comparison pulse that the nuclear control device of self correspondence sends, read the neuron that sequence number is identical with the Current neural unit cycle
Address effective marker, if effectively, this neuron result of calculation is sent into described function information table by described caching and the interface of sending
And obtain transformation result, if invalid, described transmission caching and interface send data invalid mark to the nuclear control device of self correspondence
Will；
Described function information table, is converted to the functional value of correspondence by described neuron result of calculation；
Described send caching and interface, be additionally operable to judge that described transformation result is the most effective, if effectively, described send caching and
This transformation result is exported by interface, if invalid, described transmission caching and interface send data to the nuclear control device of self correspondence
Invalid flag.
5. artificial neural networks core as claimed in claim 4, it is characterised in that described router-module wraps further
Include: neuron-router interface, path control deivce, route information table, internuclear mutual caching and interface；
Described neuron-router interface, for receiving the neuron result of calculation of described neuron computing module output；
Described path control deivce, for read the output of described neuron-router interface neuron result of calculation, and will
The Current neural unit cycle of the nuclear control device output corresponding with this neuron computing module sends to described route information table；
Described route information table, for storing the destination address of different neuron, the current god that described path control deivce is inputted
It is corresponding destination address through unit's periodic conversion；
Described path control deivce, is additionally operable to the destination address that exported by described route information table and described neuron result of calculation group
Become the first Frame, and this first Frame is sent into described internuclear mutual caching and interface；
Described internuclear mutual caching and interface, be used for sending described first Frame.
6. artificial neural networks core as claimed in claim 5, it is characterised in that described first Frame includes:
Target core address field, stores target artificial neural networks core address；
Neuron computing module address field in core, stores target artificial neural networks core internal object neuron computing module
Address；
Aixs cylinder address field, stores target nerve unit computing module internal object neuron computing unit address；
Neuron output field, stores neuron result of calculation.
7. artificial neural networks core as claimed in claim 6, it is characterised in that
Described internuclear mutual caching and interface, be additionally operable to receive extrinsic neural network packet；
Described path control deivce, is additionally operable to resolve described extrinsic neural network packet, if described extrinsic neural network packet
For sending to local packet, then target nerve is sent in the aixs cylinder address in this extrinsic neural network packet and aixs cylinder value
Unit's computing module, if described extrinsic neural network packet is not for sending to local packet, then to this extrinsic neural network
Target core address in packet is updated, and by after this renewal extrinsic neural network packet send into described internuclear alternately
Caching and interface.
8. the artificial neural networks core as described in claim 7 any claim, it is characterised in that described neuron-
Router interface uses Serial Peripheral Interface (SPI), works in from pattern, as recipient when this Serial Peripheral Interface (SPI) is as sender
Time work in holotype.
9. the artificial neural networks core as described in claim 8 any claim, it is characterised in that described internuclear alternately
Caching and interface use Serial Peripheral Interface (SPI), work in from pattern, as recipient when this Serial Peripheral Interface (SPI) is as sender
Time work in holotype.
10. the data receiver method of an artificial neural networks core as claimed in claim 9, it is characterised in that include
Following steps:
S101, the Neural Network Data bag to be received such as described router-module；
S102, described router-module receives Neural Network Data bag and resolves；
S103, described router mould judges whether the Neural Network Data bag received mails to this locality, if then performing S105, if
Otherwise perform S104；
S104, order upper and lower after pressing about first feeding will send caching after target core address decrement in this Neural Network Data bag,
Return S101；
S105, sends the aixs cylinder address in this Neural Network Data bag and aixs cylinder value into neuron computing module, this neutral net
Packet is passed to each neuron computing unit；
S106, each neuron computing unit judges whether described aixs cylinder address mates, if then performing with self storage address
S107, if otherwise returning S101；
S107, the weighted value that the output of neuron computing unit is corresponding with described aixs cylinder address, by this weighted value and described aixs cylinder value
Send into accumulator after being multiplied, return S101.
The data transmission method for uplink of 11. 1 kinds of artificial neural networks cores as claimed in claim 9, it is characterised in that include
Following steps:
S201, waits that global clock triggers and comes into force；
S202, it is judged that whether global clock triggering comes into force, if then performing S203, if otherwise returning S201；
S203, is latching to transmission caching and the interface of correspondence by each neuron computing module accumulator result；
S204, it is judged that whether each transmission caching and interface receive the comparison pulse that the nuclear control device of correspondence sends, if then holding
Row S205, if otherwise returning S204；
S205, reads the address value neuron computing module identical with the Current neural unit cycle and sends in caching and in interface
Data；
S206, it is judged that these data are the most effective, if then performing S207, if otherwise performing S215；
S207, by this data input function look-up table and export result；
S208, it is judged that this output result is the most effective, if then performing S209, if otherwise performing S215；
S209, delivers to this output result caching and waits that router receives, reset corresponding neuron accumulator；
S210, waits that router is idle, if the router free time, performs S211, if router is the most idle, return S210；
S211, router-module starts a data cached reception of neuron computing module；
S212, it is judged that neuron computing module sends whether information has processed, if then performing S213, if otherwise returning
S212；
S213, router-module has processed pulse to nuclear control device photos and sending messages；
S214, nuclear control device judges whether last neuron cycle, if then returning S201, if otherwise performing S216；
S215, to nuclear control device photos and sending messages data invalid pulse；
S216, nuclear control device drives the neuron cycle to substitute, sends and compare pulse, returns 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 true CN106056212A (en) | 2016-10-26 |
| CN106056212B 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) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106971229A (en) * | 2017-02-17 | 2017-07-21 | 清华大学 | Neural computing nuclear information processing method and system |
| CN106971228A (en) * | 2017-02-17 | 2017-07-21 | 清华大学 | Neuronal messages sending method and system |
| 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 |
| CN108334942A (en) * | 2017-12-22 | 2018-07-27 | 清华大学 | Data processing method, device, chip and the storage medium of neural network |
| CN108345936A (en) * | 2018-01-31 | 2018-07-31 | 清华大学 | A kind of neuromorphic chip, system and method based on internal state mark |
| 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 |
| CN109214616A (en) * | 2017-06-29 | 2019-01-15 | 上海寒武纪信息科技有限公司 | A kind of information processing unit, system and method |
| CN110399977A (en) * | 2018-04-25 | 2019-11-01 | 华为技术有限公司 | Pond arithmetic unit |
| WO2020244370A1 (en) * | 2019-06-05 | 2020-12-10 | 北京灵汐科技有限公司 | Heterogeneous cooperative system and communication method therefor |
| CN112163673A (en) * | 2020-09-28 | 2021-01-01 | 复旦大学 | Population routing method for large-scale brain-like computing network |
| CN112242963A (en) * | 2020-10-14 | 2021-01-19 | 广东工业大学 | Rapid high-concurrency neural pulse data packet distribution and transmission method |
| 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 |
| US11537879B2 (en) | 2017-07-03 | 2022-12-27 | Tsinghua University | Neural network weight discretizing method, system, device, and readable storage medium |
| US11656910B2 (en) | 2017-08-21 | 2023-05-23 | Shanghai Cambricon Information Technology Co., Ltd | Data sharing system and data sharing method therefor |
| US11687467B2 (en) | 2018-04-28 | 2023-06-27 | Shanghai Cambricon Information Technology Co., Ltd | Data sharing system and data sharing method therefor |
| US11726844B2 (en) | 2017-06-26 | 2023-08-15 | Shanghai Cambricon Information Technology Co., Ltd | Data sharing system and data sharing method therefor |
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 |
|---|
| 罗莉: "数字神经元芯片的设计与应用", 《计算机研究与发展》 * |
Cited By (24)
| 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 |
| US11823030B2 (en) | 2017-01-25 | 2023-11-21 | Tsinghua University | Neural network information receiving method, sending method, system, apparatus and readable storage medium |
| CN106971229A (en) * | 2017-02-17 | 2017-07-21 | 清华大学 | Neural computing nuclear information processing method and system |
| CN106971228A (en) * | 2017-02-17 | 2017-07-21 | 清华大学 | Neuronal messages sending method and system |
| 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 |
| US11726844B2 (en) | 2017-06-26 | 2023-08-15 | Shanghai Cambricon Information Technology Co., Ltd | Data sharing system and data sharing method therefor |
| CN109214616A (en) * | 2017-06-29 | 2019-01-15 | 上海寒武纪信息科技有限公司 | A kind of information processing unit, system and method |
| 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 |
| US11537879B2 (en) | 2017-07-03 | 2022-12-27 | Tsinghua University | Neural network weight discretizing method, system, device, and readable storage medium |
| US11656910B2 (en) | 2017-08-21 | 2023-05-23 | Shanghai Cambricon Information Technology Co., Ltd | Data sharing system and data sharing method therefor |
| CN108334942A (en) * | 2017-12-22 | 2018-07-27 | 清华大学 | Data processing method, device, chip and the 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 |
| CN108345936A (en) * | 2018-01-31 | 2018-07-31 | 清华大学 | A kind of 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 |
| US11687467B2 (en) | 2018-04-28 | 2023-06-27 | Shanghai Cambricon Information Technology Co., Ltd | Data sharing system and data sharing method therefor |
| WO2020244370A1 (en) * | 2019-06-05 | 2020-12-10 | 北京灵汐科技有限公司 | Heterogeneous cooperative system and communication method therefor |
| CN112163673A (en) * | 2020-09-28 | 2021-01-01 | 复旦大学 | Population routing method for large-scale brain-like computing network |
| 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 |
| CN112242963A (en) * | 2020-10-14 | 2021-01-19 | 广东工业大学 | Rapid high-concurrency neural pulse data packet distribution and transmission method |
| 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 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106056212B (en) | 2018-11-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106056212A (en) | Artificial neural network calculating core | |
| CN106056211A (en) | Neuron computing unit, neuron computing module and artificial neural network computing core | |
| EP3340126A1 (en) | Scalable neuromoric core with shared synaptic memory and variable precision synaptic memory | |
| CN105729491B (en) | The execution method, apparatus and system of robot task | |
| EP3343458A1 (en) | Population-based connectivity architecture for spiking neural networks | |
| CN109426647A (en) | For coordinating the technology of the accelerator installation resource of depolymerization | |
| CN104683405B (en) | The method and apparatus of cluster server distribution map matching task in car networking | |
| CN108021451B (en) | Self-adaptive container migration method in fog computing environment | |
| Nie et al. | Network traffic prediction based on deep belief network and spatiotemporal compressive sensing in wireless mesh backbone networks | |
| CN107710237A (en) | Deep neural network divides on server | |
| CN108431796A (en) | Distributed resource management system and method | |
| CN109788489A (en) | A kind of base station planning method and device | |
| Goh et al. | Nonlocal evolution of weighted scale-free networks | |
| CN107678858A (en) | application processing method, device, storage medium and electronic equipment | |
| CN113449839A (en) | Distributed training method, gradient communication device and computing equipment | |
| CN113642734A (en) | Distributed training method and device for deep learning model and computing equipment | |
| CN104717272A (en) | Event stream processing system and method thereof | |
| CN112199154B (en) | Reinforced learning training system and method based on distributed collaborative sampling center type optimization | |
| CN113221475A (en) | Grid self-adaption method for high-precision flow field analysis | |
| CN113452655A (en) | Distributed training method, gradient communication device and computing equipment | |
| CN104854602A (en) | Generating messages from the firing of pre-synaptic neurons | |
| Ding et al. | A hybrid-mode on-chip router for the large-scale FPGA-based neuromorphic platform | |
| CN112242963B (en) | Rapid high-concurrency neural pulse data packet distribution and transmission method and system | |
| CN106681803A (en) | Task scheduling method and server | |
| Wang et al. | C3Meta: a context-aware cloud-edge-end collaboration framework toward green metaverse |
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 |