CN104281559A - Model calculation method and device used for performing function model based on data - Google Patents
Model calculation method and device used for performing function model based on data Download PDFInfo
- Publication number
- CN104281559A CN104281559A CN201410321571.5A CN201410321571A CN104281559A CN 104281559 A CN104281559 A CN 104281559A CN 201410321571 A CN201410321571 A CN 201410321571A CN 104281559 A CN104281559 A CN 104281559A
- Authority
- CN
- China
- Prior art keywords
- computing unit
- model
- calculating
- cycle calculations
- model computing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B17/00—Systems involving the use of models or simulators of said systems
- G05B17/02—Systems involving the use of models or simulators of said systems electric
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/17—Function evaluation by approximation methods, e.g. inter- or extrapolation, smoothing, least mean square method
Abstract
The present invention relates to a model calculation method and device used for performing a function model based on data. A model calculation unit (3) used for calculating a function model based on data in an integrated control assembly (1) comprises: a calculation core (31) configured for performing calculation of algorithm purely based on hardware, wherein the algorithm performs at least one cycle calculation; and an interruption unit (32) configured for inquiring an interruption condition so as to interrupt the cycle calculation in the calculation core (31) according to existence of the interruption condition, and configured for providing a stage result of the cycle calculation and a counter value of the cycle calculation related thereto for recovering the calculation.
Description
Technical field
The present invention relates to the calculating based on the function model of data in the integrated Control Component of model computing unit, described model computing unit is configured to purely based on hardware ground Computation function model.
Background technology
By the known some controllers of prior art, these controllers with integrated Control Component, integrated Control Component with main computation unit and independent model computing unit to calculate the function model based on data.Therefore document DE 10 2,010 028 266 A1 shows a kind of Control Component, and it is with additional logical circuit as model computing unit, and model computing unit is configured to purely based on hardware ground gauge index function and additive operation and multiplying.This point achieves, and supports the calculating especially calculating the Bayesian regression method that Gaussian process model needs in a hardware cell.
Model computing unit is designed to perform the mathematical procedure calculating the function model based on data on the basis in parameter and sampling spot or training data generally.Function for the model computing unit of effective gauge index function and cumulative function is implemented especially purely within hardware, and thus it makes to calculate Gaussian process model by the computing velocity higher than the computing velocity of carrying out in the main computation unit of software control becomes possibility.
Configuration data was usually just ready and carried out the calculating based on configuration data of function model followed by the hardware of model computing unit before model computing unit calculates, wherein, configuration data comprises parameter (hyper parameter) and for calculating based on the number of sampling certificate of the function model of data or the address pointer on address field, address field stores number of sampling certificate.
Summary of the invention
According to the present invention, define by according to claim 1 for calculating in integrated Control Component based on the model computing unit of the function model of data with by the method described in claim arranged side by side.
Other favourable design proposal of the present invention illustrates in the dependent claims.
According to first aspect, specify a kind of model computing unit for calculating the function model based on data in integrated Control Component, it comprises:
-calculating core core, it is configured to the pure hardware based calculating of execution algorithm, and wherein, algorithm performs at least one cycle calculations; And
-interrupt location, it is configured to inquire interrupt condition, to interrupt cycle calculations according to the existence of interrupt condition, and is configured to provide the Counter Value relevant therewith of the phase results of cycle calculations and cycle calculations to calculate for recovering.
When requiring this calculating of the function model based on data of higher-priority in model computing unit, use traditional model computing unit with hardware implementation usually cannot interrupt the calculating of the function model based on data had compared with low priority.
A kind of thought of said method is, specifies the interruptibility of the calculating of the function model based on data in a model computing unit implemented within hardware.When have the calculating of the function model based on data of higher-priority must be had to substitute maybe must be interrupted compared with the calculating of the function model based on data of low priority time, such as may calculative interruptibility.When there being the calculating based on timeslice of sufficiently long time grating to be had another calculating based on switching time of shorter time grating to interrupt, may such as there is this situation.
In addition may occur that, that is, the storage space of inner storer is restricted and this storage space cannot store all sampling spots for the function model based on data.In this case, the calculating time-out by interrupting calculating can be used to reload other sampling spot and rewrite the number of sampling certificate no longer needed.
As mentioned above, the model computing unit implemented within hardware up to now does not specify the interruption that calculates usually.In order to avoid model computing unit stops the calculating started of the function model based on data within one period of long duration, function model based on data can be split into less submodel, and then the phase results of submodel is such as collected at together by addition.But extend calculating prolongeding time thus, because especially need the extra process interrupted each time and restart each time for calculating single in the computing unit of software control to expend.
Described method provide such advantage, that is, the computing interval that model computing unit is continuing is put at a fixed time or whenever can be stopped by the corresponding instruction of main computation unit.Another computing power based on the model computing unit of the calculating of the function model of data for there being higher-priority can be discharged thus, and need not wait for compared with low priority, before calculated, based on the end of the function model of data.Can realize thus in principle interrupting, that is, the feasible program calculating and stop is provided during calculating the function model based on data on specific time point.
In addition, model computing unit can be configured to adjustable land used storage stage result sum counter value or they be stored classifiedly in order to the recovery calculated.The phase results reached of the information continued needed for calculating and the calculating be performed by the time can be kept in this way, can again continue interrupted calculating on time point afterwards.
Can specify, algorithm have one inner with an outside cycle calculations, in these cycle calculations, define summation by the phase results of cycle calculations respectively, wherein, interrupt location be configured to the quantity implementing outside circulation can be predetermined cycle calculations after inquire interrupt condition.
According to a kind of form of implementation, a kind of configuration register write being set, wherein, in order to determine the existence of interrupt condition, write or the content of configuration register can being inquired.
A signal input part especially can be set, wherein, as the look-at-me that interrupt condition can be inquired on signal input part.
Can specify, model computing unit is configured to, the recovery of interrupted calculating before being performed by the setting of the phase results of cycle calculations and the loop counter value of cycle calculations.
Model computing unit can be configured to calculate Bayesian regression.
According to another aspect, arrange a Control Component in particular for the engine system with internal combustion engine, this Control Component comprises a main computation unit with software control and above-mentioned model computing unit.
In addition, main computation unit can write the configuration register of model computing unit, so that predetermined interrupt condition.
According to another aspect, specify a kind of function model for calculating in above-mentioned model computing unit based on data, especially the method for Gaussian process model, wherein, function model is calculated by the algorithm purely implemented within hardware, wherein, this algorithm is configured to perform at least one cycle calculations, wherein, cycle calculations is interrupted in existence according to interrupt condition, wherein, the phase results of cycle calculations and the Counter Value of cycle calculations are stored and are supplied to model computing unit for recalculating.
Accompanying drawing explanation
Next preferred form of implementation of the present invention is explained in detail by accompanying drawing.In accompanying drawing:
Fig. 1 diagrammatically illustrates the hardware architecture for integrated Control Component;
Fig. 2 is the process flow diagram of the method illustrated for interrupting the calculating in model computing unit.
Embodiment
Fig. 1 diagrammatically illustrates the hardware architecture that for such as form is the integrated form Control Component 1 of microcontroller, arranges a main computation unit 2 and one in the microcontroller in an integrated fashion and calculates the model computing unit 3 of the function model based on data for pure based on hardware.Main computation unit 2 and model computing unit 3 pass through an inner communication link 4, such as, by system bus, mutually communicate to connect.
Model computing unit 3 correspondingly is not configured to perform software code as in main computation unit 2 by hard wire substantially in principle.As alternative, such solution is also feasible, in this solution, provides director data group that is restricted, highly-specialised for the model computing unit 3 calculated based on the function model of data.Any processor is not established in model computing unit 3.This point realizes this model computing unit 3 or makes to become possibility with the area-optimized structure of integrated make structure with can optimizing resource.
Model computing unit 3 has one and calculates core core core 31, and it performs the calculating of predetermined algorithm purely within hardware.Calculate core core 31 to be connected with interrupt location 32, when there is interrupt condition, interrupt location is by the look-at-me of the calculating of algorithm.Model computing unit 3 can comprise a local SRAM 33 for stores configuration data in addition.Model computing unit 3 can comprise local DMA unit 34(DMA=Direct Memory Access direct memory access (DMA) equally).The integrated resource of Control Component 1 can be transferred by DMA unit 34, especially transfer inner storer 5.
Control Component 1 can comprise an inner storer 5 and another DMA unit 6(DMA=Direct Memory Access direct memory access (DMA)).Inner storer 5 and DMA unit 6 in a proper manner, such as, by inner communication link 4, are interconnected.Inner storer 5 can comprise a common SRAM memory (main computation unit 2, model computing unit 3 and other units shared if desired) and a flash memory for configuration data (parameter and number of sampling certificate).
Non-parametric, based on the use of the function model of data, based on a kind of Bayesian regression method.The basis of Bayesian regression is explanation in " Gaussian process of Gaussian Processes for Machine Learning(machine learning) " (publishing house of Massachusetts Institute of Technology (MIT) 2006) of C.E.Rasmussen etc. such as.Bayesian regression is a kind of method based on data, and it is based on a model.In order to set up this model, need the output data of the measurement point of training data and the related to this of output parameter.Foundation being used to complete by number of sampling certificate of model, number of sampling is according to corresponding training data wholly or in part or produced by these training datas.Determine abstract hyper parameter in addition, these hyper parameter are by the spatial parameterization of pattern function and effectively affect weighting on the single measurement point of training data to model prediction afterwards.
Abstract hyper parameter is determined by optimization method.A kind of feasible program for this optimization method is optimization marginal likelihood function
.Marginal likelihood function
describing the likelihood degree of the y value through recording of training data, illustrating as vector Y, the x value of setting models parameter H and training data.In model training,
be maximized thus, that is, find suitable hyper parameter, these hyper parameter result in moving towards change and as far as possible accurately describing training data of the pattern function determined by hyper parameter and training data.In order to simplify calculating,
logical algorithm be maximized, because logical algorithm does not change the continuity of plausibility function.
The step that the calculating correspondence of Gaussian process model schematically shows in fig. 2 is carried out.For test point u(input parameter vector) input value
first be standardized, or rather corresponding following formula:
。
At this, m
xthe corresponding mean function about the mean value of the input value of number of sampling certificate, s
ythe variance of the input value of corresponding sample point data and d are to the index of dimension D being applied to test point u.
As non-parametric, based on the modeling result of the function model of data, people obtain:
。
The model value v obtained like this is standardized by outputting standard, or rather by following formula:
。
At this, the corresponding input parameter vector at a standardized test point u(dimension D of v) on standardized model value (output valve),
corresponding to (not standardized) test point
(not standardized) model value (output valve) on (the input parameter vector of dimension D),
the sampling spot of corresponding sample point data, the quantity of the sampling spot of N corresponding sample point data, the dimension of D correspondence input data space/training data space/sampling spot data space, and I
dwith
the corresponding hyper parameter from model training.Vector Q
yit is a parameter calculated from hyper parameter and training data.In addition, m
ythe corresponding mean function about the mean value of the output valve of number of sampling certificate and s
ythe variance of the output valve of corresponding sample point data.
Perform input standardization and outputting standard, because the calculating of Gaussian process model typically occurs in a standardized space.
In order to start to calculate, computing unit 2 especially can order DMA unit 34 or another DMA unit 6, is delivered to by the configuration data relevant to there being function model to be calculated in model computing unit 3 and starts to calculate, calculating and perform by configuration data.Configuration data comprises hyper parameter and the number of sampling certificate of Gaussian process model, and number of sampling is according to the address field being assigned to model computing unit 3 being preferably assigned to internal storage 5 by address pointer.Especially also use the SRAM memory 33 for model computing unit 3, this storer especially can be arranged in model computing unit 3 or side for this reason.Inner storer 5 and SRAM memory 33 also can combine use.
Calculating in model computing unit 3 is carried out in of model computing unit 3 is by the hardware architecture of ensuing Implementation of pseudocode, the computation rule above its correspondence.From false code, calculate and to carry out in an inner circulation and an outside circulation and their phase results is accumulated.Model calculate start time, a typical value starting Parameter N start for counter is zero.
/ * the stage 1: input standardization (Eingangsnormierung) */
001:?for?（k=0;?k＜D；k++）{
002：
[k]?=?u?[k]?*?（s′
x）[k]?+?（m′
x）[k];
003：}
/ * the stage 2: calculate outside circulation */
004:?for?（j=Nstart;?j＜N;?j++）{
005:?i?=?j?*?D;
/ * stage 2a: calculate inner circulation */
006:?t?=?0.0;
007:?for?（1=0;?1＜D；1++）{
008：d?=
[1]?–?v[i+1];
009：d?=?d*d;
010:?t?+=?1′?[1]*d;
011:?}
/ * stage 2b: gauge index function */
012:?e?=?exp（-t）;
/ * stage 2c:*/
013:?y?+=?（Q′
y）[j]?*?e;
014:?}
/ * the stage 3: outputting standard */
015：z?=?m
y；
016:?z+=?y*s
y;
017:?return?z;
Present regulation, calculating above can be interrupted.Interruption can complete thus, that is, interrupt calculating, that is, after the predetermined value reaching cycle counter j after the circular flow of some.This interruption can optionally be added by setting, interruption identifier with good conditionsi completes.As alternative, calculate and also can be completed by the clear and definite interruption identifier of setting.After breaking in the calculation, phase results and the relevant cycle counter j of accumulation are kept in, thus can by setting temporary value to realize the continuity calculated on the time point of afterwards.
The process flow diagram of composition graphs 2 explains a kind of method for interrupting the calculating in model computing unit 3 in detail.For this reason, the calculating of the function model based on data is started in step sl.
Model computing unit 3 is configured to after the iteration of some, and that is, when the cycle counter j of outer loop reaches specifically predetermined loop counter value, interrupt calculating, wherein, loop counter value is less than the sum of provided sampling spot.This inquiry completes in step s 2, such as, in 014 row of above-mentioned false code.If determine not reach predetermined loop counter value (option: no) in step s 2, so continue the execution calculating or continue false code with 005 row with rebound step S2.
If determine in step s 2, reach predetermined loop counter value (option: yes), so inquire in step s3 along with continuity, whether should interrupt calculating.The interruption calculated can such as complete thus, that is, main computation unit 2 specifically interrupts one the configuration register that identifier inserts model computing unit.Then the setting interrupting identifier causes, and selects option "Yes" and model computing unit 3 not to carry out the calculating of the outer loop of the loop counter value j of increase in step s3.Otherwise, be triggered thus based on the calculating of function model (having loop counter value j+1's) the ensuing part of data, that is, interruption identifier be not set.As alternative, apply an interrupting input signal for model computing unit 3 so that wish signalling will be interrupted.Although comprise step S2 and S3 in this unshowned method, in a method, also only can use in two steps, thus or by reaching predetermined loop counter value, or arranging interrupt identifier basis on, realize calculate interruption.Also multiple interruption identifier can be inserted, they with cycle counter, with input signal and if desired to each other respectively by logic UND(and) or ODER(or) connect and be combined like this, make it possible to adjust with good conditionsi or unconditional interruption in different combination feasible programs.
If determine in step s3, should interrupt calculating (option: yes), so the method is along with step S4 continuity, otherwise (option: no) just jumps back to step S2.
In step s 4 which, the phase results y reached up to now and the Counter Value j of outer loop is kept in.Then model computing unit is located in idle mode and is provided for and calculate another function model based on data.Main computation unit 2 or DMA unit 6 especially give such possibility, that is, start new calculating, especially by the initial value of predefined counter value j and the result of calculation y for accumulation.And then start new calculating along with step S1.
Interrupted calculating before also can continue again.If calculating in model computing unit 3 should continue the calculating of the function model based on data after interrupting, Parameter N start can be started by correspondingly predefined counter together with so corresponding Counter Value and temporary loop counter value j and to be initialised (in 004 row in above-mentioned false code) and the phase results y kept in can the scheduled compensation as the functional value for obtaining the function model based on data.During the calculating interrupted before continuing, counter starts Parameter N start and y is typically not equal to zero, because carried out the circulation of some and defined corresponding Classifying Sum (Zwischensumme).
In a kind of alternative form of implementation, the inquiry of the loop counter value I of inner circulation also can perform in the circulation of inside, wherein, when interrupting, inner circulation and the loop counter value I of phase results t must be kept in, to realize the recovery calculated.Because provide the recovery possibility of the calculating of inner circulation may cause the worse efficiency of whole calculating, so can specify, perform interruption by putting corresponding interruption identifier or applying look-at-me, thus inner circulation can not be calculated the end.Then current phase results is not taken into account in the summation of 010 row of false code.The loop counter value j of outside circulation is not increased.Terminating fast of the calculating in hardware computational unit can be reached thus.
Claims (10)
1., for calculating the model computing unit (3) based on the function model of data in integrated Control Component (1), comprising:
-calculating core core (31), it is configured to the pure hardware based calculating of execution algorithm, and wherein, algorithm performs at least one cycle calculations; And
-interrupt location (32), it is configured to inquire interrupt condition, so that according to the cycle calculations that the existence of interrupt condition is interrupted in calculating core core (31), and be configured to provide the Counter Value relevant therewith of the phase results of cycle calculations and cycle calculations to calculate for recovering.
2. by model computing unit (3) according to claim 1, wherein, model computing unit (3) is configured to call ground storage stage result and loop counter value.
3. by model computing unit (3) described in claim 1 or 2, wherein, algorithm has inner with cycle calculations that is outside, in these cycle calculations, defines summation respectively by the phase results of cycle calculations,
Wherein, interrupt location (32) be configured to the quantity implementing outside circulation can be predetermined cycle calculations after interrupt calculating.
4. by model computing unit (3) described in any one of claims 1 to 3, wherein, being provided with configuration register, wherein, write or the content of configuration register can being inquired in order to determine the existence of interrupt condition.
5., by the model computing unit (3) described in any one of Claims 1-4, wherein, be provided with input end, wherein as the look-at-me that interrupt condition can be inquired on input end.
6. by the model computing unit (3) described in any one of claim 1 to 5, wherein, model computing unit (3) be configured to be performed by the setting of the phase results of cycle calculations and the loop counter value of cycle calculations before the recovery of interrupted calculating.
7. by the model computing unit (3) described in any one of claim 1 to 6, wherein, model computing unit (3) is configured to calculate Bayesian regression.
8. Control Component (1), is particularly useful for the controller of the engine system with internal combustion engine, and this Control Component comprises:
The main computation unit (2) of-software control; And
-by the model computing unit (3) described in any one of claim 1 to 7.
9., by Control Component according to claim 8 (1), wherein, main computation unit (2) writes one or more configuration registers of model computing unit (3), so that predetermined interrupt condition.
10. for calculating the method for the function model especially Gaussian process model based on data in by the model computing unit (3) described in any one of claim 1 to 7, wherein, by the algorithm Computation function model implemented within hardware, wherein this algorithm performs at least one cycle calculations, wherein interrupt cycle calculations according to the existence of interrupt condition, wherein the Counter Value of the phase results of cycle calculations and cycle calculations is stored and is supplied to model computing unit (3) for recalculating.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| DE201310213414 DE102013213414A1 (en) | 2013-07-09 | 2013-07-09 | Method and apparatus for performing a model calculation of a data-based function model |
| DE102013213414.2 | 2013-07-09 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN104281559A true CN104281559A (en) | 2015-01-14 |
Family
ID=52107247
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410321571.5A Pending CN104281559A (en) | 2013-07-09 | 2014-07-08 | Model calculation method and device used for performing function model based on data |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN104281559A (en) |
| DE (1) | DE102013213414A1 (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109643393A (en) * | 2016-09-07 | 2019-04-16 | 罗伯特·博世有限公司 | For calculating the model computing unit and controller with the Multilayer Perception model of feedforward and feedback |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2003100655A1 (en) * | 2002-05-28 | 2003-12-04 | Droplet Technology, Inc. | Systems and methods for pile-processing parallel-processors |
| CN102043761A (en) * | 2011-01-04 | 2011-05-04 | 东南大学 | Fourier transform implementation method based on reconfigurable technology |
| CN102360344A (en) * | 2011-10-10 | 2012-02-22 | 西安交通大学 | Matrix processor as well as instruction set and embedded system thereof |
| CN102859158A (en) * | 2010-04-27 | 2013-01-02 | 罗伯特·博世有限公司 | Control device and method for calculating an output parameter for a controller |
-
2013
- 2013-07-09 DE DE201310213414 patent/DE102013213414A1/en not_active Withdrawn
-
2014
- 2014-07-08 CN CN201410321571.5A patent/CN104281559A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2003100655A1 (en) * | 2002-05-28 | 2003-12-04 | Droplet Technology, Inc. | Systems and methods for pile-processing parallel-processors |
| CN102859158A (en) * | 2010-04-27 | 2013-01-02 | 罗伯特·博世有限公司 | Control device and method for calculating an output parameter for a controller |
| CN102043761A (en) * | 2011-01-04 | 2011-05-04 | 东南大学 | Fourier transform implementation method based on reconfigurable technology |
| CN102360344A (en) * | 2011-10-10 | 2012-02-22 | 西安交通大学 | Matrix processor as well as instruction set and embedded system thereof |
Non-Patent Citations (1)
| Title |
|---|
| 阿申登: "《VHDL嵌入式数字系统设计教程》", 30 April 2011 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109643393A (en) * | 2016-09-07 | 2019-04-16 | 罗伯特·博世有限公司 | For calculating the model computing unit and controller with the Multilayer Perception model of feedforward and feedback |
Also Published As
| Publication number | Publication date |
|---|---|
| DE102013213414A1 (en) | 2015-01-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102859158B (en) | For calculating control apparatus and the method for the output parameter controlled | |
| US8731737B2 (en) | Microcontroller having a computing unit and a logic circuit, and method for carrying out computations by a microcontroller for a regulation or a control in a vehicle | |
| US11521116B2 (en) | Self-optimizing multi-core integrated circuit | |
| CN109992210B (en) | Data storage method and device and electronic equipment | |
| US11288047B2 (en) | Heterogenous computer system optimization | |
| US7971082B2 (en) | Method and system for estimating power consumption of integrated circuit design | |
| CN112418416A (en) | Neural network computing system, neural network computing method and computer system | |
| CN104102129A (en) | Method and device for creating a data-based function model | |
| Genius et al. | Model-driven performance evaluation and formal verification for multi-level embedded system design | |
| US10963036B2 (en) | Idle loop detection and control for processors | |
| CN104281559A (en) | Model calculation method and device used for performing function model based on data | |
| US7761280B2 (en) | Data processing apparatus simulation by generating anticipated timing information for bus data transfers | |
| JP2021508861A (en) | Neural network processing methods, computer systems and storage media | |
| JP2019125093A (en) | Electronic control device | |
| JP2015169997A (en) | Program analyzing device, program analyzing method and program | |
| Tang et al. | FPGA implementation of RANSAC algorithm for real-time image geometry estimation | |
| US7346868B2 (en) | Method and system for evaluating design costs of an integrated circuit | |
| CN102063289B (en) | Method and estimator for estimating thread-level speculative execution capability of serial program | |
| US7124311B2 (en) | Method for controlling processor in active/standby mode by third decoder based on instructions sent to a first decoder and the third decoder | |
| CN110929857B (en) | Data processing method and device of neural network | |
| Arndt et al. | FPGA Accelerated NoC-Simulation: A Case Study on the Intel Xeon Phi Ringbus Topology | |
| Penolazzi et al. | Predicting bus contention effects on energy and performance in multi-processor SoCs | |
| WO2020192587A1 (en) | Artificial intelligence computing device and related product | |
| JP6737960B2 (en) | Model calculation unit and control device for calculating multilayer perceptron model | |
| CN104570757A (en) | Method and apparatus for operating integrated control assembly |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20150114 |