[go: up one dir, main page]

CN119806848A - A deep sleep control method and system for an end-side AI audio chip - Google Patents

A deep sleep control method and system for an end-side AI audio chip Download PDF

Info

Publication number
CN119806848A
CN119806848A CN202510294706.1A CN202510294706A CN119806848A CN 119806848 A CN119806848 A CN 119806848A CN 202510294706 A CN202510294706 A CN 202510294706A CN 119806848 A CN119806848 A CN 119806848A
Authority
CN
China
Prior art keywords
task
energy consumption
data
audio processing
processing unit
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
Application number
CN202510294706.1A
Other languages
Chinese (zh)
Other versions
CN119806848B (en
Inventor
李子青
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen Ultra Easy Technology Co ltd
Original Assignee
Shenzhen Ultra Easy Technology Co ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Shenzhen Ultra Easy Technology Co ltd filed Critical Shenzhen Ultra Easy Technology Co ltd
Priority to CN202510294706.1A priority Critical patent/CN119806848B/en
Publication of CN119806848A publication Critical patent/CN119806848A/en
Application granted granted Critical
Publication of CN119806848B publication Critical patent/CN119806848B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02DCLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Power Sources (AREA)

Abstract

The invention relates to the technical field of chip control, and provides a deep sleep control method and a deep sleep control system for an end-side AI audio chip, wherein the deep sleep control method comprises the steps of obtaining current energy consumption data and task execution information of all audio processing units in the end-side AI audio chip, constructing an energy consumption optimization map for analysis, identifying an audio processing unit in a low-load state, mapping a task to a high-efficiency audio processing unit, and carrying out dynamic sleep calculation on the audio processing unit in the low-load state after a task reconstruction sequence is obtained to obtain a deep sleep target unit and sleep parameters thereof; and configuring a power consumption control module of the terminal side AI audio chip by using the dormancy parameters to enable the deep dormancy target unit to enter a low power consumption mode. By analyzing and processing the current energy consumption data and task execution information, flexible and efficient energy consumption management is realized, and the problem that energy consumption cannot be flexibly controlled when task load is frequently changed and energy consumption optimization is not flexible is solved.

Description

Deep sleep control method and system for terminal side AI audio chip
Technical Field
The application relates to the technical field of chip control, in particular to a deep sleep control method and a deep sleep control system for an end-side AI audio chip.
Background
In recent years, with the rapid development of artificial intelligence technology, an end-side AI audio chip is widely applied in the fields of smart home, smart speakers, smart wearable devices and the like. The devices realize more intelligent audio processing and interaction functions through the terminal side AI audio chip, and meet the requirements of users on high efficiency and low power consumption.
In the related technical means, the audio chip dormancy control technology is adopted to reduce power consumption mainly through dormancy and wake-up strategies in a fixed period, so that energy consumption is reduced, the endurance time of equipment can be prolonged to some extent, and the power consumption optimization effect is limited by setting of the dormancy period.
According to the technical scheme, although the power consumption of the audio processing unit can be reduced through the sleep and wake-up strategies with fixed periods, when the task load is frequently changed, the energy consumption cannot be flexibly controlled, the power consumption optimizing effect is limited by the setting of the sleep period, and the problem that the energy consumption optimizing is not flexible exists.
Disclosure of Invention
In order to solve the problem that energy consumption cannot be flexibly controlled when task load is frequently changed and the energy consumption optimization is not flexible enough, the application provides a deep sleep control method and a deep sleep control system for an end-side AI audio chip.
The invention provides a deep sleep control method of an end side AI audio chip, which comprises the steps of obtaining current energy consumption data and task execution information of all audio processing units in the end side AI audio chip, constructing an energy consumption optimization map based on the current energy consumption data and the task execution information, carrying out topology analysis on the energy consumption optimization map, identifying an audio processing unit in a low-load state, mapping the task of the audio processing unit in the low-load state to a high-efficiency audio processing unit according to a preset task migration mapping strategy to obtain a task reconstruction sequence, carrying out dynamic sleep calculation on the audio processing unit in the low-load state based on the task reconstruction sequence to obtain a deep sleep target unit and sleep parameters thereof, configuring a power consumption control module of the end side AI audio chip by utilizing the sleep parameters of the deep sleep target unit to enable the deep sleep target unit to enter a low-power consumption mode, recording a sleep state sequence, carrying out self-adaptive calculation on the deep sleep target unit according to the sleep state sequence and task scheduling information of the end side AI audio chip to obtain a wake-up time and a corresponding energy consumption threshold, and carrying out the deep sleep control on the deep sleep target unit according to the wake-up time and the corresponding energy consumption threshold to enable the deep sleep target unit to enter the deep sleep control state.
The method comprises the steps of obtaining current energy consumption data and task execution information of all audio processing units in an end-side AI audio chip through a power monitoring module and a timestamp recording module, preprocessing the current energy consumption data and the task execution information by utilizing Z-score standardization to obtain energy consumption preprocessing data and task execution characteristic data, calculating energy consumption load levels and task execution intensity of all audio processing units based on the energy consumption preprocessing data to obtain energy consumption load evaluation data, building task dependency relations among the audio processing units according to the task execution characteristic data to obtain task topology structure data, optimizing the energy consumption load evaluation data through the task topology structure data, building energy consumption optimization based on the optimized energy consumption load evaluation data, and carrying out topology analysis on the energy consumption optimization audio processing unit to obtain the low-load analysis on the task execution characteristic data.
The method comprises the steps of analyzing the low-load state audio processing unit to obtain low-load state identification data and low-load state task data, calculating energy consumption change trends of the low-load state audio processing unit under different task migration conditions to obtain energy consumption change trend data, and identifying a transferable task set and corresponding task migration priorities according to the energy consumption change trend data to obtain task migration priority data.
The method comprises the steps of judging candidate high-efficiency audio processing units according to low-load state identification data, low-load state task data and task migration priority data in combination with a preset task migration mapping strategy, generating candidate high-efficiency audio processing unit candidate set data according to the candidate high-efficiency audio processing units, calculating task carrying capacity and task matching degree of the high-efficiency audio processing unit candidate set data to obtain task carrying capacity data and task execution matching degree data, selecting an optimal high-efficiency audio processing unit according to the task execution matching degree data, mapping the low-load state task data to the optimal high-efficiency audio processing unit to obtain task reconstruction sequence data, calculating the influence of the task reconstruction sequence data on the low-load state audio processing unit by the task migration priority data, calculating the sleep time of the task reconstruction sequence data on the low-load state audio processing unit, and obtaining sleep time corresponding to the sleep state data, and determining the sleep time of the sleep mode according to the task execution matching degree data, and obtaining the sleep time of the sleep mode.
As a preferred solution, the calculating the task carrying capacity and the task matching degree of the candidate set data of the high-performance audio processing unit obtains the following calculation formulas of the task carrying capacity data and the task execution matching degree data:
Wherein, For the task-carrying capacity of a high-performance audio processing unit,For the number of candidate high performance audio processing units,Is the firstThe processing power of the candidate audio processing units,Is the firstLoad factors of the candidate audio processing units;
Wherein, The degree of matching is performed for the task,As a function of the intensity coefficient of the task,The processing power required for the current task.
As a preferred solution, the calculation formula for calculating the influence of the task reconstruction sequence data on the low-load state audio processing unit to obtain the energy consumption data after task migration of the low-load state audio processing unit is as follows:
Wherein, As the energy consumption data after the task migration,For the number of audio processing units in the task reconstruction sequence,Is the firstPower consumption of the audio processing units after migration,Is the firstThe time required for the audio processing units to migrate the task.
The method comprises the steps of transmitting the sleep parameters to the power consumption control module of the terminal side AI audio chip, configuring the power consumption control module of the terminal side AI audio chip by using the sleep parameters to enable the deep sleep target unit to enter a low power consumption mode, and recording a sleep state sequence, wherein the power consumption control module is configured to control the start of the low power consumption mode to enable the deep sleep target unit to enter a sleep state, and the sleep state change sequence of each audio processing unit is recorded in the low power consumption mode to obtain the sleep state sequence.
The method comprises the steps of obtaining task scheduling information of an end side AI audio chip, establishing a mapping relation between the task scheduling information and the sleep state sequence, analyzing the mapping relation by utilizing a decision tree algorithm, determining relevance between task scheduling and sleep time to obtain a relevance result, predicting task load information in the next period through the task scheduling information, determining the wake-up requirement of the deep sleep target unit according to the task load information and the relevance result, calculating the wake-up time of each task based on the wake-up requirement, and calculating the energy consumption threshold of the deep sleep target unit according to the wake-up time.
The method comprises the steps of generating a wake-up control instruction based on the wake-up time and the corresponding energy consumption threshold, sending the wake-up control instruction to the deep sleep target unit through a power consumption control module, starting a recovery process, judging whether current power consumption meets a preset energy consumption threshold, if not, delaying the wake-up time, and if yes, starting a wake-up program of the deep sleep target unit, and recovering the task execution capacity of the deep sleep target unit.
The application further provides a deep sleep control system of the terminal side AI audio chip, which comprises an acquisition unit, a mapping unit, a configuration unit, a calculation unit and a calculation unit, wherein the acquisition unit is used for acquiring current energy consumption data and task execution information of all audio processing units in the terminal side AI audio chip, constructing an energy consumption optimization map based on the current energy consumption data and the task execution information, performing topology analysis on the energy consumption optimization map to identify the audio processing units in a low-load state, the mapping unit is used for mapping tasks of the audio processing units in the low-load state to the high-efficiency audio processing units according to a preset task migration mapping strategy to obtain a task reconstruction sequence, performing dynamic sleep calculation on the audio processing units in the low-load state based on the task reconstruction sequence to obtain a deep sleep target unit and sleep parameters thereof, the configuration unit is used for configuring a power consumption control module of the terminal side AI audio chip by utilizing the sleep parameters of the deep sleep target unit to enable the deep sleep target unit to enter a low-power consumption mode, recording a sleep state sequence, the calculation unit is used for performing self-adaption calculation on the wake-up depth target unit according to the sleep state sequence and the task scheduling information of the terminal side AI chip to obtain wake-up time, and the wake-up time of the corresponding to the wake-up target unit is used for enabling the task to enter a sleep state to be controlled according to the sleep time and the sleep time.
Compared with the prior art, the application has the following beneficial effects of long endurance time and flexible energy consumption optimization. The method comprises the steps of obtaining current energy consumption data and task execution information of all audio processing units in an AI audio chip at an end side, constructing an energy consumption optimization map, carrying out topology analysis to identify audio processing units in a low-load state, carrying out task reconstruction and dynamic dormancy calculation on the low-load state units according to a preset task migration mapping strategy to obtain a deep dormancy target unit and dormancy parameters thereof, configuring a power consumption control module by using the dormancy parameters to enable the deep dormancy target unit to enter a low-power consumption mode and record a dormancy state sequence, carrying out self-adaptive awakening calculation based on the dormancy state sequence and task scheduling information, determining awakening time and energy consumption threshold value, carrying out restoration control on the deep dormancy target unit to reenter a task execution state, realizing more flexible and efficient energy consumption management by dynamically adapting to the change of task loads, effectively prolonging the endurance time of equipment, and solving the problems that energy consumption cannot be flexibly controlled and energy consumption optimization is not flexible enough when task loads are frequently changed.
Drawings
In order to more clearly illustrate the embodiments of the invention or the technical solutions of the prior art, the drawings which are used in the description of the embodiments or the prior art will be briefly described, it being obvious that the drawings in the description below are only some embodiments of the invention, and that other drawings can be obtained from these drawings without inventive faculty for a person skilled in the art.
The structures, proportions, sizes, etc. shown in the drawings are shown only in connection with the present disclosure, and are not intended to limit the scope of the invention, since any modification, variation in proportions, or adjustment of the size, etc. of the structures, proportions, etc. should be considered as falling within the spirit and scope of the invention, without affecting the effect or achievement of the objective.
Fig. 1 is a flow chart of a deep sleep control method of an end-side AI audio chip according to an embodiment of the present invention;
Fig. 2 is a schematic block diagram of a deep sleep control system of an end-side AI audio chip according to an embodiment of the present invention.
Reference numerals illustrate:
10. The terminal side AI audio chip deep sleep control system comprises an acquisition unit, a mapping unit, a configuration unit, a calculation unit, a control unit and a control unit.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are some, but not all embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The flow diagrams depicted in the figures are merely illustrative and not necessarily all of the elements and operations/steps are included or performed in the order described. For example, some operations/steps may be further divided, combined, or partially combined, so that the order of actual execution may be changed according to actual situations.
It is also to be understood that the terminology used in the description of the application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.
It should be further understood that the term "and/or" as used in the present specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes such combinations.
The technical scheme of the invention is further described below by the specific embodiments with reference to the accompanying drawings.
Example 1:
As shown in fig. 1, the present application provides a deep sleep control method of an end-side AI audio chip, which includes steps S100 to S500;
Step S100, current energy consumption data and task execution information of all audio processing units in the terminal side AI audio chip are obtained, an energy consumption optimization map is constructed based on the current energy consumption data and the task execution information, topology analysis is carried out on the energy consumption optimization map, and the audio processing units in a low-load state are identified.
In the step, the current energy consumption data and task execution information of all audio processing units in the terminal side AI audio chip are collected, and an energy consumption optimization map is constructed by using a data analysis tool. Specifically, a graph theory algorithm is adopted to perform topology analysis on the energy consumption optimization graph, an audio processing unit in a low-load state is identified, and the audio processing unit is marked as a potential deep sleep target unit.
For example, at a specific moment, the energy consumption data of the audio processing units A, B and C are respectively 0.1W, 0.05W and 0.02W, and the energy consumption data of the audio processing unit D, E are respectively 0.4W and 0.35W, then A, B and C are identified as low-load state units after the spectrum analysis.
Step 200, mapping the task of the audio processing unit in the low-load state to the high-efficiency audio processing unit according to a preset task migration mapping strategy to obtain a task reconstruction sequence, and performing dynamic dormancy calculation on the audio processing unit in the low-load state based on the task reconstruction sequence to obtain a deep dormancy target unit and dormancy parameters thereof.
In this step, the task migration mapping strategy is used to redistribute the task of the audio processing unit in the low-load state to the high-performance audio processing unit. Specifically, a task reconstruction sequence is generated through a task migration mapping algorithm, and dynamic dormancy calculation is performed on the audio processing unit in a low-load state based on the reconstruction sequence, so that a deep dormancy target unit and corresponding dormancy parameters thereof are obtained.
For example, assuming that after task T1 of audio processing unit a is mapped to audio processing unit D, D's power consumption data rises to 0.45w, a is a deep sleep target unit, and its sleep parameter is set to a sleep period of 30 minutes.
And step S300, configuring a power consumption control module of the terminal side AI audio chip by utilizing the sleep parameters of the deep sleep target unit, so that the deep sleep target unit enters a low power consumption mode, and recording a sleep state sequence.
In this step, the sleep parameters of the deep sleep target unit are used to configure the power consumption control module of the end-side AI audio chip. Specifically, the sleep parameters of each deep sleep target unit are input into the power consumption control module, so that the deep sleep target unit enters a low power consumption mode, and the sleep state sequence of the deep sleep target unit is recorded.
For example, after the sleep parameter of the audio processing unit a is input to the power consumption control module, a enters a low power consumption mode, the power consumption data of the low power consumption mode is reduced to 0.01W, and the sleep state sequence of a is recorded as "30 minutes sleep".
And step 400, performing self-adaptive wake-up calculation on the deep sleep target unit according to the sleep state sequence and task scheduling information of the terminal side AI audio chip to obtain wake-up time and a corresponding energy consumption threshold.
In this step, adaptive wake-up computation is performed based on the sleep state sequence of the deep sleep target unit and the task scheduling information of the end-side AI audio chip. Specifically, by analyzing the task execution priority and the current energy consumption state of the deep sleep target unit in the task scheduling information, an appropriate wake-up opportunity and a corresponding energy consumption threshold are determined.
For example, the audio processing unit a increases the task execution priority 5 minutes before the sleep period, and the current energy consumption state approaches to the preset threshold value of 0.01W, and the wake-up time is set to be the sleep period end time, and the energy consumption threshold value is 0.02W.
And S500, performing recovery control on the deep sleep target unit according to the awakening time and the corresponding energy consumption threshold value, so that the deep sleep target unit reenters the task execution state.
In this step, the deep sleep target unit is subjected to recovery control according to its wake-up timing and energy consumption threshold. Specifically, when the preset wake-up time and the energy consumption threshold are reached, the low-power mode of the deep sleep target unit is released, and the deep sleep target unit reenters the task execution state.
For example, when the sleep period of the audio processing unit a ends and the current power consumption reaches 0.02W, resume control is performed on a so that it re-executes task T1.
In this embodiment, by acquiring current energy consumption data and task execution information of all audio processing units in the terminal-side AI audio chip, an energy consumption optimization map is constructed based on the current energy consumption data and task execution information. And then, carrying out topology analysis on the energy consumption optimization map, and identifying the audio processing unit in the low-load state. And according to a preset task migration mapping strategy, mapping the task of the audio processing unit in the low-load state to the high-efficiency audio processing unit to obtain a task reconstruction sequence. And based on the task reconstruction sequence, carrying out dynamic dormancy calculation on the audio processing unit in the low-load state to obtain a deep dormancy target unit and dormancy parameters thereof. And finally, configuring a power consumption control module of the terminal side AI audio chip by utilizing sleep parameters of the deep sleep target unit, enabling the deep sleep target unit to enter a low power consumption mode, recording a sleep state sequence, and simultaneously carrying out self-adaptive wake-up calculation on the deep sleep target unit according to the sleep state sequence and task scheduling information of the terminal side AI audio chip to obtain wake-up time and a corresponding energy consumption threshold value, thereby realizing the recovery control of the deep sleep target unit. The audio processing unit capable of accurately identifying and dynamically dormancy in a low-load state is more flexible and efficient in energy consumption management. Compared with a sleep strategy and a wake-up strategy with fixed periods, the scheme provides a better power consumption control effect under the condition that task loads are changed frequently, so that the duration of equipment is prolonged, the whole energy consumption is reduced, the working efficiency of an end-side AI audio chip is improved, the problem that the energy consumption cannot be controlled flexibly when the task loads are changed frequently is solved, and the problem that the energy consumption optimization is not flexible is solved.
Example 2:
In step S100, current energy consumption data and task execution information of each audio processing unit in the terminal AI audio chip are obtained through a power monitoring module and a timestamp recording module, and the current energy consumption data and task execution information are preprocessed by using Z-score standardization, so as to obtain energy consumption preprocessing data and task execution characteristic data.
The power consumption of the audio processing units is monitored in real time through the power monitoring module, and the current energy consumption data and task execution information of all the audio processing units in the terminal side AI audio chip are collected by combining the time information of the time stamping module. Specifically, the data are subjected to Z-score standardization processing so as to eliminate energy consumption difference among different processing units and dimension influence of task execution characteristic data, and consistency and comparability of the data are ensured.
For example, the current energy consumption data of the audio processing unit A is 0.1W, the task execution information is 20 times/second, the normalized energy consumption preprocessing data is-0.5, and the task execution characteristic data is 1.5.
And calculating the energy consumption load level and task execution intensity of each audio processing unit based on the energy consumption preprocessing data to obtain energy consumption load assessment data.
And calculating the energy consumption load level of each audio processing unit by analyzing the energy consumption preprocessing data, and calculating the task execution intensity by combining the task execution characteristic data. Specifically, the energy consumption preprocessing data are processed by adopting a weighted average algorithm to obtain the energy consumption load level of each audio processing unit, and the task execution intensity is calculated through a linear regression model by combining the task execution characteristic data.
For example, the energy consumption load level of the audio processing unit a is 0.2, the task execution intensity is 0.8, and the obtained energy consumption load evaluation data is (0.2, 0.8).
And constructing task dependency relations among the audio processing units according to the task execution characteristic data to obtain task topological structure data.
And establishing task dependency relations among the audio processing units by analyzing the task execution characteristic data. Specifically, the task execution characteristic data is processed by using a correlation analysis method, task dependency relations with higher correlation are identified, and task topological structure data is constructed based on the dependency relations.
For example, the task execution correlation between the audio processing units a and B is 0.85, and the task dependency relationship between a and B is established.
And optimizing the energy consumption load evaluation data through the task topological structure data, and constructing an energy consumption optimization map based on the optimized energy consumption load evaluation data.
And optimizing the energy consumption load evaluation data by combining the task topology structure data with the energy consumption load evaluation data. Specifically, the task topological structure data is optimized by adopting a graph optimization algorithm, optimized energy consumption load evaluation data are obtained, and an energy consumption optimization graph is constructed based on the data.
For example, the optimized energy consumption load evaluation data of the audio processing units a and B are respectively 0.15 and 0.75, and when the energy consumption optimization map is constructed, the a and the B are connected to represent the task dependency relationship.
And performing topology analysis on the energy consumption optimization map to identify the audio processing unit in a low-load state.
And identifying the audio processing unit in a low-load state by carrying out topology analysis on the constructed energy consumption optimization map. Specifically, the energy consumption optimization map is analyzed node by using a graph traversal algorithm, and an audio processing unit with low energy consumption load level and low task execution intensity is identified.
For example, the power consumption load level and task execution intensity of the audio processing unit a are 0.1 and 0.2, respectively, and the audio processing unit is identified as a low-load state.
After the step of identifying the audio processing unit in the low-load state, analyzing the audio processing unit in the low-load state to obtain low-load state identification data and low-load state task data.
And extracting the low-load state identification data and the low-load state task data by analyzing the low-load state audio processing unit in detail. Specifically, the audio processing unit in the low-load state is analyzed one by one, the current task execution state and the energy consumption data of the audio processing unit are recorded, low-load state identification data are formed, and the corresponding task information is recorded as low-load state task data.
For example, the low load state identification data of the audio processing unit a is "low load", and the task data is "task T1".
And calculating the energy consumption change trend of the audio processing unit in the low-load state under different task migration conditions, and obtaining energy consumption change trend data.
And (3) calculating the energy consumption change trend of the low-load state audio processing unit by performing simulation analysis on the low-load state audio processing unit under different task migration conditions. Specifically, an energy consumption simulation tool is adopted to simulate the energy consumption of the audio processing unit under different task migration conditions, and energy consumption change trend data is obtained.
For example, when task T1 is migrated from audio processing unit a to audio processing unit B, the power consumption of a decreases from 0.1W to 0.02W.
And identifying a transferable task set and a corresponding task transfer priority according to the energy consumption change trend data, and obtaining task transfer priority data.
And identifying the transferable task set and the corresponding task transfer priority by analyzing the energy consumption change trend data. Specifically, the task migration priority is determined by sequencing the energy consumption change trend under different task migration conditions.
For example, both tasks T1 and T2 may be migrated from the audio processing unit a, with the migration priority of T1 being higher than T2, resulting in task migration priority data { T1, T2}.
In step S200, according to the low-load state identification data, the low-load state task data and the task migration priority data in combination with a preset task migration mapping policy, a candidate high-performance audio processing unit is determined, and high-performance audio processing unit candidate set data is generated based on the candidate high-performance audio processing unit.
And judging the most suitable candidate high-efficiency audio processing unit by analyzing the low-load state identification data, the low-load state task data and the task migration priority data and combining a preset task migration mapping strategy. Specifically, according to a task migration mapping strategy, matching low-load state task data to candidate high-performance audio processing units to generate high-performance audio processing unit candidate set data.
For example, task T1 is adapted to migrate from audio processing unit A to audio processing units B and C, forming a candidate set { B, C }.
And calculating the task bearing capacity and the task matching degree of the high-efficiency audio processing unit candidate set data to obtain task bearing capacity data and task execution matching degree data.
And analyzing the candidate set data of the high-efficiency audio processing unit to calculate the task carrying capacity and the task matching degree. Specifically, the task carrying capacity of the candidate high-performance audio processing units is calculated by evaluating the processing capacity and the load condition of the candidate high-performance audio processing units, and the task matching degree is calculated by comparing task execution characteristic data.
For example, the task carrying capacities of the audio processing units B and C are 0.8 and 0.7, respectively, the task matching degrees are 0.9 and 0.85, respectively, and the obtained task carrying capacity data and task matching degree data are { (B, 0.8, 0.9), (C, 0.7, 0.85) }.
And selecting an optimal high-efficiency audio processing unit according to the task execution matching degree data, and mapping the task data in the low-load state to the optimal high-efficiency audio processing unit to obtain task reconstruction sequence data.
And selecting an optimal high-efficiency audio processing unit by analyzing task execution matching degree data, and performing task reconstruction mapping on low-load state task data. Specifically, according to the task matching degree data, a high-performance audio processing unit with the highest matching degree is selected, and task data in a low load state is mapped to the unit to generate task reconstruction sequence data.
For example, the optimal high-performance audio processing unit of task T1 is B, and task T1 is mapped to audio processing unit B to form task reconstruction sequence data { (T1, B) }.
And calculating the influence of the task reconstruction sequence data on the low-load-state audio processing unit to obtain the energy consumption data after the task migration of the low-load-state audio processing unit.
And (3) calculating the energy consumption influence of the task reconstruction sequence data on the low-load state audio processing unit by performing simulation analysis on the task reconstruction sequence data. Specifically, an energy consumption simulation tool is adopted to simulate task reconstruction sequence data, and energy consumption data of the low-load state audio processing unit after task migration is calculated.
For example, after the task T1 is migrated to B, the energy consumption of the audio processing unit a is reduced to 0.02W, and the obtained energy consumption data after the task migration of a is 0.02W.
And calculating an optimal sleep mode of the low-load state audio processing unit through the energy consumption data after task migration to obtain sleep mode data and sleep time data.
And calculating the optimal dormancy mode of the low-load state audio processing unit by analyzing the energy consumption data after task migration of the low-load state audio processing unit. Specifically, the energy consumption simulation tool is utilized to simulate different sleep modes, the sleep mode with the lowest energy consumption is selected, the sleep time is set according to the task execution characteristic data, and the sleep mode data and the sleep time data are generated.
For example, for the audio processing unit a, after simulation analysis, it is determined that its optimal sleep mode is "deep sleep", and the sleep time is "45 minutes".
And determining a deep sleep target unit based on the sleep mode data and the sleep time data, and calculating sleep parameters corresponding to the deep sleep target unit.
And determining a deep sleep target unit by comprehensively analyzing the sleep mode data and the sleep time data, and calculating sleep parameters of the deep sleep target unit. Specifically, according to the optimal sleep mode and sleep time of the low-load state audio processing unit, sleep parameters are set, and the minimum energy consumption is realized on the premise of ensuring the task execution quality.
For example, for audio processing unit a, the sleep parameters include "deep sleep mode" and "45 minutes sleep time".
The task bearing capacity and task matching degree of the high-efficiency audio processing unit candidate set data are calculated, and a calculation formula for obtaining the task bearing capacity data and task execution matching degree data is as follows:
Wherein, For the task-carrying capacity of a high-performance audio processing unit,For the number of candidate high performance audio processing units,Is the firstThe processing power of the candidate audio processing units,Is the firstThe load factor of the candidate audio processing units.
The task carrying capacity of each unit is calculated by applying the above formula to the data of the candidate high performance audio processing units. Specifically, the processing capacity of each candidate unit is multiplied by a load factor, and the total task carrying capacity is accumulated.
For example, the candidate high-performance audio processing unit B has a processing power of 0.8, and a load factor of 0.6, and the task carrying power is 0.48.
Wherein, The degree of matching is performed for the task,As a function of the intensity coefficient of the task,The processing power required for the current task.
And calculating the task execution matching degree through the formula. Specifically, the task carrying capacity is multiplied by the task intensity coefficient and divided by the processing capacity required by the current task to obtain the task matching degree.
For example, the task strength coefficient of the task T1 is 1.2, the processing power required for the current task is 0.5, and the matching degree is 1.15.
The method comprises the steps of calculating the influence of task reconstruction sequence data on an audio processing unit in a low-load state, and obtaining a calculation formula of energy consumption data after task migration of the audio processing unit in the low-load state, wherein the calculation formula is as follows:
Wherein, As the energy consumption data after the task migration,For the number of audio processing units in the task reconstruction sequence,Is the firstPower consumption of the audio processing units after migration,Is the firstThe time required for the audio processing units to migrate the task.
And calculating the energy consumption data after the task migration of the low-load state audio processing unit through the formula. Specifically, the power consumption of each migrated audio processing unit is multiplied by the time required for the migration task, and the total migrated energy consumption data is accumulated.
For example, the power consumption of the audio processing unit a is 0.1W, the time required for task migration is 30 minutes, and the power consumption after task migration is 0.05W.
In step S300, the sleep parameter is transmitted to the power consumption control module of the end-side AI audio chip, and the power consumption control module is configured to control the start of the low power consumption mode, so that the deep sleep target unit enters the sleep state.
And the power consumption control module is configured to realize the starting of a low-power consumption mode by transmitting the dormancy parameters to the power consumption control module of the terminal side AI audio chip. Specifically, according to the setting of the sleep parameter, the configuration of the power consumption control module is adjusted to enable the deep sleep target unit to enter a sleep state.
For example, the sleep parameter of the audio processing unit a is transmitted to the power consumption control module, and is set to "deep sleep mode", and a enters the sleep state after being started.
And in the low power consumption mode, recording the sleep state change sequence of each audio processing unit to obtain a sleep state sequence.
By monitoring the low power consumption mode of the audio processing unit in real time, the change sequence of the sleep state of the audio processing unit is recorded. Specifically, the state recording module is used for recording the sleep state of each audio processing unit one by one to generate a sleep state sequence.
For example, the audio processing unit a is in a sleep state with a sleep state sequence of { "00:00:00", "deep sleep", "00:30:00", "low power" }.
In step S400, task scheduling information of the terminal AI audio chip is obtained, a mapping relationship between the task scheduling information and the sleep state sequence is established, the mapping relationship is analyzed by using a decision tree algorithm, and the association between task scheduling and sleep time is determined, so as to obtain an association result.
And establishing a mapping relation by analyzing task scheduling information of the terminal AI audio chip and a dormant state sequence of the audio processing unit. Specifically, task scheduling information is acquired through a task scheduling module and is associated with a sleep state sequence, a decision tree algorithm is utilized to analyze the mapping relation, and the association between task scheduling and sleep time is determined.
For example, the correlation between the scheduling information of the task T1 and the sleep state "deep sleep" of the audio processing unit a is 0.95, and the obtained correlation result is { T1, a, "deep sleep", 0.95}.
And predicting task load information in the next period through task scheduling information, and determining the wake-up requirement of the deep sleep target unit according to the task load information and the association result.
And (3) estimating task load information in the next period by analyzing and predicting task scheduling information, and determining the wake-up requirement of the deep sleep target unit by combining the association result. Specifically, the task scheduling information is modeled by using a prediction model, future task loads are predicted, and whether the deep sleep target unit needs to be awakened or not is determined by combining the association result.
For example, it is predicted that the load of task T1 will increase in the next cycle, while audio processing unit A is currently in a "deep sleep" state, requiring a wake-up, with the wake-up demand being "high".
And calculating the wake-up time of each task based on the wake-up requirement, and calculating the energy consumption threshold of the deep sleep target unit according to the wake-up time.
And calculating the optimal wake-up time of each task by quantitatively analyzing the wake-up requirement, and further calculating the energy consumption threshold of the deep sleep target unit. Specifically, according to task load and wake-up requirements, wake-up time of each task is calculated by combining an energy consumption model, and a corresponding energy consumption threshold is determined.
For example, task T1 has an optimal wake-up time of "01:00:00" and a corresponding energy consumption threshold of 0.05W.
In step S500, a wake-up control command is generated based on the wake-up timing and the corresponding energy consumption threshold, and the power consumption control module sends the wake-up control command to the deep sleep target unit to start the recovery process.
And generating a corresponding wake-up control instruction by analyzing the calculated wake-up time and the corresponding energy consumption threshold, sending the wake-up control instruction to the deep sleep target unit through the power consumption control module, and starting a recovery process of the deep sleep target unit. Specifically, a wake-up control instruction is transmitted to the power consumption control module, and a wake-up resume operation of the deep sleep target unit is started according to the instruction.
For example, at "01:00:00", a wake-up control instruction is sent to the audio processing unit a, causing it to resume from the "deep sleep" state.
Judging whether the current power consumption meets a preset energy consumption threshold, and if not, delaying the awakening time.
And judging whether the current power consumption of the deep sleep target unit meets a preset energy consumption threshold or not by monitoring the current power consumption of the deep sleep target unit. Specifically, the energy consumption monitoring module monitors the power consumption data of the audio processing unit in real time, and if the current power consumption does not reach the preset energy consumption threshold, the wake-up time is adjusted, and the wake-up operation is delayed.
For example, if the current power consumption of the audio processing unit a is 0.03W and the preset threshold value of 0.05W is not reached, the wake-up opportunity is delayed to "01:10:00".
If the preset energy consumption threshold is met, a wake-up program of the deep sleep target unit is started, and the task execution capacity of the deep sleep target unit is restored.
And after the current power consumption of the deep sleep target unit reaches or exceeds a preset energy consumption threshold value, the energy consumption monitoring module starts a wake-up program of the deep sleep target unit. Specifically, the power consumption control module sends an instruction to start a wake-up program, so that the deep sleep target unit resumes the task execution capacity.
For example, when the audio processing unit a is in the "01:10:00", the current power consumption reaches 0.06W, the preset threshold value 0.05W is met, the wake-up program is started, and the execution capacity of the task T1 is restored.
In this embodiment, current energy consumption data and task execution information of each audio processing unit in the terminal AI audio chip are obtained through the power monitoring module and the timestamp recording module, and preprocessing is performed by using Z-score standardization, so as to obtain energy consumption preprocessing data and task execution characteristic data. And calculating the energy consumption load level and task execution intensity of each audio processing unit based on the energy consumption preprocessing data, and constructing task dependency relations among the audio processing units to form task topological structure data. And then optimizing the energy consumption load evaluation data through task topological structure data, constructing an energy consumption optimizing map, performing topological analysis on the energy consumption optimizing map, and identifying the audio processing unit in a low-load state. Analyzing the audio processing unit in the low-load state, calculating the energy consumption change trend of the audio processing unit under different task migration conditions, and identifying a migratable task set and task migration priority. And judging candidate high-efficiency audio processing units according to the low-load state identification data, the low-load state task data and the task migration priority data and combining with a preset task migration mapping strategy, and generating high-efficiency audio processing unit candidate set data. And selecting an optimal high-efficiency audio processing unit by calculating the task bearing capacity and the task matching degree of the candidate set of the high-efficiency audio processing unit, mapping low-load state task data to the optimal unit, and generating task reconstruction sequence data. And calculating the influence of the task reconstruction sequence data on the low-load state audio processing unit to obtain the energy consumption data after task migration, calculating an optimal sleep mode according to the energy consumption data, and determining a deep sleep target unit and sleep parameters thereof. And transmitting the dormancy parameters to a power consumption control module, enabling the deep dormancy target unit to enter a low power consumption mode, and recording a dormancy state sequence. Based on the mapping relation between task scheduling information and the dormancy state sequence, determining the relevance between task scheduling and dormancy time by utilizing a decision tree algorithm, predicting task load information in the next period, calculating a wakeup time and a corresponding energy consumption threshold value, generating a wakeup control instruction, and carrying out recovery control by a power consumption control module. In the wake-up process, if the current power consumption does not reach the preset energy consumption threshold, the wake-up time is delayed, and if the current power consumption meets the preset energy consumption threshold, a wake-up program is started, so that the deep sleep target unit resumes the task execution capacity. By the mode, accurate energy consumption management of each audio processing unit in the opposite-end AI audio chip is realized, and the energy efficiency and the working efficiency of the system are improved.
Example 3:
as shown in fig. 2, the present application further provides a deep sleep control system 10 of an end-side AI audio chip, which includes an acquisition unit 11, a mapping unit 12, a configuration unit 13, a calculation unit 14, and a control unit 15.
The acquiring unit 11 is mainly configured to acquire current energy consumption data and task execution information of all audio processing units in the terminal AI audio chip, construct an energy consumption optimization map based on the current energy consumption data and the task execution information, perform topology analysis on the energy consumption optimization map, and identify the audio processing units in a low-load state.
The mapping unit 12 is mainly configured to map the task of the audio processing unit in the low-load state to the high-performance audio processing unit according to a preset task migration mapping policy, obtain a task reconstruction sequence, and perform dynamic dormancy calculation on the audio processing unit in the low-load state based on the task reconstruction sequence, so as to obtain a deep dormancy target unit and dormancy parameters thereof.
The configuration unit 13 is mainly configured to configure the power consumption control module of the end-side AI audio chip by using the sleep parameter of the deep sleep target unit, so that the deep sleep target unit enters a low power consumption mode, and records a sleep state sequence.
The computing unit 14 is mainly configured to perform adaptive wake-up computation on the deep sleep target unit according to the sleep state sequence and task scheduling information of the end AI audio chip, so as to obtain a wake-up opportunity and a corresponding energy consumption threshold.
The control unit 15 is mainly configured to perform recovery control on the deep sleep target unit according to the wake-up time and the corresponding energy consumption threshold, so that the deep sleep target unit reenters the task execution state.
In this embodiment, the current energy consumption data and task execution information of all the audio processing units in the terminal AI audio chip are first acquired by the acquiring unit 11, and an energy consumption optimization map is constructed and topologically analyzed to identify the audio processing units in a low-load state. Then, the mapping unit 12 maps the task of the audio processing unit in the low-load state to the high-efficiency audio processing unit according to a preset task migration mapping strategy, generates a task reconstruction sequence, performs dynamic dormancy calculation, and determines a deep dormancy target unit and dormancy parameters thereof. Next, the configuration unit 13 configures the power consumption control module of the end-side AI audio chip with the sleep parameters of the deep sleep target unit, so that these units enter the low power consumption mode and record the sleep state sequence thereof. The computing unit 14 performs adaptive wake-up computation according to the sleep state sequence and task scheduling information of the terminal AI audio chip, and obtains a wake-up opportunity and an energy consumption threshold. Finally, the control unit 15 performs recovery control on the deep sleep target unit according to the calculated wake-up time and the energy consumption threshold value, so that the deep sleep target unit reenters the task execution state. Through the system, accurate energy consumption management of the audio processing unit can be realized, task scheduling is more flexible and efficient, energy consumption performance of the system is optimized, and working efficiency of the terminal side AI audio chip and endurance time of equipment are improved.
It should be noted that, for convenience and brevity of description, a corresponding process in the foregoing embodiment of the deep sleep control method of the end-side AI audio chip may be referred to for a specific working process of the above-described system and each unit, which will not be described herein.
The structures, proportions, sizes, etc. shown in the drawings are shown only in connection with the present disclosure, and are not intended to limit the scope of the invention, since any modification, variation in proportions, or adjustment of the size, etc. of the structures, proportions, etc. should be considered as falling within the spirit and scope of the invention, without affecting the effect or achievement of the objective.
While the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those skilled in the art that the foregoing embodiments may be modified or equivalents may be substituted for some of the features thereof, and that the modifications or substitutions do not depart from the spirit and scope of the embodiments of the invention.

Claims (10)

1.一种端侧AI音频芯片的深度休眠控制方法,其特征在于,包括:1. A deep sleep control method for a terminal-side AI audio chip, comprising: 获取端侧AI音频芯片内所有音频处理单元的当前能耗数据和任务执行信息,基于所述当前能耗数据和任务执行信息,构建能耗优化图谱,对所述能耗优化图谱进行拓扑分析,识别处于低负载状态的音频处理单元;Obtain current energy consumption data and task execution information of all audio processing units in the end-side AI audio chip, build an energy consumption optimization map based on the current energy consumption data and task execution information, perform topological analysis on the energy consumption optimization map, and identify audio processing units in a low-load state; 根据预设的任务迁移映射策略将所述低负载状态的音频处理单元任务映射至高效能音频处理单元,得到任务重构序列,基于所述任务重构序列对所述低负载状态的音频处理单元进行动态休眠计算,得到深度休眠目标单元及其休眠参数;According to a preset task migration mapping strategy, the audio processing unit task in the low-load state is mapped to a high-performance audio processing unit to obtain a task reconstruction sequence, and based on the task reconstruction sequence, a dynamic sleep calculation is performed on the audio processing unit in the low-load state to obtain a deep sleep target unit and its sleep parameters; 利用所述深度休眠目标单元的休眠参数对端侧AI音频芯片的功耗控制模块进行配置,使所述深度休眠目标单元进入低功耗模式,并记录休眠状态序列;The power consumption control module of the end-side AI audio chip is configured using the sleep parameters of the deep sleep target unit to make the deep sleep target unit enter a low power consumption mode and record a sleep state sequence; 根据所述休眠状态序列和端侧AI音频芯片的任务调度信息,对所述深度休眠目标单元进行自适应唤醒计算,得到唤醒时机和对应的能耗阈值;According to the sleep state sequence and the task scheduling information of the end-side AI audio chip, adaptively wake up the deep sleep target unit to obtain a wake-up timing and a corresponding energy consumption threshold; 根据所述唤醒时机和对应的能耗阈值对所述深度休眠目标单元进行恢复控制,使所述深度休眠目标单元重新进入任务执行状态。The deep sleep target unit is restored and controlled according to the wake-up timing and the corresponding energy consumption threshold, so that the deep sleep target unit re-enters the task execution state. 2.根据权利要求1所述的端侧AI音频芯片的深度休眠控制方法,其特征在于,所述获取端侧AI音频芯片内所有的当前能耗数据和任务执行信息,基于所述当前能耗数据和任务执行信息,构建能耗优化图谱,对所述能耗优化图谱进行拓扑分析,识别处于低负载状态的音频处理单元的步骤,包括:2. The deep sleep control method of the end-side AI audio chip according to claim 1 is characterized in that the step of obtaining all current energy consumption data and task execution information in the end-side AI audio chip, constructing an energy consumption optimization map based on the current energy consumption data and task execution information, performing topological analysis on the energy consumption optimization map, and identifying the audio processing unit in a low-load state includes: 通过功率监控模块和时间戳记录模块获取端侧AI音频芯片中各音频处理单元的当前能耗数据和任务执行信息,利用Z-score标准化对所述当前能耗数据和所述任务执行信息进行预处理,得到能耗预处理数据和任务执行特征数据;The current energy consumption data and task execution information of each audio processing unit in the end-side AI audio chip are obtained through the power monitoring module and the timestamp recording module, and the current energy consumption data and the task execution information are preprocessed by using Z-score standardization to obtain energy consumption preprocessing data and task execution feature data; 基于所述能耗预处理数据计算各音频处理单元的能耗负载水平和任务执行强度,得到能耗负载评估数据;Calculating the energy consumption load level and task execution intensity of each audio processing unit based on the energy consumption preprocessing data to obtain energy consumption load evaluation data; 根据所述任务执行特征数据构建音频处理单元之间的任务依赖关系,得到任务拓扑结构数据;Constructing task dependency relationships between audio processing units according to the task execution feature data to obtain task topology structure data; 通过所述任务拓扑结构数据对所述能耗负载评估数据进行优化处理,并基于优化后的能耗负载评估数据构建能耗优化图谱;Optimizing the energy consumption load evaluation data through the task topology data, and constructing an energy consumption optimization map based on the optimized energy consumption load evaluation data; 对所述能耗优化图谱进行拓扑分析,识别出低负载状态的音频处理单元。A topological analysis is performed on the energy consumption optimization map to identify audio processing units in a low-load state. 3.根据权利要求2所述的端侧AI音频芯片的深度休眠控制方法,其特征在于,所述对所述能耗优化图谱进行拓扑分析,识别出低负载状态的音频处理单元的步骤之后,包括:3. The deep sleep control method of the end-side AI audio chip according to claim 2 is characterized in that after the step of performing topological analysis on the energy consumption optimization map to identify the audio processing unit in a low-load state, it includes: 对所述低负载状态的音频处理单元进行解析,得到低负载状态标识数据和低负载状态任务数据;Parsing the audio processing unit in the low-load state to obtain low-load state identification data and low-load state task data; 计算所述低负载状态的音频处理单元在不同任务迁移条件下的能耗变化趋势,得到能耗变化趋势数据;Calculating the energy consumption change trend of the audio processing unit in the low-load state under different task migration conditions to obtain energy consumption change trend data; 根据所述能耗变化趋势数据识别出可迁移任务集合和对应的任务迁移优先级,得到任务迁移优先级数据。A set of migratable tasks and corresponding task migration priorities are identified according to the energy consumption change trend data to obtain task migration priority data. 4.根据权利要求3所述的端侧AI音频芯片的深度休眠控制方法,其特征在于,所述根据预设的任务迁移映射策略将所述低负载状态的音频处理单元任务映射至高效能音频处理单元,得到任务重构序列,基于所述任务重构序列对所述低负载状态的音频处理单元进行动态休眠计算,得到深度休眠目标单元及其休眠参数的步骤,包括:4. The deep sleep control method of the end-side AI audio chip according to claim 3 is characterized in that the step of mapping the audio processing unit task in the low-load state to the high-performance audio processing unit according to a preset task migration mapping strategy to obtain a task reconstruction sequence, and dynamically performing sleep calculation on the audio processing unit in the low-load state based on the task reconstruction sequence to obtain a deep sleep target unit and its sleep parameters includes: 根据所述低负载状态标识数据、低负载状态任务数据和任务迁移优先级数据结合预设的任务迁移映射策略,判断出候选的高效能音频处理单元,基于所述候选的高效能音频处理单元,生成高效能音频处理单元候选集数据;According to the low-load state identification data, the low-load state task data and the task migration priority data combined with a preset task migration mapping strategy, a candidate high-efficiency audio processing unit is determined, and based on the candidate high-efficiency audio processing unit, a candidate set of high-efficiency audio processing unit data is generated; 计算所述高效能音频处理单元候选集数据的任务承载能力和任务匹配度,得到任务承载能力数据和任务执行匹配度数据;Calculating the task carrying capacity and task matching degree of the high-efficiency audio processing unit candidate set data to obtain task carrying capacity data and task execution matching degree data; 根据所述任务执行匹配度数据,选择最优的高效能音频处理单元,并将所述低负载状态任务数据映射至所述最优的高效能音频处理单元,得到任务重构序列数据;According to the task execution matching degree data, an optimal high-efficiency audio processing unit is selected, and the low-load state task data is mapped to the optimal high-efficiency audio processing unit to obtain task reconstruction sequence data; 计算所述任务重构序列数据对所述低负载状态的音频处理单元的影响,得到低负载状态音频处理单元的任务迁移后能耗数据;Calculating the influence of the task reconstruction sequence data on the audio processing unit in the low-load state, and obtaining the energy consumption data of the audio processing unit in the low-load state after task migration; 通过所述任务迁移后能耗数据计算所述低负载状态音频处理单元的最优休眠模式,得到休眠模式数据和休眠时间数据;Calculate the optimal sleep mode of the audio processing unit in the low-load state through the energy consumption data after the task migration, and obtain sleep mode data and sleep time data; 基于所述休眠模式数据和休眠时间数据,确定深度休眠目标单元,并计算出所述深度休眠目标单元对应的休眠参数。Based on the sleep mode data and the sleep time data, a deep sleep target unit is determined, and a sleep parameter corresponding to the deep sleep target unit is calculated. 5.根据权利要求4所述的端侧AI音频芯片的深度休眠控制方法,其特征在于,所述计算所述高效能音频处理单元候选集数据的任务承载能力和任务匹配度,得到任务承载能力数据和任务执行匹配度数据的计算公式如下:5. The deep sleep control method of the end-side AI audio chip according to claim 4 is characterized in that the task carrying capacity and task matching degree of the candidate set data of the high-efficiency audio processing unit are calculated to obtain the calculation formula of the task carrying capacity data and the task execution matching degree data as follows: 其中,为高效能音频处理单元的任务承载能力,为候选高效能音频处理单元的数量,为第个候选音频处理单元的处理能力,为第个候选音频处理单元的负载因子;in, The task-carrying capacity of the high-performance audio processing unit, is the number of candidate high-performance audio processing units, For the The processing capabilities of the candidate audio processing units, For the The load factor of the candidate audio processing units; 其中,为任务执行匹配度,为任务强度系数,为当前任务所需的处理能力。in, For task execution matching, is the task intensity coefficient, The processing power required for the current task. 6.根据权利要求4所述的端侧AI音频芯片的深度休眠控制方法,其特征在于,所述计算所述任务重构序列数据对所述低负载状态的音频处理单元的影响,得到低负载状态音频处理单元的任务迁移后能耗数据的计算公式如下:6. The deep sleep control method of the end-side AI audio chip according to claim 4 is characterized in that the calculation of the influence of the task reconstruction sequence data on the audio processing unit in the low-load state to obtain the calculation formula of the energy consumption data of the audio processing unit in the low-load state after task migration is as follows: 其中,为任务迁移后能耗数据,为任务重构序列中的音频处理单元数量,为第个迁移后音频处理单元的功耗,为第个音频处理单元迁移任务所需的时间。in, Energy consumption data after task migration. The number of audio processing units in the task reconstruction sequence, For the The power consumption of the migrated audio processing unit, For the The time required for the migration task of each audio processing unit. 7.根据权利要求1所述的端侧AI音频芯片的深度休眠控制方法,其特征在于,所述利用所述休眠参数对端侧AI音频芯片的功耗控制模块进行配置,使所述深度休眠目标单元进入低功耗模式,并记录休眠状态序列的步骤,包括:7. The deep sleep control method of the end-side AI audio chip according to claim 1, characterized in that the step of using the sleep parameters to configure the power consumption control module of the end-side AI audio chip to make the deep sleep target unit enter a low power consumption mode and record a sleep state sequence comprises: 将所述休眠参数传输至端侧AI音频芯片的功耗控制模块,并配置功耗控制模块以控制低功耗模式的启动,使所述深度休眠目标单元进入休眠状态;Transmitting the sleep parameters to the power consumption control module of the AI audio chip on the terminal side, and configuring the power consumption control module to control the start of the low power consumption mode so that the deep sleep target unit enters a sleep state; 在所述低功耗模式下,记录各音频处理单元的休眠状态变化序列,得到休眠状态序列。In the low power consumption mode, the sleep state change sequence of each audio processing unit is recorded to obtain a sleep state sequence. 8.根据权利要求1所述的端侧AI音频芯片的深度休眠控制方法,其特征在于,所述根据所述休眠状态序列和端侧AI音频芯片的任务调度信息,对所述深度休眠目标单元进行自适应唤醒计算,得到唤醒时机和对应的能耗阈值的步骤,包括:8. The deep sleep control method of the end-side AI audio chip according to claim 1 is characterized in that the step of performing adaptive wake-up calculation on the deep sleep target unit according to the sleep state sequence and the task scheduling information of the end-side AI audio chip to obtain the wake-up timing and the corresponding energy consumption threshold comprises: 获取端侧AI音频芯片的任务调度信息,建立所述任务调度信息与所述休眠状态序列之间的映射关系,并利用决策树算法对所述映射关系进行分析,确定任务调度与休眠时机之间的关联性,得到关联结果;Acquire task scheduling information of the end-side AI audio chip, establish a mapping relationship between the task scheduling information and the sleep state sequence, and use a decision tree algorithm to analyze the mapping relationship to determine the correlation between task scheduling and sleep timing, and obtain a correlation result; 通过所述任务调度信息预测下一周期内的任务负载信息,并根据所述任务负载信息和所述关联结果,确定所述深度休眠目标单元的唤醒需求;Predicting the task load information in the next cycle through the task scheduling information, and determining the wake-up requirement of the deep sleep target unit according to the task load information and the association result; 基于所述唤醒需求计算出各任务的唤醒时机,并根据所述唤醒时机计算所述深度休眠目标单元的能耗阈值。The wake-up timing of each task is calculated based on the wake-up requirement, and the energy consumption threshold of the deep sleep target unit is calculated according to the wake-up timing. 9.根据权利要求1所述的端侧AI音频芯片的深度休眠控制方法,其特征在于,所述根据所述唤醒时机和对应的能耗阈值对所述深度休眠目标单元进行恢复控制,使所述深度休眠目标单元重新进入任务执行状态的步骤,包括:9. The deep sleep control method of the end-side AI audio chip according to claim 1, characterized in that the step of restoring and controlling the deep sleep target unit according to the wake-up timing and the corresponding energy consumption threshold so that the deep sleep target unit re-enters the task execution state comprises: 基于所述唤醒时机和对应的能耗阈值生成唤醒控制指令,通过功耗控制模块向深度休眠目标单元发送所述唤醒控制指令,启动恢复过程;Generate a wake-up control instruction based on the wake-up timing and the corresponding energy consumption threshold, and send the wake-up control instruction to the deep sleep target unit through the power consumption control module to start the recovery process; 判断当前功耗是否满足预设的能耗阈值,若不满足,则延迟唤醒时机;Determine whether the current power consumption meets the preset energy consumption threshold. If not, delay the wake-up time. 若满足预设能耗阈值,则启动深度休眠目标单元的唤醒程序,恢复所述深度休眠目标单元的任务执行能力。If the preset energy consumption threshold is met, a wake-up procedure of the deep sleep target unit is started to restore the task execution capability of the deep sleep target unit. 10.一种端侧AI音频芯片的深度休眠控制系统,其特征在于,包括:10. A deep sleep control system for a terminal-side AI audio chip, characterized by comprising: 获取单元,用于获取端侧AI音频芯片内所有音频处理单元的当前能耗数据和任务执行信息,基于所述当前能耗数据和任务执行信息,构建能耗优化图谱,对所述能耗优化图谱进行拓扑分析,识别处于低负载状态的音频处理单元;An acquisition unit, used to acquire current energy consumption data and task execution information of all audio processing units in the end-side AI audio chip, construct an energy consumption optimization map based on the current energy consumption data and task execution information, perform topological analysis on the energy consumption optimization map, and identify audio processing units in a low-load state; 映射单元,用于根据预设的任务迁移映射策略将所述低负载状态的音频处理单元任务映射至高效能音频处理单元,得到任务重构序列,基于所述任务重构序列对所述低负载状态的音频处理单元进行动态休眠计算,得到深度休眠目标单元及其休眠参数;A mapping unit, configured to map the audio processing unit task in the low-load state to a high-performance audio processing unit according to a preset task migration mapping strategy, obtain a task reconstruction sequence, and perform dynamic sleep calculation on the audio processing unit in the low-load state based on the task reconstruction sequence to obtain a deep sleep target unit and its sleep parameters; 配置单元,用于利用所述深度休眠目标单元的休眠参数对端侧AI音频芯片的功耗控制模块进行配置,使所述深度休眠目标单元进入低功耗模式,并记录休眠状态序列;A configuration unit, configured to configure a power consumption control module of the end-side AI audio chip using the sleep parameters of the deep sleep target unit, so that the deep sleep target unit enters a low power consumption mode, and records a sleep state sequence; 计算单元,用于根据所述休眠状态序列和端侧AI音频芯片的任务调度信息,对所述深度休眠目标单元进行自适应唤醒计算,得到唤醒时机和对应的能耗阈值;A calculation unit, configured to perform adaptive wake-up calculation on the deep sleep target unit according to the sleep state sequence and task scheduling information of the end-side AI audio chip, and obtain a wake-up timing and a corresponding energy consumption threshold; 控制单元,用于根据所述唤醒时机和对应的能耗阈值对所述深度休眠目标单元进行恢复控制,使所述深度休眠目标单元重新进入任务执行状态。The control unit is used to perform recovery control on the deep sleep target unit according to the wake-up timing and the corresponding energy consumption threshold, so that the deep sleep target unit re-enters the task execution state.
CN202510294706.1A 2025-03-13 2025-03-13 A method and system for deep sleep control of an edge AI audio chip Active CN119806848B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202510294706.1A CN119806848B (en) 2025-03-13 2025-03-13 A method and system for deep sleep control of an edge AI audio chip

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202510294706.1A CN119806848B (en) 2025-03-13 2025-03-13 A method and system for deep sleep control of an edge AI audio chip

Publications (2)

Publication Number Publication Date
CN119806848A true CN119806848A (en) 2025-04-11
CN119806848B CN119806848B (en) 2025-11-25

Family

ID=95274858

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202510294706.1A Active CN119806848B (en) 2025-03-13 2025-03-13 A method and system for deep sleep control of an edge AI audio chip

Country Status (1)

Country Link
CN (1) CN119806848B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120084789A1 (en) * 2010-09-30 2012-04-05 Francesco Iorio System and Method for Optimizing the Evaluation of Task Dependency Graphs
US20180152505A1 (en) * 2016-11-30 2018-05-31 Microsoft Technology Licensing, Llc Data migration reservation system and method
CN112380005A (en) * 2020-11-10 2021-02-19 深圳供电局有限公司 Data center energy consumption management method and system
CN116189695A (en) * 2023-02-22 2023-05-30 武汉光华芯科技有限公司 Audio encoding and decoding chip and power consumption management mode switching method thereof
CN117519980A (en) * 2023-11-22 2024-02-06 联通(广东)产业互联网有限公司 Energy-saving data center

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120084789A1 (en) * 2010-09-30 2012-04-05 Francesco Iorio System and Method for Optimizing the Evaluation of Task Dependency Graphs
US20180152505A1 (en) * 2016-11-30 2018-05-31 Microsoft Technology Licensing, Llc Data migration reservation system and method
CN112380005A (en) * 2020-11-10 2021-02-19 深圳供电局有限公司 Data center energy consumption management method and system
CN116189695A (en) * 2023-02-22 2023-05-30 武汉光华芯科技有限公司 Audio encoding and decoding chip and power consumption management mode switching method thereof
CN117519980A (en) * 2023-11-22 2024-02-06 联通(广东)产业互联网有限公司 Energy-saving data center

Also Published As

Publication number Publication date
CN119806848B (en) 2025-11-25

Similar Documents

Publication Publication Date Title
CN120065746B (en) Building energy consumption optimization method based on artificial intelligence
CN117850494A (en) Data center energy-saving control method based on AI algorithm
CN110070181A (en) A kind of optimization method of the deep learning for edge calculations equipment
CN112114648B (en) Wearable device power management method, system and computer equipment
Pughat et al. Performance analysis of an improved dynamic power management model in wireless sensor node
CN112835715B (en) Method and device for determining unmanned aerial vehicle task offloading strategy based on reinforcement learning
CN119356836B (en) Automatic resource allocation method for bearer neural network
Zhou et al. Deadline-aware deep-recurrent-q-network governor for smart energy saving
CN119201622B (en) Adaptive operator parallel partitioning method for heterogeneous embedded chips in intelligent Internet of Things
CN119759198A (en) Power consumption management method and system for IoT devices based on big data
CN117891681A (en) A method and system for dynamic configuration of storage devices based on energy consumption prediction
CN120560939A (en) Server power consumption management method, device, electronic device and storage medium
CN118586291A (en) A method for optimizing processor performance under temperature constraints based on deep reinforcement learning
CN119806848B (en) A method and system for deep sleep control of an edge AI audio chip
CN113783707B (en) Sensor power supply method and device based on intelligent gateway and computer equipment
CN119136288B (en) A dual-band low-power Wi-Fi keep-alive wake-up method and system for the Internet of Things
KR102130509B1 (en) Method for determining optimum transmission interval of sensor data and apparatus thereof
CN119417177A (en) Power plant unit load distribution system and method
CN120065802A (en) Power consumption control method and device of power distribution network monitoring control chip and storage medium
Heo et al. Predictive energy-aware adaptive sampling with deep reinforcement learning
CN118092817A (en) Tablet computer space intelligent management method and system
CN116933628A (en) Enterprise electricity load monitoring method and system
Heydary et al. Optimization framework and clustering-based algorithm for energy-aware adaptive sensing
CN119903928B (en) A client-based AI model dynamic optimization and adaptive reasoning method
CN119233376B (en) A high-energy-efficiency sensor scheduling method and related equipment for collaborative sensing

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant