WO2020238121A1 - 控制峰值功耗的方法及系统 - Google Patents

控制峰值功耗的方法及系统 Download PDF

Info

Publication number
WO2020238121A1
WO2020238121A1 PCT/CN2019/122382 CN2019122382W WO2020238121A1 WO 2020238121 A1 WO2020238121 A1 WO 2020238121A1 CN 2019122382 W CN2019122382 W CN 2019122382W WO 2020238121 A1 WO2020238121 A1 WO 2020238121A1
Authority
WO
WIPO (PCT)
Prior art keywords
power consumption
load
module
controlled
peak power
Prior art date
Application number
PCT/CN2019/122382
Other languages
English (en)
French (fr)
Inventor
张慧明
Original Assignee
芯原微电子(上海)股份有限公司
芯原控股有限公司
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 芯原微电子(上海)股份有限公司, 芯原控股有限公司 filed Critical 芯原微电子(上海)股份有限公司
Priority to KR1020217030998A priority Critical patent/KR20210134355A/ko
Priority to EP19930416.3A priority patent/EP3816769A4/en
Priority to JP2021570476A priority patent/JP7304972B2/ja
Priority to US17/263,901 priority patent/US20220342474A1/en
Publication of WO2020238121A1 publication Critical patent/WO2020238121A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F1/00—Details not covered by groups G06F3/00 - G06F13/00 and G06F21/00
    • G06F1/26—Power supply means, e.g. regulation thereof
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F1/00—Details not covered by groups G06F3/00 - G06F13/00 and G06F21/00
    • G06F1/26—Power supply means, e.g. regulation thereof
    • G06F1/32—Means for saving power
    • G06F1/3203—Power management, i.e. event-based initiation of a power-saving mode
    • G06F1/3234—Power saving characterised by the action undertaken
    • G06F1/324—Power saving characterised by the action undertaken by lowering clock frequency
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F1/00—Details not covered by groups G06F3/00 - G06F13/00 and G06F21/00
    • G06F1/26—Power supply means, e.g. regulation thereof
    • G06F1/32—Means for saving power
    • G06F1/3203—Power management, i.e. event-based initiation of a power-saving mode
    • G06F1/3206—Monitoring of events, devices or parameters that trigger a change in power modality
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F1/00—Details not covered by groups G06F3/00 - G06F13/00 and G06F21/00
    • G06F1/26—Power supply means, e.g. regulation thereof
    • G06F1/32—Means for saving power
    • G06F1/3203—Power management, i.e. event-based initiation of a power-saving mode
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/30—Monitoring
    • G06F11/34—Recording or statistical evaluation of computer activity, e.g. of down time, of input/output operation ; Recording or statistical evaluation of user activity, e.g. usability assessment
    • G06F11/3409—Recording or statistical evaluation of computer activity, e.g. of down time, of input/output operation ; Recording or statistical evaluation of user activity, e.g. usability assessment for performance assessment
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
    • G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
    • G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
    • G06F9/505—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals considering the load
    • Y—GENERAL 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
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
    • Y—GENERAL 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
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/50—Reducing energy consumption in communication networks in wire-line communication networks, e.g. low power modes or reduced link rate

Definitions

  • the invention relates to power consumption control, in particular to a method and system for controlling peak power consumption.
  • AI Artificial Intelligence
  • IoT Internet of Things
  • the traditional power control method uses frequency modulation and voltage regulation algorithms, one of which is static frequency modulation and voltage regulation.
  • the central processing unit changes the value of the voltage and frequency control registers to achieve the purpose of controlling the frequency and voltage.
  • the central processing unit has a large load, which is mainly used to process the process control and data synchronization of the convolutional network.
  • There is a large delay in the control of voltage and frequency and the control effect has a large lag.
  • switching between tasks completed through the interrupt mechanism generally takes 1 to 2 milliseconds. Time delay will cause frequency and voltage control lag, if it is adjusted from a higher frequency to a lower frequency, adjusting the lag will cause excessive power consumption.
  • the central processing unit If it is adjusted from a lower frequency to a higher frequency, the adjustment lag will lead to performance degradation, a phenomenon of stuck in the application, and the balance point of performance and power consumption cannot be achieved. In addition, controlling the frequency and voltage by the central processing unit will cause the central processing unit to degrade the performance of the main processing tasks during frequent task switching; the central processing unit will also consume more power consumption.
  • Another frequency and voltage regulation algorithm uses a dedicated microprocessor to control the frequency and voltage.
  • This method is now one of the commonly used methods in AI chip systems.
  • the dedicated microprocessor only serves the process of power consumption and voltage control, and the response time to the control register can be increased by 5 to 10 times.
  • this method requires additional microprocessor hardware resources, which is a huge waste of chip area.
  • the commonly used microprocessor has an area of 0.5 to 1 square millimeter under the 28nm process. This is a large chip area and cost overhead.
  • the microprocessor is always working, which is also a great power consumption overhead.
  • the purpose of the present invention is to provide a method and system for controlling peak power consumption, which is used to solve the problem that power consumption control in the prior art cannot take into account area, cost, and power consumption. .
  • the present invention provides a method for controlling peak power consumption.
  • the method for controlling peak power consumption at least includes:
  • the load of each monitoring window in the current power consumption statistical period is less than the load threshold and is greater than the load lower limit, if the current power consumption load is in the normal power consumption state, the original power consumption state will be maintained, and the frequency level will not change; If the peak power consumption state is lower than the normal power consumption state, the module to be controlled enters the previous peak power consumption state, the frequency level is increased by one level, and the module to be controlled is controlled based on the frequency level corresponding to the previous peak power consumption state Increase the load of the module to be controlled;
  • the load of the module to be controlled is the sum of the product of the load weight of each load unit and the actual load.
  • the method for controlling peak power consumption includes n power consumption states, where the load of the module to be controlled increases by 1/n times the load threshold when the power consumption state is increased by one level, and n is a positive integer.
  • each power consumption statistical period includes k monitoring windows, the first monitoring window is removed from the current power consumption statistical period and the k+1 monitoring window is added to obtain the next power consumption statistical period, where k is A positive integer not less than 3.
  • the method for controlling peak power consumption further includes: when the load of each monitoring window in the current power consumption statistical period is lower than the load lower limit, the frequency of the working clock is at the frequency corresponding to the current power consumption state. One frequency band is dropped in the level.
  • the frequency of the working clock is reduced to a minimum frequency.
  • the method for controlling peak power consumption is used for power consumption control of the AI chip.
  • the present invention also provides a system for controlling peak power consumption. Based on the above method for controlling peak power consumption, the system for controlling peak power consumption at least includes:
  • the module to be controlled includes multiple load units
  • a global control module connected to the module to be controlled and the load information statistics module, provides the module to be controlled with control signals and load weight information of each load unit, and obtains a load threshold based on the working state of the module to be controlled;
  • the load information statistics module is connected to the global control module and the module to be controlled, monitors the load information of each load unit in the module to be controlled, and calculates the load information based on the load weight information of each load unit in the global control module State the load of the module to be controlled;
  • a frequency and voltage regulation control module connected to the load information statistics module, and adjust the power consumption state of the module to be controlled based on the load of the module to be controlled;
  • the clock control module is connected between the frequency modulation and voltage regulation control module and the module to be controlled, generates a clock signal of corresponding frequency based on the output signal of the frequency modulation and voltage regulation control module, and controls the module to be controlled based on the clock The signal works.
  • the system for controlling peak power consumption is applied to an AI chip.
  • the module to be controlled includes a programmable arithmetic unit, a neural network accelerator or a tensor processor.
  • the frequency and voltage regulation control module includes a comparison unit and a plurality of power consumption state control units; the comparison unit compares the load of the module to be controlled in each power consumption statistical period with the load threshold; each The power consumption state control unit controls the module to be controlled to enter a corresponding power consumption state based on the comparison result output by the comparison unit.
  • the clock control module includes a digital frequency division unit.
  • the method and system for controlling peak power consumption of the present invention have the following beneficial effects:
  • the method and system for controlling peak power consumption of the present invention monitor the load situation in real time, reduce power consumption by dynamically and intelligently controlling the working frequency, and reach the balance point of performance and power consumption, so that the peak power consumption does not exceed the load threshold and works at the highest level. Frequency, effectively improve work efficiency under the premise of realizing power consumption control.
  • Fig. 1 shows a schematic flow chart of the method for controlling peak power consumption of the present invention.
  • Figure 2 shows a schematic diagram of the structure of the system for controlling peak power consumption of the present invention.
  • FIG. 3 shows a schematic diagram of the frequency division principle of the clock control module of the present invention.
  • the present invention provides a method for controlling peak power consumption.
  • the method for controlling peak power consumption includes:
  • the module to be controlled is the computing unit of the AI chip.
  • any module that requires power consumption control is applicable to the module to be controlled in the present invention. Examples are limited.
  • the load condition of the module to be controlled in the normal working state is analyzed and the load threshold is set.
  • the working state of the module to be controlled in different application scenarios is different, and the load threshold is also different, It can be set as required.
  • the load of the module to be controlled is the actual load of the module to be controlled; when the module to be controlled is multiple (2 or more) load units
  • the load of the module to be controlled is the sum of the product of the load weight of each load unit and the actual load.
  • each power consumption statistical period includes k monitoring windows
  • the load of the module 11 to be controlled in each monitoring window is compared with the load threshold, and the power consumption statistical period is taken as a whole to obtain a comparison result.
  • the monitoring window of each power consumption statistical period adopts weighted control, that is, the first monitoring window is removed in the current power consumption statistical period and the k+1 monitoring window is added to obtain each monitoring window (in time In order to order, the earliest monitoring window is removed and added to the latest monitoring window).
  • k is set to 4 to improve the accuracy of monitoring (generally setting 3 monitoring windows can improve the accuracy of monitoring) .
  • the number of monitoring windows in each power consumption statistical period can be set as required. Under the premise that accuracy is not required or can be ensured, the number of monitoring windows in each power consumption statistical period can be appropriately reduced. It is not limited to this embodiment.
  • the module to be controlled enters the lowest peak power consumption state, the frequency level is reduced to the lowest level, and is based on the frequency corresponding to the lowest peak power consumption state Level control of the working clock of the module to be controlled.
  • n is set to 4, which are arranged as the first peak power consumption state, the second peak power consumption state, and the third peak power consumption state in descending order of frequency.
  • the normal power consumption state in this embodiment, the frequency levels corresponding to the first peak power consumption state, the second peak power consumption state, and the third peak power consumption state are the frequency division signals corresponding to the frequency of the normal power consumption state.
  • each monitoring window in the current power consumption statistical period is greater than the load threshold, regardless of the current power consumption state of the module to be controlled, it is adjusted to the first peak power consumption state, and is based on the first peak power consumption state.
  • the frequency level corresponding to the power consumption state controls the working clock of the module to be controlled.
  • the normal power consumption state remains unchanged when the current power consumption load is in the normal power consumption state.
  • the current power consumption load state is in the first peak power consumption state, the second peak power consumption state, or In the third peak power consumption state, the module to be controlled is adjusted to the previous power consumption state, the corresponding frequency level is increased by one level, and the load of the module to be controlled increases correspondingly; that is, it is currently in the first peak power consumption state Then it rises to the second peak power consumption state, the current second peak power consumption state rises to the third peak power consumption state, and the current third peak power consumption state rises to the normal power consumption state.
  • the power consumption state is increased by one level, and the load of the module to be controlled is increased by 1/n times of the load threshold.
  • the load threshold is increased by 1/4.
  • the method for controlling peak power consumption further includes: when the load of each monitoring window in the current power consumption statistical period is lower than the load lower limit, the frequency of the working clock is at the current The frequency level corresponding to the power consumption state drops by one frequency band.
  • the load of the module to be controlled is relatively low and does not need to work in a full frequency state.
  • the frequency of the module to be controlled drops to the next frequency band, that is, assuming that the current frequency is 12/16 of the original frequency in the normal working state, it drops To 11/16 of the original frequency in the normal working state, this embodiment supports a minimum frequency reduction of one sixteenth, and the lowest frequency reduction frequency can be set according to actual use. If the load of the next power consumption statistical period is not lower than the lower limit of the load, the working frequency of the module to be controlled is increased by a frequency level in the current period to ensure that the module to be controlled has sufficient performance.
  • the clock frequency can be automatically reduced to the lowest frequency.
  • the module to be controlled starts to work in a normal power consumption state; the power consumption state is updated according to the load in a power consumption statistical period.
  • the module to be controlled If the load of the module to be controlled is greater than the load threshold in a power consumption statistical period, the module to be controlled enters the first peak power consumption state, that is, enters the first peak power consumption state from the normal power consumption state, The frequency is reduced by three levels; the power consumption status is continuously updated according to the load in the next power consumption statistical period.
  • step S6 Continue to loop from step S2 to step S5 to ensure that the peak power consumption does not exceed the load threshold at the highest frequency at which the module to be controlled works.
  • the invention reduces power consumption by dynamically controlling the voltage and frequency of the AI chip, while providing sufficient computing performance.
  • the present invention provides a system 1 for controlling peak power consumption.
  • the system 1 for controlling peak power consumption is used to implement the method for controlling peak power consumption.
  • the system 1 for controlling peak power consumption includes :
  • the module to be controlled 11 the global control module 12, the load information statistics module 13, the frequency and voltage regulation control module 14 and the clock control module 15.
  • the module to be controlled 11 includes multiple load units.
  • the module to be controlled 11 is set as a computing unit of the AI chip, and each load unit includes but is not limited to a programmable computing unit, a neural network accelerator, and a tensor processor.
  • the global control module 12 is connected to the module to be controlled 11 and the load information statistics module 13, to provide the module to be controlled 11 with control signals and load weight information of each load unit, and based on all The working state of the module 11 to be controlled obtains the load threshold.
  • the global control module 12 is based on the control information provided by the central processing unit (not shown in the figure) and the driver (not shown in the figure) (each control information is used to determine the module to be controlled 11 Working state) to obtain the load threshold, load weight information and mode control information of each load unit, and send the load weight information and mode control information of each load unit to the module to be controlled 11; at the same time, the global control module 12
  • the load statistical information of the module 11 to be controlled is fed back to the driver, so that the driver can adjust the control strategy according to the internal conditions of the AI chip.
  • the load information statistics module 13 is connected to the global control module 12 and the to-be-controlled module 11, monitors the load information of each load unit in the to-be-controlled module 11, and is based on the global control module
  • the load weight information of each load unit in 12 calculates the load of the module 11 to be controlled.
  • the load information statistics module 13 monitors the load information of each main assembly line (each load unit) of the module 11 to be controlled, and collects statistics on the load weight information of each load unit provided by the global control module 12 Contributed load, according to the importance of each load unit’s impact on the overall performance and the power consumption of the load unit itself, according to the load weight of the upper level load unit, the load information of all load units is integrated to obtain the entire module to be controlled 11 load.
  • the main pipeline of the calculation unit of the AI chip includes a programmable arithmetic unit, a neural network accelerator, and a tensor processor.
  • the neural network accelerator is used to process convolution operations in the neural network, and the tensor processor is used to process
  • the programmable arithmetic unit performs process control and auxiliary calculations; from the perspective of performance, the neural network accelerator is the most important, 80% of the operations are completed by the neural network accelerator, and the other 15% are processed by the tensor
  • the other 5% control and auxiliary calculations are completed by the programmable arithmetic unit, so the load information statistics module 13 has load coefficients of 80%, 15% and 5% for these three main modules.
  • the load of the module 11 to be controlled is 80%*load1+ 15%*load2+5%*load3.
  • the frequency and voltage regulation control module 14 is connected to the load information statistics module 13 to adjust the power consumption state of the module to be controlled 11 based on the load of the module to be controlled 11.
  • the frequency and voltage regulation control module 14 includes a comparison unit 141 and a plurality of power consumption state control units. In this embodiment, it includes three power consumption state control units, which are respectively The first power consumption state control unit 142, the second power consumption state control unit 143, and the third power consumption state control unit 144.
  • the comparison unit 141 compares the load of the module 11 to be controlled in each power consumption statistical period with the load threshold. In this embodiment, four monitoring windows are set in the power consumption statistical period.
  • Each power consumption state control unit controls the module to be controlled 11 to enter a corresponding power consumption state based on the comparison result output by the comparison unit 141.
  • the frequency and voltage regulation control module 14 can also analyze the current minimum load condition to implement frequency reduction operations, which will not be repeated here.
  • the clock control module 15 is connected between the frequency and voltage regulation control module 14 and the module to be controlled 11, and generates a clock signal of corresponding frequency based on the signal output by the frequency and voltage regulation control module 14. , And control the module to be controlled 11 to work based on the clock signal.
  • the clock control module 15 includes a digital frequency dividing unit for dividing the original frequency by 16 in a normal working state to obtain the frequency and frequency corresponding to the corresponding frequency level. Frequency reduction frequency.
  • the frequency levels corresponding to the first peak power state control unit 142, the second peak power state control unit 143, and the third peak power state control unit 144 are in order Increase, respectively, 1/4, 1/2 and 3/4 of the original frequency under normal working conditions.
  • the corresponding frequency of each frequency level control unit can be set according to needs.
  • the clock control module 15 may adopt equal phase frequency division, which is not limited to this embodiment.
  • the method and system for controlling peak power consumption of the present invention improve the precise control of dynamic power consumption in the AI chip. Without affecting the performance of the chip, the frequency of the chip can be dynamically controlled to achieve the effect of reducing peak power consumption; The CPU occupancy rate and area overhead in traditional power control methods are too large.
  • the method and system for controlling peak power consumption of the present invention perform automatic frequency control by real-time monitoring of the load conditions of each module of the AI chip; it can also perform frequency control and voltage control through a combination of external software active control and internal hardware automatic control. Regulation.
  • the present invention provides a method and system for controlling peak power consumption, including: monitoring the load of the module to be controlled in the current power consumption statistical period, when the load of each monitoring window in the current power consumption statistical period is greater than the load threshold
  • the module to be controlled enters the lowest peak power consumption state, and controls the working clock of the module to be controlled based on the frequency level corresponding to the lowest peak power consumption state; the load of each monitoring window in the current power consumption statistical period is less than the load
  • the threshold value is greater than the lower limit of the load, if the current power consumption load is in the normal power consumption state, the original power consumption state is maintained, and the frequency level remains unchanged; if the current peak power consumption state is lower than the normal power consumption state, the module to be controlled Enter the previous peak power consumption state, the frequency level is increased by one level, and based on the frequency level corresponding to the previous peak power consumption state, the working clock of the module to be controlled is controlled to increase the load of the module to be controlled; When the load unevenness of each monitoring window in the consumption statistical period is less
  • the method and system for controlling peak power consumption of the present invention monitor the load situation in real time, reduce power consumption by dynamically and intelligently controlling the working frequency, and reach the balance point of performance and power consumption, so that the peak power consumption does not exceed the load threshold to work at the highest Frequency, effectively improve work efficiency under the premise of realizing power consumption control. Therefore, the present invention effectively overcomes various shortcomings in the prior art and has high industrial value.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Power Sources (AREA)

Abstract

本发明提供一种控制峰值功耗的方法及系统，通过实时监控负载情况来动态控制频率，进而降低功耗，在控制峰值功耗的同时提供足够的计算性能。本发明的控制峰值功耗的方法及系统实时监控负载情况，通过动态智能控制工作频率来降低功耗，达到性能和功耗的平衡点，使得峰值功耗不超出负载阈值的情况下工作在最高频率，在实现功耗控制的前提下有效提高工作效率。

Description

控制峰值功耗的方法及系统 技术领域
本发明涉及功耗控制，特别是涉及一种控制峰值功耗的方法及系统。
背景技术
随着人工智能(Artificial Intelligence，AI)以及物联网(Internet of Things，IoT)技术的蓬勃发展，对AI运算能力的要求越来越大，随着算力的加大，功耗控制变成了AI芯片至关重要的一个组成部分。如何在提供强大计算能力的同时，将功耗控制在电源设计阈值范围以内，已经成为芯片设计的必须解决的问题之一，平衡芯片计算能力和动态功耗也是芯片竞争力的集中表现。
传统的功耗控制方法采用调频调压算法，其中一种调频调压算法是静态调频和调压的方式，通过中央处理器改变电压和频率的控制寄存器的值来达到控制频率的电压的目的。在AI芯片中，中央处理器的负载较大，主要用于处理卷积网络的流程控制和数据同步。对电压和频率的控制存在很大延迟，控制效果具有很大的滞后性。在单核处理器情况下，通过中断机制所完成的任务之间的切换一般需要1～2毫秒。时间延迟会造成频率和电压控制滞后，如果是从较高频率向较低频率调整，调整滞后会导致过多的功耗消耗。如果是从较低频率向较高频率调整，调整滞后会导致性能下降，在应用中会出现卡顿的现象，不能达到性能和功耗的平衡点。另外，通过中央处理器来控制频率和电压，会导致中央处理器在频繁的任务切换中主要处理任务的性能下降；中央处理器也会消耗更多功耗。
另一种调频调压算法是利用专用的微处理器来控制频率和电压，这种方式是现在AI芯片系统中常用的方法之一。专用的微处理器只服务于功耗和电压控制的进程，对控制寄存器的响应时间可以提高5～10倍。但是这种方法需要额外的微处理器硬件资源，是芯片面积的巨大浪费，一般普遍采用的微处理器，在28nm工艺下，面积也在0.5～1平方毫米。这是很大的芯片面积和成本开销。微处理器处于永远工作状态，也是很大的功耗开销。
因此，如何在不增加面积开销、成本开销及功耗开销的情况下实现功耗控制，已成为本领域技术人员亟待解决的问题之一。
发明内容
鉴于以上所述现有技术的缺点，本发明的目的在于提供一种控制峰值功耗的方法及系统，用于解决现有技术中功耗控制不能兼顾面积开销、成本开销及功耗开销的问题。
为实现上述目的及其他相关目的，本发明提供一种控制峰值功耗的方法，所述控制峰值功耗的方法至少包括：
监测当前功耗统计周期内待控制模块的负载，当前功耗统计周期内各监测窗口的负载均大于负载阈值时，所述待控制模块进入最低峰值功耗状态，并基于最低峰值功耗状态对应的频率等级控制所述待控制模块的工作时钟；
当前功耗统计周期内各监测窗口的负载均小于所述负载阈值，且均大于负载下限时，若当前功耗负载处于正常功耗状态则保持原有功耗状态，频率等级不变；若当前峰值功耗状态低于正常功耗状态则所述待控制模块进入上一级峰值功耗状态，频率等级上升一级，并基于上一级峰值功耗状态对应的频率等级控制所述待控制模块的工作时钟，增大所述待控制模块的负载；
当前功耗统计周期内各监测窗口的负载不均小于所述负载阈值或不均大于所述负载阈值时，保持原有功耗状态，频率等级不变，所述待控制模块的负载不变；
继续监测后续各功耗统计周期内所述待控制模块的负载，并在各功耗统计周期内基于负载状况调整所述待控制模块的工作时钟的频率，使得所述待控制模块的峰值功耗不超出负载阈值的情况下工作在最高频率。
可选地，所述待控制模块包括多个负载单元时，所述待控制模块的负载为各负载单元的负载权重与实际负载乘积的和。
可选地，所述控制峰值功耗的方法包括n个功耗状态，功耗状态上升一级所述待控制模块的负载增大所述负载阈值的1/n倍，n为正整数。
可选地，各功耗统计周期均包括k个监测窗口，在当前功耗统计周期中去除第1个监测窗口并加入第k+1个监测窗口以得到下一功耗统计周期，其中k为不小于3的正整数。
更可选地，所述控制峰值功耗的方法还包括：当前功耗统计周期内各监测窗口的负载均低于所述负载下限时，所述工作时钟的频率在当前功耗状态对应的频率等级中下降一个频段。
更可选地，当所述待控制模块进入空闲状态时，所述工作时钟的频率降低至最低频率。
可选地，所述控制峰值功耗的方法用于AI芯片的功耗控制。
为实现上述目的及其他相关目的，本发明还提供一种控制峰值功耗的系统，基于上述控制峰值功耗的方法，所述控制峰值功耗的系统至少包括：
待控制模块，包括多个负载单元；
全局控制模块，连接所述待控制模块及所述负载信息统计模块，为所述待控制模块提供 控制信号及各负载单元的负载权重信息，并基于所述待控制模块的工作状态获取负载阈值；
负载信息统计模块，连接所述全局控制模块及所述待控制模块，监控所述待控制模块中各负载单元的负载信息，并基于所述全局控制模块中各负载单元的负载权重信息，计算所述待控制模块的负载；
调频调压控制模块，连接所述负载信息统计模块，基于所述待控制模块的负载调整所述待控制模块的功耗状态；
时钟控制模块，连接于所述调频调压控制模块与所述待控制模块之间，基于所述调频调压控制模块输出信号产生相应频率的时钟信号，并控制所述待控制模块基于所述时钟信号工作。
可选地，所述控制峰值功耗的系统应用于AI芯片中。
更可选地，所述待控制模块包括可编程运算单元，神经网络加速器或张量处理器。
可选地，所述调频调压控制模块包括比较单元及多个功耗状态控制单元；所述比较单元将各功耗统计周期内所述待控制模块的负载与所述负载阈值进行比较；各功耗状态控制单元基于所述比较单元输出的比较结果控制所述待控制模块进入相应的功耗状态。
可选地，所述时钟控制模块包括数字分频单元。
如上所述，本发明的控制峰值功耗的方法及系统，具有以下有益效果：
本发明的控制峰值功耗的方法及系统实时监控负载情况，通过动态智能控制工作频率来降低功耗，达到性能和功耗的平衡点，使得峰值功耗不超出负载阈值的情况下工作在最高频率，在实现功耗控制的前提下有效提高工作效率。
附图说明
图1显示为本发明的控制峰值功耗的方法流程示意图。
图2显示为本发明的控制峰值功耗的系统的结构示意图。
图3显示为本发明的时钟控制模块的分频原理示意图。
元件标号说明
1                      控制峰值功耗的系统
11                     待控制模块
12                     全局控制模块
13                     负载信息统计模块
14                     调频调压控制模块
141                    比较单元
142～144               第一～第三功耗状态控制单元
15                     时钟控制模块
具体实施方式
以下通过特定的具体实例说明本发明的实施方式，本领域技术人员可由本说明书所揭露的内容轻易地了解本发明的其他优点与功效。本发明还可以通过另外不同的具体实施方式加以实施或应用，本说明书中的各项细节也可以基于不同观点与应用，在没有背离本发明的精神下进行各种修饰或改变。
请参阅图1～图3。需要说明的是，本实施例中所提供的图示仅以示意方式说明本发明的基本构想，遂图式中仅显示与本发明中有关的组件而非按照实际实施时的组件数目、形状及尺寸绘制，其实际实施时各组件的型态、数量及比例可为一种随意的改变，且其组件布局型态也可能更为复杂。
如图1所示，本发明提供一种控制峰值功耗的方法，所述控制峰值功耗的方法包括：
1)监测当前功耗统计周期内待控制模块的负载，并将当前功耗统计周期内的负载与负载阈值进行比较。
具体地，在本实施例中，所述待控制模块为AI芯片的计算单元，在实际使用中，任意需要进行功耗控制的模块均适用于本发明的所述待控制模块，不以本实施例为限。
具体地，根据不同应用场景分析得到正常工作状态下所述待控制模块的负载情况，并设定所述负载阈值，不同应用场景下所述待控制模块的工作状态不同，负载阈值也不相同，可根据需要进行设定。
具体地，当所述待控制模块为单个负载单元时，所述待控制模块的负载即为所述待控制模块的实际负载；当所述待控制模块为多个(2个及以上)负载单元时，所述待控制模块的负载为各负载单元的负载权重与实际负载乘积的和。
具体地，各功耗统计周期均包括k个监测窗口，将所述待控制模块11在各监测窗口的负载均与所述负载阈值进行比较，并以功耗统计周期为整体得出比较结果。各功耗统计周期的监测窗口采取加权控制，即在当前功耗统计周期中去除第1个监测窗口并加入第k+1个监测窗口以得到下一功耗统计周期的各监测窗口(以时间为顺序，将最早的监测窗口去除加入最新的监测窗口)，在本实施例中，k设置为4，以此提高监测的准确性(一般地设置3个监测 窗口即可提高监测的准确性)。在实际使用中，各功耗统计周期内监测窗口的数量可根据需要设置，在对准确性要求不高或可确保准确性的前提下，可适当减少各功耗统计周期内监测窗口的数量，不以本实施例为限。
2)当前功耗统计周期内各监测窗口的负载均大于所述负载阈值时，所述待控制模块进入最低峰值功耗状态，频率等级降低为最低等级，并基于最低峰值功耗状态对应的频率等级控制所述待控制模块的工作时钟。
具体地，包括n个功耗状态，在本实施例中，n设定为4，根据频率从小到大依次排列为第一峰值功耗状态、第二峰值功耗状态、第三峰值功耗状态及正常功耗状态；在本实施例中，第一峰值功耗状态、第二峰值功耗状态及第三峰值功耗状态对应的频率等级为正常功耗状态对应频率的各分频信号。
具体地，若当前功耗统计周期内各监测窗口的负载均大于负载阈值，则无论所述待控制模块当前处于功耗状态，均调整为第一峰值功耗状态，并基于所述第一峰值功耗状态对应的频率等级控制所述待控制模块的工作时钟。
3)当前功耗统计周期内各监测窗口的负载均小于所述负载阈值，且均大于负载下限时，若当前功耗负载处于正常功耗状态则保持原有功耗状态，频率等级不变；若当前功耗负载状态低于正常功耗状态则所述待控制模块进入上一级峰值功耗状态，频率等级上升一级，并基于上一级峰值功耗状态对应的频率等级控制所述待控制模块的工作时钟，增大所述待控制模块的负载。
具体地，若当前功耗统计周期内各监测窗口的负载均小于负载阈值，且均大于所述负载下限，则在当前功耗负载处于正常功耗状态时保持正常功耗状态不变。
具体地，若当前功耗统计周期内各监测窗口的负载均小于负载阈值，且均大于所述负载下限，则在当前功耗负载状态处于第一峰值功耗状态、第二峰值功耗状态或第三峰值功耗状态时将所述待控制模块调整为上一级功耗状态，相应的频率等级上升一级，所述待控制模块的负载对应增大；即当前处于第一峰值功耗状态则上升至第二峰值功耗状态，当前处于第二峰值功耗状态则上升至第三峰值功耗状态，当前处于第三峰值功耗状态则上升至正常功耗状态。
作为本发明的一种实现方式，功耗状态上升一级，所述待控制模块的负载增大所述负载阈值1/n倍，在本实施例中，增大1/4的负载阈值。
4)当前功耗统计周期内各监测窗口的负载不均小于所述负载阈值或不均大于所述负载阈值时，保持原有功耗状态，频率等级不变，所述待控制模块的负载不变。
5)继续监测后续各功耗统计周期内所述待控制模块的负载，并在各功耗统计周期内基于负载状况调整所述待控制模块的工作时钟的频率，使得所述待控制模块的峰值功耗不超出负载阈值的情况下工作在最高频率。
6)作为本发明的一种实现方式，所述控制峰值功耗的方法还包括：当前功耗统计周期内各监测窗口的负载均低于所述负载下限时，所述工作时钟的频率在当前功耗状态对应的频率等级中下降一个频段。
具体地，所述待控制模块的负载低于负载下限时，所述待控制模块的负载较低，不需要工作在满频状态。如果当前功耗统计周期内4个监测窗口的负载均低于负载下限，所述待控制模块的频率降到下一个频段，即假设当前频率为正常工作状态下原始频率的12/16，则下降至正常工作状态下原始频率的11/16，本实施例支持最低十六分之一的降频，实际使用中可根据需要设定最低降频的频率。如果下一个功耗统计周期的负载没有低于负载下限，则在当前周期把所述待控制模块的工作频率上调一个频率等级，以保证所述待控制模块有足够的性能。
7)作为本发明的一种实现方式，如果所述待控制模块进入空闲状态，则时钟频率可以自动降到最低频率。
更具体地，下面给出一个具体的实例：
S1：在启动阶段，所述待控制模块以正常功耗状态开始工作；根据一个功耗统计周期内的负载对功耗状态进行更新。
S2：如果在一个功耗统计周期内，所述待控制模块的负载大于负载阈值，则所述待控制模块进入第一峰值功耗状态，即从正常功耗状态进入第一峰值功耗状态，频率降低三个等级；根据下一个功耗统计周期内的负载对功耗状态继续更新。
S3：如果在下一个功耗统计周期内，所述待控制模块的负载大于负载阈值，则所述待控制模块继续保持第一峰值功耗状态；如果所述待控制模块的负载小于负载阈值且大于负载下限，则所述待控制模块进入第二峰值功耗状态，当前负载增加1/4负载阈值，当前前频率提高一个等级；根据下一个功耗统计周期内的负载对功耗状态继续更新。
S4：如果在下一个功耗统计周期内，所述待控制模块的负载大于负载阈值，则所述待控制模块继续保持第二峰值功耗状态；如果所述待控制模块的负载小于负载阈值且大于负载下限，则所述待控制模块进入第三峰值功耗状态，当前负载增加1/4负载阈值，当前前频率提高一个等级；根据下一个功耗统计周期内的负载对功耗状态继续更新。
S5：如果在下一个功耗统计周期内，所述待控制模块的负载大于负载阈值，则所述待控 制模块继续保持第三峰值功耗状态；如果所述待控制模块的负载小于负载阈值且大于负载下限，则所述待控制模块进入正常功耗状态，当前负载增加1/4负载阈值，当前频率提高一个等级；负载信息统计模块会根据一个功耗统计周期内的负载对功耗负载状态继续更新。
S6：继续循环步骤S2到步骤S5，保证峰值功耗不超过负载阈值的情况下，所述待控制模块工作的最高的频率上。
本发明通过动态控制AI芯片的电压和频率来降低功耗，同时提供足够的计算性能。
如图2所示，本发明提供一种控制峰值功耗的系统1，所述控制峰值功耗的系统1用于实现所述控制峰值功耗的方法，所述控制峰值功耗的系统1包括：
待控制模块11、全局控制模块12、负载信息统计模块13、调频调压控制模块14ji时钟控制模块15。
如图2所示，所述待控制模块11包括多个负载单元。
具体地，在本实施例中，所述待控制模块11设定为AI芯片的计算单元，各负载单元包括但不限于可编程运算单元，神经网络加速器及张量处理器。
如图2所示，所述全局控制模块12连接所述待控制模块11及所述负载信息统计模块13，为所述待控制模块11提供控制信号及各负载单元的负载权重信息，并基于所述待控制模块11的工作状态获取负载阈值。
具体地，在本实施例中，所述全局控制模块12基于中央处理器(图中未显示)及驱动(图中未显示)提供的控制信息(各控制信息用于确定所述待控制模块11的工作状态)得到负载阈值、各负载单元的负载权重信息和模式控制信息，并将各负载单元的负载权重信息和模式控制信息发送至所述待控制模块11；同时，所述全局控制模块12将所述待控制模块11的负载统计信息反馈给驱动，以便于驱动根据AI芯片内部情况调整控制策略。
如图2所示，所述负载信息统计模块13连接所述全局控制模块12及所述待控制模块11，监控所述待控制模块11中各负载单元的负载信息，并基于所述全局控制模块12中各负载单元的负载权重信息，计算所述待控制模块11的负载。
具体地，所述负载信息统计模块13监控所述待控制模块11各个主要流水线(各负载单元)的负载信息，根据所述全局控制模块12提供的各负载单元的负载权重信息统计汇总各个负载单元贡献的负载，根据每个负载单元对总体性能影响的重要程度和负载单元本身的功耗，按照上一级负载单元的负载权重将所有负载单元的负载信息综合在一起得到整个待控制模块11的负载。在本实施例中，AI芯片的计算单元的主要流水线包括可编程运算单元，神 经网络加速器及张量处理器，其中神经网络加速器用来处理神经网络中的卷积运算，张量处理器用来处理神经网络中的其他数学运算，可编程运算单元进行流程控制和辅助计算；从性能角度考虑，神经网络加速器最重要，80％的运算由神经网络加速器来完成，另外15％的运算由张量处理器完成，另外的5％控制和辅助计算由可编程运算单元完成，所以负载信息统计模块13对这三个主要模块的负载系数为80％，15％和5％。假设神经网络加速器本身的功耗为load1，运算由张量处理器本身的功耗为load2，可编程运算单元本身的功耗为load3，则所述待控制模块11的负载为80％*load1+15％*load2+5％*load3。
如图2所示，所述调频调压控制模块14连接所述负载信息统计模块13，基于所述待控制模块11的负载调整所述待控制模块11的功耗状态。
具体地，作为本发明的一种实现方式，所述调频调压控制模块14包括比较单元141及多个功耗状态控制单元，在本实施例中，包括三个功耗状态控制单元，分别为第一功耗状态控制单元142、第二功耗状态控制单元143及第三功耗状态控制单元144。所述比较单元141将各功耗统计周期内所述待控制模块11的负载与所述负载阈值进行比较，在本实施例中，所述功耗统计周期内设置有4个监测窗口。各功耗状态控制单元基于所述比较单元141输出的比较结果控制所述待控制模块11进入相应的功耗状态。各功耗状态的跳变过程及跳变条件参考所述控制峰值功耗的方法，在此不一一赘述。所述调频调压控制模块14还可以对当前最低负载情况进行分析，实现降频操作，在此不一一赘述。
如图2所示，所述时钟控制模块15连接于所述调频调压控制模块14与所述待控制模块11之间，基于所述调频调压控制模块14输出的信号产生相应频率的时钟信号，并控制所述待控制模块11基于所述时钟信号工作。
具体地，如图3所示，在本实施例中，所述时钟控制模块15包括数字分频单元，用于对正常工作状态下原始频率进行16分频，以得到相应频率等级对应的频率及降频频率，在本实施例中，所述第一峰值功耗状态控制单元142、所述第二峰值功耗状态控制单元143及所述第三峰值功耗状态控制单元144对应的频率等级依次增大，分别为正常工作状态下原始频率的1/4，1/2及3/4，在实际使用中可根据需要设定各频率等级控制单元对应的频率。在实际使用中，所述时钟控制模块15可采用等相位分频，不以本实施例为限。
本发明的控制峰值功耗的方法及系统提高了AI芯片中动态功耗的精准控制，在不影响芯片性能的情况下，可以动态控制芯片的频率达到降低峰值功耗的效果；同时也解决了传统功耗控制方式中的CPU占用率问题和面积开销太大的问题。
本发明的控制峰值功耗的方法及系统通过对AI芯片各个模块负载情况的实时监控，进行 自动的频率调控；也可以通过外部软件主动控制和内部硬件自动控制相结合的方式进行频率调控和电压调控。
综上所述，本发明提供一种控制峰值功耗的方法及系统，包括：监测当前功耗统计周期内待控制模块的负载，当前功耗统计周期内各监测窗口的负载均大于负载阈值时，所述待控制模块进入最低峰值功耗状态，并基于最低峰值功耗状态对应的频率等级控制所述待控制模块的工作时钟；当前功耗统计周期内各监测窗口的负载均小于所述负载阈值，且均大于负载下限时，若当前功耗负载处于正常功耗状态则保持原有功耗状态，频率等级不变；若当前峰值功耗状态低于正常功耗状态则所述待控制模块进入上一级峰值功耗状态，频率等级上升一级，并基于上一级峰值功耗状态对应的频率等级控制所述待控制模块的工作时钟，增大所述待控制模块的负载；当前功耗统计周期内各监测窗口的负载不均小于所述负载阈值或不均大于所述负载阈值时，保持原有功耗状态，频率等级不变，所述待控制模块的负载不变；继续监测后续各功耗统计周期内所述待控制模块的负载，并在各功耗统计周期内基于负载状况调整所述待控制模块的工作时钟的频率，使得所述待控制模块的峰值功耗不超出负载阈值的情况下工作在最高频率。本发明的控制峰值功耗的方法及系统实时监控负载情况，通过动态智能控制工作频率来降低功耗，达到性能和功耗的平衡点，使得峰值功耗不超出负载阈值的情况下工作在最高频率，在实现功耗控制的前提下有效提高工作效率。所以，本发明有效克服了现有技术中的种种缺点而具高度产业利用价值。
上述实施例仅例示性说明本发明的原理及其功效，而非用于限制本发明。任何熟悉此技术的人士皆可在不违背本发明的精神及范畴下，对上述实施例进行修饰或改变。因此，举凡所属技术领域中具有通常知识者在未脱离本发明所揭示的精神与技术思想下所完成的一切等效修饰或改变，仍应由本发明的权利要求所涵盖。

Claims (12)

  1. 一种控制峰值功耗的方法，其特征在于，所述控制峰值功耗的方法至少包括：
    监测当前功耗统计周期内待控制模块的负载，当前功耗统计周期内各监测窗口的负载均大于负载阈值时，所述待控制模块进入最低峰值功耗状态，并基于最低峰值功耗状态对应的频率等级控制所述待控制模块的工作时钟；
    当前功耗统计周期内各监测窗口的负载均小于所述负载阈值，且均大于负载下限时，若当前功耗负载处于正常功耗状态则保持原有功耗状态，频率等级不变；若当前峰值功耗状态低于正常功耗状态则所述待控制模块进入上一级峰值功耗状态，频率等级上升一级，并基于上一级峰值功耗状态对应的频率等级控制所述待控制模块的工作时钟，增大所述待控制模块的负载；
    当前功耗统计周期内各监测窗口的负载不均小于所述负载阈值或不均大于所述负载阈值时，保持原有功耗状态，频率等级不变，所述待控制模块的负载不变；
    继续监测后续各功耗统计周期内所述待控制模块的负载，并在各功耗统计周期内基于负载状况调整所述待控制模块的工作时钟的频率，使得所述待控制模块的峰值功耗不超出负载阈值的情况下工作在最高频率。
  2. 根据权利要求1所述的控制峰值功耗的方法，其特征在于：所述待控制模块包括多个负载单元时，所述待控制模块的负载为各负载单元的负载权重与实际负载乘积的和。
  3. 根据权利要求1所述的控制峰值功耗的方法，其特征在于：所述控制峰值功耗的方法包括n个功耗状态，功耗状态上升一级所述待控制模块的负载增大所述负载阈值的1/n倍，n为正整数。
  4. 根据权利要求1所述的控制峰值功耗的方法，其特征在于：各功耗统计周期均包括k个监测窗口，在当前功耗统计周期中去除第1个监测窗口并加入第k+1个监测窗口以得到下一功耗统计周期，其中k为不小于3的正整数。
  5. 根据权利要求1～4任意一项所述的控制峰值功耗的方法，其特征在于：所述控制峰值功耗的方法还包括：当前功耗统计周期内各监测窗口的负载均低于所述负载下限时，所述工作时钟的频率在当前功耗状态对应的频率等级中下降一个频段。
  6. 根据权利要求1～4任意一项所述的控制峰值功耗的方法，其特征在于：当所述待控制模 块进入空闲状态时，所述工作时钟的频率降低至最低频率。
  7. 根据权利要求1所述的控制峰值功耗的方法，其特征在于：所述控制峰值功耗的方法用于AI芯片的功耗控制。
  8. 一种控制峰值功耗的系统，基于如权利要求1～7任意一项所述的控制峰值功耗的方法，其特征在于，所述控制峰值功耗的系统至少包括：
    待控制模块，包括多个负载单元；
    全局控制模块，连接所述待控制模块及所述负载信息统计模块，为所述待控制模块提供控制信号及各负载单元的负载权重信息，并基于所述待控制模块的工作状态获取负载阈值；
    负载信息统计模块，连接所述全局控制模块及所述待控制模块，监控所述待控制模块中各负载单元的负载信息，并基于所述全局控制模块中各负载单元的负载权重信息，计算所述待控制模块的负载；
    调频调压控制模块，连接所述负载信息统计模块，基于所述待控制模块的负载调整所述待控制模块的功耗状态；
    时钟控制模块，连接于所述调频调压控制模块与所述待控制模块之间，基于所述调频调压控制模块输出信号产生相应频率的时钟信号，并控制所述待控制模块基于所述时钟信号工作。
  9. 根据权利要求8所述的控制峰值功耗的系统，其特征在于：所述控制峰值功耗的系统应用于AI芯片中。
  10. 根据权利要求8或9所述的控制峰值功耗的系统，其特征在于：所述待控制模块包括可编程运算单元，神经网络加速器或张量处理器。
  11. 根据权利要求8所述的控制峰值功耗的系统，其特征在于：所述调频调压控制模块包括比较单元及多个功耗状态控制单元；所述比较单元将各功耗统计周期内所述待控制模块的负载与所述负载阈值进行比较；各功耗状态控制单元基于所述比较单元输出的比较结果控制所述待控制模块进入相应的功耗状态。
  12. 根据权利要求8所述的控制峰值功耗的系统，其特征在于：所述时钟控制模块包括数字分频单元。
PCT/CN2019/122382 2019-05-29 2019-12-02 控制峰值功耗的方法及系统 WO2020238121A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
KR1020217030998A KR20210134355A (ko) 2019-05-29 2019-12-02 피크 전력 소비의 제어 방법 및 시스템
EP19930416.3A EP3816769A4 (en) 2019-05-29 2019-12-02 PROCEDURE AND SYSTEM FOR REGULATING PEAK POWER CONSUMPTION
JP2021570476A JP7304972B2 (ja) 2019-05-29 2019-12-02 ピーク消費電力を制御する方法及びシステム
US17/263,901 US20220342474A1 (en) 2019-05-29 2019-12-02 Method and system for controlling peak power consumption

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910457063.2 2019-05-29
CN201910457063.2A CN112015259B (zh) 2019-05-29 2019-05-29 控制峰值功耗的方法及系统

Publications (1)

Publication Number Publication Date
WO2020238121A1 true WO2020238121A1 (zh) 2020-12-03

Family

ID=73501730

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/122382 WO2020238121A1 (zh) 2019-05-29 2019-12-02 控制峰值功耗的方法及系统

Country Status (6)

Country Link
US (1) US20220342474A1 (zh)
EP (1) EP3816769A4 (zh)
JP (1) JP7304972B2 (zh)
KR (1) KR20210134355A (zh)
CN (1) CN112015259B (zh)
WO (1) WO2020238121A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111782577B (zh) * 2019-04-04 2023-03-24 安徽寒武纪信息科技有限公司 数据处理装置及方法以及相关产品

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120011377A1 (en) * 2010-07-09 2012-01-12 Elitegroup Computer Systems Co., Ltd. Computer system capable of adjusting operating frequency dynamically
CN102609319A (zh) * 2011-01-20 2012-07-25 中国移动通信有限公司 一种处理器调频方法、装置及设备
CN103678247A (zh) * 2012-08-29 2014-03-26 三星电子株式会社 动态电压频率调节方法和装置
CN103955264A (zh) * 2014-05-15 2014-07-30 乐视致新电子科技（天津）有限公司 动态调节处理器工作频率的方法及系统
CN107341091A (zh) * 2017-07-04 2017-11-10 郑州云海信息技术有限公司 分布式存储系统功耗管理方法及装置

Family Cites Families (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP4033066B2 (ja) * 2003-05-07 2008-01-16 ソニー株式会社 周波数制御装置、情報処理装置、周波数制御方法及びプログラム
JP2005196430A (ja) * 2004-01-07 2005-07-21 Hiroshi Nakamura 半導体装置および半導体装置の電源電圧／クロック周波数制御方法
CN101365192A (zh) * 2007-08-10 2009-02-11 大唐移动通信设备有限公司 调整嵌入式处理器和移动终端处理器的频率的方法、装置
US8160753B2 (en) * 2008-07-31 2012-04-17 Microsemi Corp.—Analog Mixed Signal Group Ltd. Time integrated guard band
KR101533572B1 (ko) * 2009-05-20 2015-07-03 삼성전자주식회사 전력 관리 방법
US8410768B2 (en) * 2010-01-19 2013-04-02 Delta Electronics, Inc. Switch-mode power supply having reduced audible noise
CN102314208B (zh) * 2010-06-30 2016-08-03 重庆重邮信科通信技术有限公司 一种动态调整嵌入式设备频率电压的方法及装置
CN102662822B (zh) * 2012-04-26 2015-02-04 华为技术有限公司 负载监控装置和方法
US20140095696A1 (en) * 2012-10-01 2014-04-03 Emerson Electric Co. Evaluating utilization of one or more computer systems
CN103176944B (zh) * 2013-03-01 2016-01-20 福州瑞芯微电子股份有限公司 基于不同制造工艺实现的低功耗多核soc及其设计方法
CN103197754B (zh) * 2013-04-01 2016-05-25 华为技术有限公司 一种降低芯片功耗的方法和装置
CN103559092B (zh) * 2013-11-12 2016-08-17 中国人民解放军国防科学技术大学 异构并行计算结点的峰值功耗控制方法
JP2015130035A (ja) * 2014-01-07 2015-07-16 富士通株式会社 半導体装置および制御方法
US9244747B2 (en) * 2014-03-13 2016-01-26 Qualcomm Incorporated System and method for providing dynamic clock and voltage scaling (DCVS) aware interprocessor communication
CN105511591B (zh) * 2015-12-31 2018-10-12 天津飞腾信息技术有限公司 基于双阈值功耗自适应的dvfs调节算法
JP2018005569A (ja) * 2016-07-01 2018-01-11 富士通株式会社 半導体装置及び半導体装置の制御方法
CN109117262B (zh) * 2017-06-22 2022-01-11 深圳市中兴微电子技术有限公司 一种基带处理芯片cpu动态调频方法及无线终端
CN107340500A (zh) * 2017-06-30 2017-11-10 中国航空工业集团公司雷华电子技术研究所 一种雷达信号处理平台动态功耗控制方法及系统
US10565020B2 (en) * 2017-08-29 2020-02-18 International Business Machines Corporation Adjustment of the number of central processing units to meet performance requirements of an I/O resource
US11029972B2 (en) * 2019-02-01 2021-06-08 Dell Products, Lp Method and system for profile learning window optimization

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120011377A1 (en) * 2010-07-09 2012-01-12 Elitegroup Computer Systems Co., Ltd. Computer system capable of adjusting operating frequency dynamically
CN102609319A (zh) * 2011-01-20 2012-07-25 中国移动通信有限公司 一种处理器调频方法、装置及设备
CN103678247A (zh) * 2012-08-29 2014-03-26 三星电子株式会社 动态电压频率调节方法和装置
CN103955264A (zh) * 2014-05-15 2014-07-30 乐视致新电子科技（天津）有限公司 动态调节处理器工作频率的方法及系统
CN107341091A (zh) * 2017-07-04 2017-11-10 郑州云海信息技术有限公司 分布式存储系统功耗管理方法及装置

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3816769A4 *

Also Published As

Publication number Publication date
EP3816769A1 (en) 2021-05-05
CN112015259B (zh) 2022-06-21
JP7304972B2 (ja) 2023-07-07
JP2022535357A (ja) 2022-08-08
CN112015259A (zh) 2020-12-01
US20220342474A1 (en) 2022-10-27
KR20210134355A (ko) 2021-11-09
EP3816769A4 (en) 2021-11-03

Similar Documents

Publication Publication Date Title
US7343508B2 (en) Dynamic clock control circuit for graphics engine clock and memory clock and method
EP2194442B1 (en) A multi-core processor and a method of data communication between the cores
Gu et al. Energy efficient scheduling of servers with multi-sleep modes for cloud data center
EP2049969A2 (en) Self-monitoring and self-adjusting power consumption computer control system
KR20020018309A (ko) 시피유 스케쥴링 방법
da Rosa et al. Power consumption reduction in MPSoCs through DFS
US9310872B2 (en) Processor frequency mainly depending on a target frame rate while processing a graphics application
WO2020238121A1 (zh) 控制峰值功耗的方法及系统
Kim et al. 2.4 A 7nm high-performance and energy-efficient mobile application processor with tri-cluster CPUs and a sparsity-aware NPU
WO2017148253A1 (zh) 一种节能管理实现方法、装置及网络设备
CN116627237B (zh) 一种基于芯粒Chiplet的电源管理芯片架构系统
US11853111B2 (en) System and method for controlling electrical current supply in a multi-processor core system via instruction per cycle reduction
CN105933702A (zh) 一种基于任务敏感的功耗控制方法
US11360541B2 (en) Programmable voltage regulation for data processor
US20140125401A1 (en) System for controlliing gatings of a multi-core processor
CN112214295A (zh) 一种多cpu/gpu异构服务器集群的低能耗作业调度方法
JP2017021513A (ja) マルチコアプロセッサ、マルチコアプロセッサのクロック制御方法およびクロック制御プログラム
CN113254216A (zh) 边缘计算模组及其功耗控制方法
CN117215394B (zh) 一种面向多核处理器的片上温度和能耗控制装置
JP4972522B2 (ja) データ処理システム
EP3770727A1 (en) Technology for managing per-core performance states
CN116991682A (zh) 一种高性能异构并行计算机的最大功耗控制方法
Wu et al. A Low Power Consumption Period Task Scheduling Algorithm in Real-Time System
CN116529694A (zh) 用于通用串行总线(usb)类型-c端口的功率管理

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19930416

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 20217030998

Country of ref document: KR

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 2021570476

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE