Server board card, performance adjusting method of server board card and chip

Abstract

The invention discloses a server board card, a performance adjusting method of the server board card and a chip, and belongs to the field of servers. The server board card comprises a management processor and a computing chip, and the management processor is used for managing elements in the server board card; the management processor is used for acquiring historical operation parameters of the elements, and the historical operation parameters comprise parameters used for reflecting the operation performance of the elements; the calculation chip is used for predicting a current operation parameter of the element according to the historical operation parameter of the element, and the current operation parameter is used for representing an operation parameter corresponding to a current moment in an operation parameter change trend corresponding to the historical operation parameter; and the management processor is also used for adjusting the performance of the element according to the current operation parameter of the element. According to the method and the device, the limitation of adjusting the performance of the board card based on the preset performance level is broken, the performance adjustment with finer granularity is realized, the power supply use and performance of the server board card are balanced, and the response capability of the server board card for different performance requirements is enhanced.

Description

Technical Field

[0001] The present application relates to the field of servers, and in particular to a server board, a performance adjustment method for the server board, and a chip. Background Art

[0002] As server performance continues to improve, power consumption and heat dissipation issues are becoming increasingly prominent. In order to reasonably reduce server energy consumption and reduce the heat dissipation burden, performance adjustment strategies are currently set for server boards.

[0003] In the related art, a static performance adjustment strategy is set for the server card. Specifically, the server card is preset with multiple performance levels, and the performance of the server card is different when it is at different performance levels. During the operation of the server, the server card can be switched between the multiple performance levels according to the actual situation to achieve the purpose of balancing performance and energy consumption.

[0004] Adjusting the performance of the server board through the above solution is limited by multiple preset performance levels, which may lead to excessive power consumption when high performance is not needed, and insufficient power support when high performance is needed. There are certain limitations on the performance adjustment of the server board. Summary of the invention

[0005] The present application provides a server card, a performance adjustment method for the server card, and a chip, which can break the limitation of adjusting the performance of the card based on a preset performance level and achieve more fine-grained performance adjustment. The technical solution is as follows:

[0006] According to one aspect of the present application, a server card is provided, the server card comprising a management processor and a computing chip, the management processor being used to manage components in the server card;

[0007] The management processor is used to obtain historical operating parameters of the component, wherein the historical operating parameters include parameters used to reflect the operating performance of the component;

[0008] The computing chip is used to predict the current operating parameters of the component according to the historical operating parameters of the component, and the current operating parameters are used to characterize the operating parameters corresponding to the current moment in the operating parameter change trend corresponding to the historical operating parameters;

[0009] The management processor is further configured to adjust the performance of the component according to the current operating parameters of the component.

[0010] According to another aspect of the present application, a performance adjustment method for a server card is provided, the method being applied to the server card, the method comprising:

[0011] Acquire historical operating parameters of components in the server board, wherein the historical operating parameters include parameters used to reflect operating performance of the components;

[0012] Predicting the current operating parameters of the component according to the historical operating parameters of the component, the current operating parameters being used to characterize the operating parameters corresponding to the current moment in the operating parameter change trend corresponding to the historical operating parameters;

[0013] The performance of the component is adjusted according to the current operating parameters of the component.

[0014] According to another aspect of the present application, a chip is provided, the chip comprising a management processor and a computing chip, the management processor being used to manage components in a server board;

[0015] The management processor is used to obtain historical operating parameters of the component, wherein the historical operating parameters include parameters used to reflect the operating performance of the component;

[0016] The computing chip is used to predict the current operating parameters of the component according to the historical operating parameters of the component, and the current operating parameters are used to characterize the operating parameters corresponding to the current moment in the operating parameter change trend corresponding to the historical operating parameters;

[0017] The management processor is further configured to adjust the performance of the component according to the current operating parameters of the component.

[0018] According to another aspect of the present application, a server is provided, wherein the server comprises the server board as described in the above aspect.

[0019] The beneficial effects of the technical solution provided by this application include at least:

[0020] The historical operating parameters of the components in the server board are collected by the management processor, so that the current operating parameters of the components are predicted by the calculation chip. Since the current operating parameters are determined based on the operating parameter change trend corresponding to the historical operating parameters, the predicted current operating parameters can reflect the real-time performance requirements of the components in the server board. By adjusting the performance of the components in the server board based on the current operating parameters, the performance of the components in the server board can be dynamically adjusted according to the real-time operating status of the server board, breaking the limitation of adjusting the board performance based on the preset performance level, and adjusting the components to the appropriate performance mode in advance, achieving more fine-grained performance adjustment, balancing the power usage and performance of the server board, and enhancing the responsiveness of the server board when facing different performance requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0022] Figure 1 is a schematic diagram of a server board provided by an exemplary embodiment of the present application;

[0023] Figure 2 is a schematic diagram of a computing core provided by an exemplary embodiment of the present application;

[0024] Figure 3 is a schematic diagram of a server board provided by an exemplary embodiment of the present application;

[0025] Figure 4 is a schematic diagram of an AI reasoning board provided by an exemplary embodiment of the present application;

[0026] Figure 5 is a schematic diagram of a system software scheduling framework of a server board provided by an exemplary embodiment of the present application;

[0027] Figure 6 is a schematic diagram of a system runtime scheduling framework of a server card provided by an exemplary embodiment of the present application;

[0028] Figure 7 is a schematic diagram of a process of adjusting the performance of a server board provided by an exemplary embodiment of the present application;

[0029] Figure 8 It is a flow chart of a method for adjusting the performance of a server board provided by an exemplary embodiment of the present application;

[0030] Fig. 9 It is a flow chart of a method for adjusting the performance of a server board provided by an exemplary embodiment of the present application;

[0031] Fig.10 It is a flow chart of a method for adjusting the performance of a server board provided by an exemplary embodiment of the present application;

[0032] Fig.11 is a schematic diagram of a process of adjusting the performance of a component of a server board provided by an exemplary embodiment of the present application;

[0033] Fig.12 is a structural block diagram of a computer system provided by an exemplary embodiment of the present application;

[0034] Fig.13is a structural block diagram of a computer system provided by an exemplary embodiment of the present application;

[0035] Fig.14 is a structural block diagram of a computer system provided by an exemplary embodiment of the present application;

[0036] Fig.15 It is a structural schematic diagram of a performance adjustment device for a server card provided by an exemplary embodiment of the present application;

[0037] Fig.16 It is a structural schematic diagram of a performance adjustment device for a server card provided by an exemplary embodiment of the present application;

[0038] Fig.17 It is a structural diagram of a computer device provided by an exemplary embodiment of the present application.

[0039] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application. DETAILED DESCRIPTION

[0040] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below in conjunction with the accompanying drawings.

[0041] First, the nouns involved in this application are introduced:

[0042] Server card: A server card is a card installed in a server, which can provide corresponding functions for the server. For example, the card installed in the server can control the operation of the hardware connected to the server. For example, the operation of the display, external devices, etc. can be controlled. For example, the card installed in the server can provide computing power for the server to process corresponding tasks. For example, the server card performs computing processing on the information input to the server, thereby outputting corresponding results.

[0043] In some embodiments, the server board in the present application is used to process tasks related to artificial intelligence (AI), such as AI reasoning tasks. In this case, the server board may also be referred to as an AI board, an AI reasoning board, etc., and the embodiments of the present application are not limited to this.

[0044] Artificial intelligence: It is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that machines have the functions of perception, reasoning and decision-making.

[0045] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. Basic artificial intelligence technologies generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, pre-trained model technology, operation / interaction systems, mechatronics, etc. Among them, pre-trained models are also called large models and basic models. After fine-tuning, they can be widely used in downstream tasks in various major directions of artificial intelligence. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0046] Machine Learning (ML): It is a multi-disciplinary interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithmic complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning by teaching. The pre-trained model is the latest development in deep learning, which integrates the above technologies. In some embodiments, the prediction model in this application can be regarded as a machine learning model.

[0047] Computer Vision (CV) Technology: Computer vision is a science that studies how to make machines "see". To put it more specifically, it refers to machine vision that uses cameras and computers to replace human eyes to identify, detect and measure targets, and further performs image processing to make computer processing into images that are more suitable for human eye observation or transmission to instruments for detection. As a scientific discipline, computer vision studies related theories and technologies, and attempts to establish an artificial intelligence system that can obtain information from images or multidimensional data. Large model technology has brought important changes to the development of computer vision technology. Pre-trained models in the visual field such as swin-transformer, ViT, V-MOE, and MAE can be quickly and widely applied to downstream specific tasks after fine tuning. Computer vision technology usually includes image processing, image recognition, image semantic understanding, image retrieval, OCR (Optical Character Recognition), video processing, video semantic understanding, video content / behavior recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous positioning and mapping, and other technologies, as well as common biometric recognition technologies such as face recognition and fingerprint recognition. In some embodiments, the server board in the present application supports processing computer vision tasks.

[0048] With the research and advancement of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, driverless cars, autonomous driving, drones, digital twins, virtual humans, robots, artificial intelligence generated content (AIGC), conversational interaction, smart medical care, smart customer service, game AI, etc. It is believed that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0049] Figure 1 FIG. 1 is a schematic diagram of a server card provided by an exemplary embodiment of the present application. Figure 1As shown, the server board 100 includes a management processor 101 and a computing chip 102. In some embodiments, the management processor 101 is electrically connected to the computing chip 102. The management processor 101 is used to manage the components in the server board, and the components in the server board 100 include one or more hardware for supporting the operation of the server board 100. In some embodiments, the management processor 101 is electrically connected to the components in the server board. In some embodiments, the components in the server board 100 include a first component that is on the chip of the computing chip 102 and a second component that is off the chip of the computing chip 102. The management processor 101 is used to manage the first component and the second component at the same time, or the management processor 101 includes multiple, respectively used to manage the first component and the second component. In some embodiments, the management processor 101 for managing the first component is located on the chip of the computing chip 102 and is electrically connected to the computing chip 102. The management processor 101 for managing the second component is located off the chip of the computing chip 102 and is electrically connected to the management processor 101 for managing the first component. In some embodiments, the management processor 101 for managing the first element is electrically connected to the first element. The management processor 101 for managing the second element is electrically connected to the second element. The computing chip 102 is used to provide computing power for the server board 100. Optionally, the computing chip 102 includes multiple computing cores. In some embodiments, the server board 100 in the present application is used to process AI-related tasks. In this case, the computing chip 102 may also be referred to as an AI chip, an AI reasoning chip, etc., which is not limited in the embodiments of the present application.

[0050] The management processor 101 is used to obtain the historical operating parameters of the component. The historical operating parameters include parameters used to reflect the operating performance of the component. For example, the above-mentioned components include at least one of the hardware such as the central processing unit (CPU), double rate synchronous dynamic random access memory (DDR), high performance compute vision group (HCVG), video decoder (VDEC), baseboard management controller (BMC), flash memory (Flash), fan, etc. The historical operating parameters of the component include the power state of the component in the historical period, such as sleep state, standby state, running state, etc., and the performance state of the component in the historical period, such as operating frequency, read and write speed, voltage, temperature, etc. The historical period is set by the developer.

[0051] The computing chip 102 is used to predict the current operating parameters of the component based on the historical operating parameters of the component. The current operating parameters of the component are used to characterize the operating parameters corresponding to the current moment in the operating parameter change trend corresponding to the historical operating parameters of the component. That is, the current operating parameters of the component are obtained based on the predicted change trend of the historical operating parameters of the component. For example, if the parameter change trend of the historical operating parameters is that the performance gradually decreases, then the performance corresponding to the current operating parameters is lower than the historical operating parameters; if the parameter change trend of the historical operating parameters is that the performance gradually increases, then the performance corresponding to the current operating parameters is higher than the historical operating parameters. After predicting the current operating parameters, the management processor 101 is also used to adjust the performance of the component according to the current operating parameters of the component. That is, the management processor 101 will adjust the operating parameters of the component to the current operating parameters, and use this to control the operation of the component, thereby adjusting the performance of the component.

[0052] In summary, the server board provided in this embodiment collects the historical operating parameters of the components in the server board through the management processor, and thus predicts the current operating parameters of the components through the calculation chip. Since the current operating parameters are determined based on the operating parameter change trend corresponding to the historical operating parameters, the predicted current operating parameters can reflect the real-time performance requirements of the components in the server board. By adjusting the performance of the components in the server board based on the current operating parameters, it is possible to dynamically adjust the performance of the components in the server board according to the real-time operating status of the server board, breaking the limitation of adjusting the board performance based on the preset performance level, and adjusting the components to the appropriate performance mode in advance, achieving more fine-grained performance adjustment, balancing the power usage and performance of the server board, and enhancing the responsiveness of the server board when facing different performance requirements.

[0053] In an optional implementation, Figure 2 FIG. 1 is a schematic diagram of a computing core provided by an exemplary embodiment of the present application. Figure 2 As shown, the computing chip 102 includes multiple computing cores 103, and the computing cores 103 are used to perform computing-related tasks. In some embodiments, the multiple computing cores 103 are electrically connected to the management processor 101. In some embodiments, the server board 100 in the present application is used to process AI-related tasks. In this case, the computing core 103 can also be called an AI core (AI Core), an AI processor (AI Processor), etc., and the embodiments of the present application are not limited to this.

[0054] The above-mentioned multiple computing cores 103 include a first computing core 1031, which is any computing core among the multiple computing cores 103. The first computing core 1031 is used to run a prediction model, which is an AI algorithm model. Optionally, the prediction model is used to process at least one of a regression task and a classification task, for example, a regression task is used to predict the usage rate of a component (such as a CPU), and a classification task is used to predict whether the component needs to enter a sleep mode. The first computing core 1031 is used to run a prediction model to predict current operating parameters based on historical operating parameters.

[0055] In some embodiments, continue to refer to Figure 2 . The multiple computing cores 103 also include a second computing core 1032, which is any computing core in the multiple computing cores 103, and the second computing core 1032 is different from the first computing core 1031. In the case where a prediction model is deployed in the first computing core 1031, the first computing core 1031 is used to run the prediction model to predict the current operating parameters based on the historical operating parameters. In the case where the prediction model is not deployed in the first computing core 1031, the second computing core 1032 is used to deploy the prediction model in the first computing core 1031. After the prediction model is deployed in the first computing core 1031, the first computing core 1031 is used to run the prediction model to predict the current operating parameters based on the historical operating parameters. That is, in the case where the prediction model is not deployed in the first computing core 1031, the prediction model can be deployed in the first computing core 1031 through the second computing core 1032, and then the current operating parameters of the component can be predicted through the first computing core 1031.

[0056] In some embodiments, the second computing core 1032 is deployed with a prediction model to be trained. The second computing core 1032 is used to train the prediction model to be trained through training samples and training labels, so as to obtain a prediction model that has been trained. Among them, the training samples and training labels are obtained by preprocessing historical operation parameters. Optionally, for the preprocessing process of historical operation parameters, since the historical operation parameters are sampled at the sampling time within the historical period, they are time series, so it is possible to perform time series feature extraction on the historical operation parameters to obtain training samples. For example, by performing time series feature extraction on the historical operation parameters, the average value and standard deviation of the historical operation parameters can be obtained, so as to reflect the short-term change trend of the historical operation parameters and obtain training samples. By performing linear regression modeling on the historical operation parameters, training labels can be obtained. For example, the slope and intercept of the linear regression modeling can reflect the long-term change trend of the historical operation parameters, so as to obtain training labels. After obtaining the prediction model that has been trained, the second computing core 1032 is also used to transmit the prediction model that has been trained to the first computing core 1031.

[0057] In an optional implementation, Figure 3 FIG. 1 is a schematic diagram of a server card provided by an exemplary embodiment of the present application. Figure 3 As shown, the server board 100 further includes a system scheduling processor 104 and a core scheduling processor (also referred to as a cluster scheduling processor) 105. The system scheduling processor 104 is used to schedule components in the computing chip 102, and the core scheduling processor 105 is used to schedule multiple computing cores 103. The system scheduling processor 104 and the core scheduling processor 105 are located in the computing chip 102. In some embodiments, the system scheduling processor 104 is electrically connected to the management processor 101, the management processor 101 is electrically connected to the core scheduling processor 105, and the core scheduling processor 105 is electrically connected to the multiple computing cores 103.

[0058] The system scheduling processor 104 is used to pre-process the historical operating parameters to obtain the above-mentioned training samples and training labels. In some embodiments, the system scheduling processor 104 is used to perform time series feature extraction on the historical operating parameters to obtain training samples. The system scheduling processor 104 is also used to perform linear regression modeling on the historical operating parameters to obtain training labels. For the process of pre-processing the historical operating parameters to obtain training samples and training labels, reference can be made to the relevant content in the previous text, and the embodiments of the present application are not repeated here. The system scheduling processor 104 is also used to transmit the training samples and training labels to the core scheduling processor 105. The core scheduling processor 105 is used to transmit the training samples and training labels to the second computing core 1032. The second computing core 1032 is used to transmit the trained prediction model to the first computing core 1031 through the core scheduling processor 105. That is, the second computing core 1032 transmits the trained prediction model to the core scheduling processor 105, and then the core scheduling processor 105 transmits it to the first computing core 1031, so as to predict the current operating parameters through the first computing core 1031. In some embodiments, the system scheduling processor 104 is also used to perform the above step of determining whether the prediction model is deployed in the first computing core 1031.

[0059] In some embodiments, the server card 100 includes a first component that is on-chip of the computing chip 102, and a second component that is off-chip of the computing chip 102. Figure 3The first element inside the computing chip 102 includes an element located inside the computing chip 102, and the second element outside the computing chip 102 includes an element located outside the computing chip 102. The management processor 101 is used to obtain at least one of the first historical operating parameters of the first element and the second historical operating parameters of the second element. The computing chip 102 is used to predict the current operating parameters of the first element according to the first historical operating parameters and predict at least one of the current operating parameters of the second element according to the second historical operating parameters.

[0060] In some embodiments, continue to refer to Figure 3 , the management processor 101 includes a first management processor 1011 located on the chip of the computing chip 102, and a second management processor 1012 located off the chip of the computing chip 102. The first management processor 1011 is used to manage the first component on the chip of the computing chip 102, and the second management processor 1012 is used to manage the second component off the chip of the computing chip 102. The first management processor 1011 is used to obtain the first historical operating parameters of the first component. The second management processor 1012 is used to obtain the second historical operating parameters of the second component. In some embodiments, the second management processor 1012 is used to transmit the second historical operating parameters of the second component to the first management processor 1011. That is, the historical operating parameters of the components on and off the chip of the computing chip 102 are all summarized to the first management processor 1011. The first management processor 1011 is used to transmit the first historical operating parameters to the computing chip 102, so that the computing chip 102 predicts the current operating parameters of the first component based on the first historical operating parameters. The first management processor 1011 is further configured to transmit the second historical operating parameter to the computing chip 102 so that the computing chip 102 can predict the current operating parameter of the second component based on the second historical operating parameter.

[0061] In some embodiments, when the current operating parameter is the operating parameter of the first element, the first management processor 1011 is used to adjust the performance of the first element according to the current operating parameter. That is, the first management processor 1011 adjusts the operating parameter of the first element to the current operating parameter, and controls the operation of the first element thereby adjusting the performance of the first element. When the current operating parameter is the operating parameter of the second element, the first management processor 1011 is used to transmit the current operating parameter to the second management processor 1012. The second management processor 1012 is used to adjust the performance of the second element according to the current operating parameter. That is, the second management processor 1012 adjusts the operating parameter of the second element to the current operating parameter, and controls the operation of the second element thereby adjusting the performance of the second element.

[0062] In some embodiments, when the server board 100 receives the first instruction, the server board 100 starts executing the steps of obtaining the historical operating parameters, predicting the current operating parameters based on the historical operating parameters, and adjusting the performance of the component based on the current operating parameters. When the server board 100 receives the second instruction, the server board 100 stops executing the steps of obtaining the historical operating parameters, predicting the current operating parameters based on the historical operating parameters, and adjusting the performance of the component based on the current operating parameters. In some embodiments, the first instruction and the second instruction are received and detected by the system scheduling processor 104.

[0063] In summary, the server board provided in this embodiment collects the historical operating parameters of the components in the server board through the management processor, and thus predicts the current operating parameters of the components through the calculation chip. Since the current operating parameters are determined based on the operating parameter change trend corresponding to the historical operating parameters, the predicted current operating parameters can reflect the real-time performance requirements of the components in the server board. By adjusting the performance of the components in the server board based on the current operating parameters, it is possible to dynamically adjust the performance of the components in the server board according to the real-time operating status of the server board, breaking the limitation of adjusting the board performance based on the preset performance level, and adjusting the components to the appropriate performance mode in advance, achieving more fine-grained performance adjustment, balancing the power usage and performance of the server board, and enhancing the responsiveness of the server board when facing different performance requirements.

[0064] The server card provided in this embodiment also predicts the current operating parameters of the component by running the prediction model through the first computing core. Since the first computing core is located inside the server card, the speed of predicting the current operating parameters of the component can be improved.

[0065] The server board provided in this embodiment also deploys the prediction model in the first computing core through the second computing core, thereby realizing the deployment of the prediction model without external information.

[0066] The server board provided in this embodiment also trains the prediction model to be trained through the second computing core, providing a solution for efficiently training the model through the computing core inside the server board.

[0067] The server board provided in this embodiment also realizes the deployment of the trained prediction model to the first computing core through the collaboration of the system scheduling processor and the core scheduling processor.

[0068] The server board provided in this embodiment also pre-processes the historical operating parameters of the components through the system scheduling processor, providing a solution for efficiently determining the training data of the prediction model.

[0069] The server card provided in this embodiment also obtains historical operating parameters of components inside and outside the computing chip to predict corresponding current operating parameters, thereby achieving comprehensive prediction of current operating parameters of components in the server card.

[0070] The server card provided in this embodiment also manages the on-chip and off-chip components of the computing chip through different management processors, thereby achieving accurate control of the components in the server card.

[0071] The server board provided in this embodiment can also transmit the predicted current operating parameters to the corresponding management processor through the collaboration between the first management processor and the second management processor.

[0072] The server board provided in this embodiment also adjusts the performance of the component based on the predicted current operating parameters through the management processor corresponding to the component, thereby achieving accurate adjustment of the performance of the component.

[0073] In some embodiments, the server card in the present application is used for AI-related computing, such as for running AI reasoning tasks. In this case, the server card can be referred to as an AI card or an AI reasoning card. For example, Figure 4 FIG. 1 is a schematic diagram of an AI inference board provided by an exemplary embodiment of the present application. Figure 4 As shown, the AI ​​inference board 400 is mainly composed of two parts: off-chip (outside the AI ​​inference chip 402) and on-chip (inside the AI ​​inference chip 402).

[0074] For off-chip (on-chip):

[0075] It mainly includes a peripheral second management processor 4012 (e.g., a microcontroller unit), a memory 411 (e.g., a double data rate synchronous dynamic random access memory (DDR)), a fan 412, and a baseboard management controller (Baseboard Management Controller, BMC) 413, etc. Among them, the second management processor 4012 is used to detect the historical operating parameters (e.g., temperature, voltage, etc.) of the off-chip components, and input the detected data to the AI ​​reasoning chip 402 (also called AIDIE) through the serial port of the second management processor 4012.

[0076] For the film:

[0077] The AI ​​reasoning chip 402 is the main component for implementing AI reasoning and training, in which the first management processor 4011 is used to manage the components in the AI ​​board, such as responsible for system startup (Boot), initialization, system maintenance and power consumption control. The system scheduling processor 404 is used to perform system scheduling (scheduling of on-chip components), such as scheduling tasks issued by the host (HOST) to different hardware. The core scheduling processor 405 is used to perform scheduling and synchronization resource allocation for the computing core 403 and the buffer (DMA) 408. The computing core 403 is used to complete AI task calculations and some computer vision (Computer Vision, CV) calculations. This process involves an image computing group 406 and a video decoder 407.

[0078] The above-mentioned AI reasoning board provided in the embodiment of the present application can improve the performance while saving the energy consumption of the AI ​​reasoning board by adjusting the performance of the components in the AI ​​reasoning board. For the specific implementation of adjusting the performance of the components in the AI ​​reasoning board, refer to the relevant contents in the aforementioned embodiments, and the embodiments of the present application will not be repeated here. In combination with the aforementioned embodiments, it can be seen that the performance adjustment of the components is mainly divided into two parts: on-chip and off-chip to perform adjustment control. After the predicted current operating parameters are obtained by the first management processor on the chip, the components on the chip are regulated, while the first management processor on the chip needs to send the predicted current operating parameters to the second management processor outside the chip through the serial port, and then the components outside the chip are regulated. Model prediction is implemented by the first AI core in the AI ​​reasoning chip on the chip. The scheme of adjusting the performance of the components of the AI ​​reasoning board can improve the operating efficiency of the board and reduce energy consumption, and can also improve the processing power and response speed of the board to provide a better user experience.

[0079] Combined with reference Figure 3 and Figure 4 , the system software scheduling framework of server boards is introduced. Figure 5 FIG. 1 is a schematic diagram of a system software scheduling framework of a server card provided by an exemplary embodiment of the present application. Figure 5 As shown, the MCU 501 inside the computing chip includes five parts: a core scheduling processor, a system scheduling processor, a first management processor, a high-performance image computing group, and a video decoder. After the server board is powered on, the following scheduling steps will be performed:

[0080] (1) After power-on, the first management processor executes a boot program (RomCode) and loads the first firmware to initialize the phase-locked loop (PLL) / fuse (FUSE), peripheral component interconnect (PCI) bus, and high-bandwidth memory (HBM);

[0081] (2) The first management processor waits to receive the second part of the firmware sent by the driver, and executes the second part of the firmware to initialize the system scheduling processor, the core scheduling processor, the high-performance image computing group, and the video decoder;

[0082] (3) The first management processor initializes other peripherals (such as general purpose input and output (GPIO), bidirectional two-wire synchronous serial bus (I2C), mailbox (MailBox), watchdog timer (WDT), etc.);

[0083] (4) After receiving and loading other IP firmware sent by the driver, the first management processor releases a reset signal of other intellectual property (IP) cores to enable them to operate normally;

[0084] (5) Finally, the first management processor cyclically monitors and reports the historical operating parameters of the components in the server board, such as temperature parameters, load, etc.

[0085] Combined with reference Figure 3 and Figure 4 , the system runtime scheduling framework of the server board is introduced. Figure 6 FIG. 1 is a schematic diagram of a system runtime scheduling framework of a server card provided by an exemplary embodiment of the present application. Figure 6 As shown, the upper layer software encapsulates the task into a command packet. After receiving the relevant packet, the system scheduling processor 602 parses and executes the following steps:

[0086] (1) Runtime calls the driver interface to send commands, such as notifying the computing chip to enable a service that dynamically adjusts the performance of a component;

[0087] (2) The driver interface detects the status of the task stream 601, selects an idle task stream 601 to send a command, and notifies the system scheduling processor 602 that there is a new instruction;

[0088] (3) The system scheduling processor 602 traverses all task flows 601 and then sends the received commands to the corresponding hardware;

[0089] (4) According to the hardware information in the instruction packet issued by the system scheduling processor 602, the corresponding hardware in the high-performance image computing group, the video decoder, the core scheduling processor, and the computing core is selected to execute the instruction issued by the system scheduling processor 602;

[0090] (5) After the command is executed, the execution result is returned.

[0091] In combination with the introduction of the server board in the foregoing embodiment, the performance adjustment method of the server board provided in the embodiment of the present application is introduced below. Figure 7 FIG. 1 is a schematic diagram of a process for adjusting the performance of a server card provided by an exemplary embodiment of the present application. Figure 7 As shown, a computing chip, a first management processor 702, a second management processor 705, a system scheduling processor 703 and a core scheduling processor are provided in the server board. The computing chip is used to perform computing tasks, and the computing chip includes multiple computing cores 704. The first management processor 702 is used to manage the components 701 within the chip of the computing chip, and the second management processor 705 is used to manage the components 701 outside the chip of the computing chip. The system scheduling processor 703 is used to schedule the components 701 in the server board, and the core scheduling processor is used to schedule multiple computing cores 704.

[0092] In the process of adjusting the performance of the server board, the first management processor 702 will obtain the first historical operating parameters of the component 701 in the chip, and the first management processor 702 will also obtain the second historical operating parameters of the component 701 outside the chip through the second management processor 705. Among them, the above historical operating parameters include parameters used to reflect the operating performance of the component 701. After that, the first management processor 702 will perform scheduling, thereby predicting the current operating parameters of the component 701 in the chip according to the first historical operating parameters, and predicting the current operating parameters of the component 701 outside the chip according to the second historical operating parameters. The current operating parameters are used to characterize the operating parameters corresponding to the current moment in the operating parameter change trend corresponding to the historical operating parameters. In the case where the current operating parameters are the operating parameters of the component 701 in the chip, the first management processor 702 will adjust the performance of the component 701 in the chip according to the current operating parameters; in the case where the current operating parameters are the operating parameters of the component 701 outside the chip, the first management processor 702 will transmit the current operating parameters of the component 701 outside the chip to the second management processor 705, and the second management processor 705 will adjust the performance of the component 701 outside the chip according to the current operating parameters.

[0093] During the operation parameter prediction process scheduled by the first management processor 702, the system scheduling processor 703 will determine whether a prediction model is deployed in the first computing core 704. If the prediction model is not deployed in the first computing core 704, the model needs to be trained. The first management processor 702 will transmit the above historical operation parameters (first historical operation parameters and / or first historical operation parameters) to the system scheduling processor 703, and the system scheduling processor 703 will pre-process the historical operation parameters to obtain training samples and training labels. After that, the system scheduling processor 703 transmits the training samples and training labels to the core scheduling processor, and the core scheduling processor transmits the training samples and training labels to the second computing core 704. The second computing core 704 will train the prediction model to be trained deployed in the second computing core 704 through the training samples and training labels. After that, the core scheduling processor will transmit the trained prediction model to the first computing core 704. When the prediction model training is completed and the prediction model is deployed in the first computing core 704, the prediction model can be called by the first computing core 704 to predict the current operation parameters according to the historical operation parameters through the prediction model.

[0094] The current operating parameters of the components in the server board are predicted by collecting the historical operating parameters of the components. Since the current operating parameters are determined based on the operating parameter change trend corresponding to the historical operating parameters, the predicted current operating parameters can reflect the real-time performance requirements of the components in the server board. By adjusting the performance of the components in the server board based on the current operating parameters, the performance of the components in the server board can be dynamically adjusted according to the real-time operating status of the server board, breaking the limitation of adjusting the board performance based on the preset performance level, and adjusting the components to the appropriate performance mode in advance, achieving more fine-grained performance adjustment, balancing the power usage and performance of the server board, and enhancing the responsiveness of the server board when facing different performance requirements.

[0095] Figure 8 FIG. 1 is a flow chart of a method for adjusting the performance of a server card provided by an exemplary embodiment of the present application. The method can be used for a server card. Figure 8 As shown, the method includes:

[0096] Step 802: Obtain historical operating parameters of components in the server board.

[0097] The historical operating parameters include parameters for reflecting the operating performance of the component. The components in the server board include one or more hardware for supporting the operation of the server board. In some embodiments, the server board includes a computing chip, which is used to perform computing-related tasks, and the components in the server board include a first component in the chip of the computing chip and a second component outside the chip of the computing chip. For example, the above-mentioned components include at least one of the hardware such as CPU, DDR, HCVG, VDEC, BMC, flash memory, fan, etc., and the historical operating parameters of the component include the power state of the component in the historical period, such as sleep state, standby state, running state, etc., and the performance state of the component in the historical period, such as operating frequency, read and write speed, voltage, temperature, etc. The historical period is set by the developer. In some embodiments, the server board also obtains the number of currently executed tasks and the priority of the tasks to predict the current operating parameters of the component based on this.

[0098] Step 804: Predict the current operating parameters of the component based on the historical operating parameters of the component.

[0099] The current operating parameters are used to characterize the operating parameters corresponding to the current moment in the operating parameter change trend corresponding to the historical operating parameters. That is, the current operating parameters of the component are obtained based on the predicted change trend of the historical operating parameters of the component. In some embodiments, the operating parameter change trend corresponding to the historical operating parameters is a continuous value within a certain numerical range, so the current operating parameter is an arbitrary value within a continuous numerical range. That is, the current operating parameter is not one of a plurality of preset values, but is random within a certain numerical range. For example, if the parameter change trend of the historical operating parameters is that the performance gradually decreases, then the performance corresponding to the current operating parameters is lower than the historical operating parameters; if the parameter change trend of the historical operating parameters is that the performance gradually increases, then the performance corresponding to the current operating parameters is higher than the historical operating parameters.

[0100] Step 806: Adjust the performance of the component according to the current operating parameters of the component.

[0101] After predicting the current operating parameters of the component, the server board will adjust the operating parameters of the component to the current operating parameters, and use this to control the operation of the component, thereby adjusting the performance of the component.

[0102] In summary, the method provided in this embodiment predicts the current operating parameters of the components in the server board by collecting the historical operating parameters of the components. Since the current operating parameters are determined based on the operating parameter change trend corresponding to the historical operating parameters, the predicted current operating parameters can reflect the real-time performance requirements of the components in the server board. By adjusting the performance of the components in the server board based on the current operating parameters, the performance of the components in the server board can be dynamically adjusted according to the real-time operating status of the server board, breaking the limitation of adjusting the board performance based on the preset performance level, and adjusting the components to the appropriate performance mode in advance, achieving more fine-grained performance adjustment, balancing the power usage and performance of the server board, and enhancing the responsiveness of the server board when facing different performance requirements.

[0103] Fig. 9 FIG. 1 is a flow chart of a method for adjusting the performance of a server card provided by an exemplary embodiment of the present application. The method can be used for a server card. Fig. 9 As shown, the method includes:

[0104] Step 902: Obtain historical operating parameters of components in the server board.

[0105] The historical operation parameters include parameters for reflecting the operating performance of the component. The historical operation parameters include parameters for reflecting the operating performance of the component. The component in the server board includes one or more hardware for supporting the operation of the server board. Optionally, the historical operation parameters of the component include at least one of the power state of the component in the historical period and the performance state of the component in the historical period. The power state is used to reflect the power consumption of the component, for example, the power state includes at least one of the sleep state, the standby state, and the operating state, the power consumption of the sleep state is lower than the power consumption of the standby state, and the power consumption of the standby state is lower than the power consumption of the operating state. The performance state is used to reflect the release of the operating performance of the component, for example, the performance state includes at least one of the operating frequency, read and write speed, voltage, and temperature of the component.

[0106] Optionally, a computing chip is provided in the server board, and the computing chip is used to perform computing-related tasks. The server board includes a first component that is inside the computing chip, and a second component that is outside the computing chip. When acquiring the historical operating parameters of the components, the server board acquires at least one of a first historical operating parameter of the first component and a second historical operating parameter of the second component.

[0107] Step 904: Train the prediction model to be trained through the second computing core to deploy the prediction model in the first computing core.

[0108] In some embodiments, a computing chip is provided in the server board, and the computing chip includes multiple computing cores, and the computing cores are used to perform computing-related tasks. The above-mentioned multiple computing cores include a first computing core, and the first computing core is any computing core among the multiple computing cores. The first computing core among the multiple computing cores is used to run a prediction model, and the prediction model is an AI algorithm model. Optionally, the prediction model is used to process at least one of a regression task and a classification task, for example, a regression task is used to predict the utilization rate of a component (such as a CPU), and a classification task is used to predict whether the component needs to enter a sleep mode. The first computing core is used to run a prediction model to predict the current operating parameters based on historical operating parameters. The current operating parameters are used to characterize the operating parameters corresponding to the current moment in the operating parameter change trend corresponding to the historical operating parameters. That is, the current operating parameters of the component are obtained based on the predicted change trend of the historical operating parameters of the component.

[0109] Optionally, the multiple computing cores further include a second computing core, which is any computing core among the multiple computing cores, and the second computing core is different from the first computing core. When the prediction model is not deployed in the first computing core, for example, when the server board is powered on again, the server board deploys the prediction model in the first computing core through the second computing core, thereby running the prediction model through the first computing core to predict the current operating parameters according to the historical operating parameters.

[0110] The prediction model to be trained is deployed in the second computing core. The server board can obtain training samples and training labels by preprocessing the historical operating parameters. Then, the prediction model to be trained is trained by the second computing core according to the training samples and training labels, so that the trained prediction model can be transmitted to the first computing core. Optionally, the server board performs time series feature extraction on the historical operating parameters to obtain training samples. Linear regression modeling is performed on the historical operating parameters to obtain training labels.

[0111] For the preprocessing process of historical operating parameters, since the historical operating parameters are sampled at the sampling time within the historical period, they are time series, so it is possible to perform time series feature extraction on the historical operating parameters to obtain training samples. For example, by performing time series feature extraction on the historical operating parameters, the mean and standard deviation of the historical operating parameters can be obtained, which can reflect the short-term change trend of the historical operating parameters and obtain training samples. By performing linear regression modeling on the historical operating parameters, training labels can be obtained. The long-term change trend of the historical operating parameters can be reflected according to the slope and intercept of the linear regression modeling, so that the training labels can be obtained. Optionally, linear regression modeling refers to performing linear regression analysis on the historical operating parameters sampled at different sampling times within the historical period, that is, fitting the historical operating parameters sampled at different sampling times to obtain the linear change relationship of the historical operating parameters over time, and the training labels can be determined according to the linear change relationship, for example, the operating parameters corresponding to a certain moment after the historical period in the linear change relationship are determined as training labels.

[0112] After predicting the current operating parameters of the component, the server board will adjust the operating parameters of the component to the current operating parameters, and use this to control the operation of the component, thereby adjusting the performance of the component.

[0113] In summary, the method provided in this embodiment trains a prediction model by collecting historical operating parameters of components in the server board, thereby predicting the current operating parameters of the components. Since the current operating parameters are determined based on the operating parameter change trend corresponding to the historical operating parameters, the predicted current operating parameters can reflect the real-time performance requirements of the components in the server board. By adjusting the performance of the components in the server board based on the current operating parameters, it is possible to dynamically adjust the performance of the components in the server board according to the real-time operating status of the server board, breaking the limitation of adjusting the board performance based on the preset performance level, and adjusting the components to the appropriate performance mode in advance, achieving more fine-grained performance adjustment, balancing the power usage and performance of the server board, and enhancing the responsiveness of the server board when facing different performance requirements.

[0114] The method provided in this embodiment also deploys the prediction model in the first computing core through the second computing core, thereby realizing the deployment of the prediction model without external information.

[0115] The method provided in this embodiment also trains the prediction model to be trained through the second computing core, providing a solution for efficiently training the model through the computing core inside the server board.

[0116] The method provided in this embodiment also provides a solution for efficiently determining training data of a prediction model by preprocessing historical operating parameters of a component.

[0117] Fig.10 FIG. 1 is a flow chart of a method for adjusting the performance of a server card provided by an exemplary embodiment of the present application. The method can be used for a server card. Fig.10 As shown, the method includes:

[0118] Step 1002: Obtain historical operating parameters of components in the server board.

[0119] The historical operation parameters include parameters for reflecting the operating performance of the component. The historical operation parameters include parameters for reflecting the operating performance of the component. The component in the server board includes one or more hardware for supporting the operation of the server board. Optionally, the historical operation parameters of the component include at least one of the power state of the component in the historical period and the performance state of the component in the historical period. The power state is used to reflect the power consumption of the component, for example, the power state includes at least one of the sleep state, the standby state, and the operating state, the power consumption of the sleep state is lower than the power consumption of the standby state, and the power consumption of the standby state is lower than the power consumption of the operating state. The performance state is used to reflect the release of the operating performance of the component, for example, the performance state includes at least one of the operating frequency, read and write speed, voltage, and temperature of the component.

[0120] Optionally, a computing chip is provided in the server board, and the computing chip is used to perform computing-related tasks. The server board includes a first component that is inside the computing chip, and a second component that is outside the computing chip. When acquiring the historical operating parameters of the components, the server board acquires at least one of a first historical operating parameter of the first component and a second historical operating parameter of the second component.

[0121] Step 1004: Run the prediction model through the first computing core to predict the current operating parameters based on the historical operating parameters of the component.

[0122] The current operating parameters are used to characterize the operating parameters corresponding to the current moment in the operating parameter change trend corresponding to the historical operating parameters. That is, the current operating parameters of the component are obtained based on the predicted change trend of the historical operating parameters of the component.

[0123] In some embodiments, a computing chip is provided in the server board, and the computing chip includes multiple computing cores, and the computing cores are used to perform computing-related tasks. The above-mentioned multiple computing cores include a first computing core, and the first computing core is any computing core among the multiple computing cores. The first computing core among the multiple computing cores is used to run a prediction model, and the prediction model is an AI algorithm model. Optionally, the prediction model is used to process at least one of a regression task and a classification task, for example, a regression task is used to predict the utilization rate of a component (such as a CPU), and a classification task is used to predict whether the component needs to enter a sleep mode. The first computing core is used to run a prediction model to predict current operating parameters based on historical operating parameters.

[0124] In the case where the server card includes a first component and a second component, the server card predicts at least one of a current operating parameter of the first component based on the first historical operating parameter and a current operating parameter of the second component based on the second historical operating parameter.

[0125] Optionally, when a prediction model is deployed in the first computing core, the server board runs the prediction model through the first computing core to predict the current operating parameters based on the historical operating parameters. The implementation of deploying the prediction model in the first computing core can refer to the relevant content in the aforementioned embodiment, and the embodiment of the present application is not repeated here.

[0126] Step 1006: Adjust the performance of the component according to the current operating parameters of the component.

[0127] After predicting the current operating parameters of the component, the server board will adjust the operating parameters of the component to the current operating parameters, and use this to control the operation of the component, thereby adjusting the performance of the component.

[0128] In the case where the server card includes a first component and a second component, if the current operating parameter is the operating parameter of the first component, the server card will adjust the performance of the first component according to the current operating parameter. In the case where the current operating parameter is the operating parameter of the second component, the server card will adjust the performance of the second component according to the current operating parameter.

[0129] Optionally, when the server board receives the first instruction, the server board starts executing the steps of obtaining historical operating parameters, predicting current operating parameters based on historical operating parameters, and adjusting the performance of the component based on the current operating parameters. When the server board receives the second instruction, the server board stops executing the steps of obtaining historical operating parameters, predicting current operating parameters based on historical operating parameters, and adjusting the performance of the component based on the current operating parameters.

[0130] In summary, the method provided in this embodiment predicts the current operating parameters of the components in the server board by collecting the historical operating parameters of the components. Since the current operating parameters are determined based on the operating parameter change trend corresponding to the historical operating parameters, the predicted current operating parameters can reflect the real-time performance requirements of the components in the server board. By adjusting the performance of the components in the server board based on the current operating parameters, the performance of the components in the server board can be dynamically adjusted according to the real-time operating status of the server board, breaking the limitation of adjusting the board performance based on the preset performance level, and adjusting the components to the appropriate performance mode in advance, achieving more fine-grained performance adjustment, balancing the power usage and performance of the server board, and enhancing the responsiveness of the server board when facing different performance requirements.

[0131] The method provided in this embodiment also predicts the current operating parameters of the component by running the prediction model through the first computing core. Since the first computing core is located inside the server board, the speed of predicting the current operating parameters of the component can be improved.

[0132] The method provided in this embodiment also obtains historical operating parameters of components inside and outside the computing chip to predict corresponding current operating parameters, thereby achieving comprehensive prediction of current operating parameters of components in the server board.

[0133] The method provided in this embodiment also adjusts the performance of corresponding components based on the current operating parameters predicted for different components, thereby achieving accurate adjustment of the performance of the components.

[0134] The method provided in this embodiment also detects the first instruction and the second instruction to start or shut down the process of dynamically adjusting the performance of the component, thereby realizing the process of whether to execute the performance of the adjustment component can be customized as needed.

[0135] The method provided in the embodiment of the present application can dynamically adjust the power consumption and performance of the server board (components of the server board) according to the real-time status and usage of the server board. Fig.11 It is a schematic diagram of a process of adjusting the performance of components of a server board provided by an exemplary embodiment of the present application.

[0136] like Fig.11 As shown:

[0137] In steps S1-S4, after the server board is powered on, the boot program (Romcode) will be run first to initialize the first management processor and run its corresponding firmware. After the first management processor works normally, it will receive and run the firmware of other IP cores such as the system scheduling processor, the core scheduling processor, the high-performance image computing group, and the video decoder. The first management processor will be in the system detection management state. The system scheduling processor will handle the task scheduling at the chip level. The upper-level software will encapsulate the task into a command packet. After receiving the relevant packet, the system scheduling processor will parse and execute it. The core scheduling processor is responsible for the release of the buffer, computing core, GSYNC (Good Synchronize) / CSYNC (Composite Synchronize) (used for task synchronization in the computing core to ensure that the next task is executed synchronously after the previous task is completed) resources and schedule the data flow synchronization of the buffer / computing core.

[0138] When the server board is ready, the first management processor will parse the instruction sent by the operating system (OS) to enable dynamic performance adjustment, and determine whether there is a deployed and trained prediction model in the first computing core of the current board. If the current server board is powered on for the first time, the prediction model will not be detected. At this time, the second computing core needs to be used to train the prediction model. The specific process is as follows.

[0139] In steps S5-S12, the first management processor reads the historical operating parameters of the hardware layer's CPU, DDR, HCVG, VDEC, BMC, flash memory, fan and other hardware, including power status (such as sleep, standby, and run), performance status (such as processor / memory frequency, read and write speed, voltage, etc.), collects the read data and sends it to the system scheduling processor for preprocessing.

[0140] The system scheduling processor will use the time series feature extraction method to pre-process the data. This method can mainly reflect the short-term trend based on the mean and standard deviation of the time data, and the long-term trend based on the slope and intercept of the linear regression of the data. Finally, the collected data can be pre-processed through the above method to obtain training samples and training labels for subsequent analysis and modeling. After the data is processed, the system scheduling processor sends it to the core scheduling processor, which uses acceleration engines such as the Packet Processing Engine (PPE) to quickly send the data to the second computing core.

[0141] The second computing core is only a computing core in the server board, and the second computing core can be used to train the prediction model. The ultimate goal of the prediction model is to predict the current operating parameters (power usage and performance requirements) in the future (at the moment) based on the historical operating parameters of the components of the server board (reflecting the real-time status and usage), so as to provide suggestions for the power saving and performance adjustment of the components. This can be a regression task (for example, predicting the CPU usage) or a classification task (for example, predicting whether the device needs to enter sleep mode). The method of training the model uses the TensorFlow framework of deep learning to implement model training. In the TensorFlow framework, model training is completed by using the fit method, which uses the provided training samples and training labels to update the weights of the model. The training data is usually provided in the form of NumPy arrays, because NumPy arrays can efficiently store and operate large amounts of numerical data. For example, if you have a dataset with 1000 training examples, each with 10 features, then the training data will be a NumPy array of shape (1000, 10) and the training labels will be a NumPy array of shape (1000).

[0142] During the training process, several key parameters need to be specified:

[0143] Batch Size: This is the number of training samples used to update the model weights in each model iteration. Choosing the right batch size is crucial to both training efficiency and model performance. A smaller batch size may slow down the training process, but it helps the model learn more carefully and sometimes achieves better generalization performance; a larger batch size may speed up training but may result in poor model performance. Therefore, the choice of batch size needs to be determined based on the specific data set and hardware conditions.

[0144] Epochs: This represents the number of times the model will completely traverse the entire dataset. Choosing an appropriate number of epochs is also important, because too few epochs may cause the model to underfit, while too many epochs may cause overfitting. Therefore, the setting of the number of epochs needs to be determined based on the size and complexity of the dataset, as well as the learning ability of the model.

[0145] Validation Data: At the end of each training cycle, providing validation data allows you to evaluate the performance of the model, which not only helps in managing the training process but can also be used to tune the model parameters to avoid overfitting.

[0146] The following is an example code to train a model using the fit method:

[0147] model.fit(x_train,y_train,batch_size=32,epochs=10,validation_data=(x_val,y_val))

[0148] In this code, x_train and y_train represent training samples and training labels respectively, batch_size=32 sets the number of samples used for each weight update to 32, epochs=10 sets the number of times the model will traverse the data set to 10 times, and (x_val, y_val) is the verification data used to evaluate the performance of the model. After the training is completed and verified, the trained model can be deployed to the first computing core through the core scheduling processor to predict the current operating parameters (reflecting power consumption and performance) in real time.

[0149] In steps S13-S24, if a trained prediction model is deployed in the first computing core, the collected data can be predicted in real time. The specific process is: the first management processor will continue to collect the historical operating parameters of the components in the server board, including the data of the hardware on and off the chip, and the subsequent method is shown above. The system scheduling processor pre-processes the data and then sends it to the core scheduling processor. The core scheduling processor quickly sends the data to the first computing core through the PPE. The prediction model in the first computing core will predict the input data, and then send the prediction results to the first management processor through the interruption of the mailbox (MailBox). The first management processor will make a judgment based on the predicted data, specifically whether it belongs to an on-chip or off-chip component. If it is an on-chip component, it will be uniformly managed by the first management processor, and the predicted data will be set through the driver interface of the server board (such as the operating frequency of HCVG and VDEC on the chip, the allocation of RAM, etc.). The off-chip data needs to be passed to the second management processor outside the chip through the serial port. The second management processor outside the chip will set the operating parameters of the corresponding component according to the data (such as the speed of the fan, the frequency of DDR and other related information), so that the operating parameters of the server board components in the next stage are adjusted according to the prediction results. In addition, it is necessary to evaluate the accuracy of the prediction results. The actual value after adjustment can be read through PLMP and compared with the preset value. Evaluating the accuracy of the prediction results can help the system understand the performance of the model and adjust the model parameters when necessary. The above real-time prediction process can help the system more intelligently adjust the performance of the components, such as the current CPU usage, battery power, and management resources, thereby improving the operating efficiency and response speed of the server board.

[0150] In addition, each time the performance is dynamically adjusted, the first management processor will detect whether the OS has transmitted an instruction packet to stop the dynamic performance adjustment. If the instruction packet is received, it means that the current server board does not need performance adjustment (for example, some test cases need to be run at full performance, etc.), and the adjustment process will be closed at this time. If the instruction packet is not received, the data will be collected cyclically to perform prediction and dynamically adjust the operating parameters of the components of the entire server board to improve the performance of the board and reduce the energy consumption of the board.

[0151] For AI prediction scenarios:

[0152] In an optional example, the server board provided in this application is used for a classification task scenario as an example for introduction. Fig.12 is a block diagram of a computer system provided by an exemplary embodiment of the present application. Fig.12 and Figure 3 The computer system includes a terminal 121 and a server 122, wherein the server board 100 provided in the present application is provided in the server 122. The terminal 121 and the server 122 are connected via a communication network 123. In some embodiments, a client involved in the classification task is deployed in the terminal 121, and the terminal 121 communicates with the server 122 through the client.

[0153] The terminal 121 is used to send the data to be classified to the server 122, and the data to be classified includes data to be classified. After receiving the data to be classified, the server 122 will transmit the data to be classified to the core scheduling processor 105 through the system scheduling processor 104 in the server board 100, and then the core scheduling processor 105 will transmit the data to be classified to the computing core 103 deployed with the classification model. The classification model is, for example, a pre-trained machine learning model that supports the classification of input data. The computing core 103 will run the classification model to perform predictions on the data to be classified, thereby obtaining the corresponding classification results and returning them. After that, the server 122 will send the classification results corresponding to the data to be classified to the terminal 121.

[0154] During the process of the server 122 executing the classification task, if the process of dynamically adjusting the performance is started, the first management processor 1011 of the server board 100 will obtain the historical operating parameters of the components on the chip of the computing chip 102, and obtain the historical operating parameters of the components outside the chip of the computing chip 102 through the second management processor 1012, and transmit them to the system scheduling processor 104. The system scheduling processor 104 will transmit the historical operating parameters to the first computing core, and the first computing core will run the prediction model to obtain the current operating parameters of the components, and feed them back to the first management processor 1011. For the current operating parameters of the components on the chip of the computing chip 102, the first management processor 1011 will set them as the operating parameters of the components on the chip. For the current operating parameters of the components outside the chip of the computing chip 102, the first management processor 1011 will transmit them to the second management processor 1012, and the second management processor 1012 will set them as the operating parameters of the components outside the chip.

[0155] For AI training scenarios:

[0156] In an optional example, the server board provided in this application is used for a model training scenario as an example. Fig.13 is a block diagram of a computer system provided by an exemplary embodiment of the present application. Fig.13 and Figure 3 The computer system includes a terminal 131 and a server 132. The server board 100 provided by the present application is disposed in the server 132. The terminal 131 and the server 132 are connected via a communication network 133.

[0157] The terminal 131 is used to send the code of the machine learning model and the training data to the server 132, and the training data is used to train the machine learning model. After receiving the above information, the server 132 will transmit the training data and the code of the machine learning model to the core scheduling processor 105 through the system scheduling processor 104 in the server board 100, and then the core scheduling processor 105 will deploy the machine learning model to the computing core 103 and transmit the training data to the computing core 103. The computing core 103 will train the machine learning model with the training data. After that, the server 132 will send the training results of the machine learning model to the terminal 131.

[0158] During the process of executing the training of the machine learning model by the server 132, if the process of dynamically adjusting the performance is started, the first management processor 1011 of the server board 100 will obtain the historical operating parameters of the components on the chip of the computing chip 102, and obtain the historical operating parameters of the components outside the chip of the computing chip 102 through the second management processor 1012, and transmit them to the system scheduling processor 104. The system scheduling processor 104 will transmit the historical operating parameters to the first computing core, and the first computing core will run the prediction model to obtain the current operating parameters of the components, and feed them back to the first management processor 1011. For the current operating parameters of the components on the chip of the computing chip 102, the first management processor 1011 will set them as the operating parameters of the components on the chip. For the current operating parameters of the components outside the chip of the computing chip 102, the first management processor 1011 will transmit them to the second management processor 1012, and the second management processor 1012 will set them as the operating parameters of the components outside the chip.

[0159] For image recognition scenarios:

[0160] In an optional example, the server board provided in this application is used for an image recognition scenario as an example. Fig.14 is a block diagram of a computer system provided by an exemplary embodiment of the present application. Fig.14 and Figure 3 The computer system includes a terminal 141 and a server 142, wherein the server board 100 provided in the present application is provided in the server 142. The terminal 141 and the server 142 are connected via a communication network 143. In some embodiments, a client involved in image recognition is deployed in the terminal 141, and the terminal 141 communicates with the server 142 through the client.

[0161] The terminal 141 is used to send a picture to the server 142. After receiving the picture, the server 142 transmits the picture to the core scheduling processor 105 through the system scheduling processor 104 in the server board 100, and then the core scheduling processor 105 transmits the picture to the high-performance image computing group. The high-performance image computing group can perform classification or target recognition on the picture based on CV. Then the server 142 sends the recognition result corresponding to the picture to the terminal 141.

[0162] During the process of the server 142 executing the image recognition task, if the process of dynamically adjusting the performance is turned on, the first management processor 1011 of the server board 100 will obtain the historical operating parameters of the components on the chip of the computing chip 102, and obtain the historical operating parameters of the components outside the chip of the computing chip 102 through the second management processor 1012, and transmit them to the system scheduling processor 104. The system scheduling processor 104 will transmit the historical operating parameters to the first computing core, and the first computing core will run the prediction model to obtain the current operating parameters of the components, and feed them back to the first management processor 1011. For the current operating parameters of the components on the chip of the computing chip 102, the first management processor 1011 will set them as the operating parameters of the components on the chip. For the current operating parameters of the components outside the chip of the computing chip 102, the first management processor 1011 will transmit them to the second management processor 1012, and the second management processor will set them as the operating parameters of the components outside the chip.

[0163] It should be noted that before collecting relevant data of users and during the process of collecting relevant data of users, this application can display a prompt interface, pop-up window or output voice prompt information, and the prompt interface, pop-up window or voice prompt information is used to prompt the user that its relevant data is currently being collected, so that this application only starts to execute the relevant steps of obtaining relevant data of users after obtaining the confirmation operation issued by the user to the prompt interface or pop-up window, otherwise (that is, when the confirmation operation issued by the user to the prompt interface or pop-up window is not obtained), the relevant steps of obtaining relevant data of users are terminated, that is, the relevant data of users are not obtained. In other words, all user data collected by this application are collected with the consent and authorization of the user, and the collection, use and processing of relevant user data need to comply with the relevant laws, regulations and standards of relevant countries and regions.

[0164] It should be noted that the sequence of the method steps provided in the embodiments of the present application can be appropriately adjusted, and the steps can be increased or decreased accordingly according to the circumstances. Any technician familiar with the technical field can easily think of the changed methods within the technical scope disclosed in the present application, and they should be included in the protection scope of the present application, so they will not be repeated here.

[0165] Fig.15 FIG. 1 is a schematic diagram of a performance adjustment device for a server card provided by an exemplary embodiment of the present application. The device can be used for a server card. Fig.15 As shown, the device comprises:

[0166] An acquisition module 1501 is used to acquire historical operating parameters of components in the server board, wherein the historical operating parameters include parameters used to reflect the operating performance of the components;

[0167] A prediction module 1502, configured to predict the current operating parameters of the component according to the historical operating parameters of the component, wherein the current operating parameters are used to represent the operating parameters corresponding to the current moment in the operating parameter change trend corresponding to the historical operating parameters;

[0168] The adjustment module 1503 is used to adjust the performance of the component according to the current operating parameters of the component.

[0169] In an optional design, a computing chip is provided in the server board, and the computing chip includes multiple computing cores. The first computing core among the multiple computing cores is used to run the prediction model, and the prediction model is an AI algorithm model; the prediction module 1502 is used to run the prediction model through the first computing core to predict the current operating parameters based on the historical operating parameters.

[0170] In an optional design, the multiple computing cores include a second computing core, which is different from the first computing core; the prediction module 1502 is used to run the prediction model through the first computing core to predict the current operating parameters based on the historical operating parameters when the prediction model is deployed in the first computing core; deploy the prediction model in the first computing core through the second computing core when the prediction model is not deployed in the first computing core; and run the prediction model through the first computing core to predict the current operating parameters based on the historical operating parameters.

[0171] In an optional design, the prediction model to be trained is deployed in the second computing core; the prediction module 1502 is used to preprocess the historical operating parameters to obtain training samples and training labels; the prediction model to be trained is trained according to the training samples and the training labels by the second computing core; and the trained prediction model is transmitted to the first computing core.

[0172] In an optional design, the prediction module 1502 is used to perform time series feature extraction on the historical operating parameters to obtain the training samples; and perform linear regression modeling on the historical operating parameters to obtain the training labels.

[0173] In an optional design, a computing chip is provided in the server board, and the server board includes a first component belonging to the computing chip on a chip, and a second component belonging to the computing chip on a chip outside the computing chip; the acquisition module 1501 is used to obtain a first historical operating parameter of the first component; obtain a second historical operating parameter of the second component; the prediction module 1502 is used to predict the current operating parameter of the first component based on the first historical operating parameter; and predict the current operating parameter of the second component based on the second historical operating parameter.

[0174] In an optional design, the adjustment module 1503 is used to adjust the performance of the first element according to the current operating parameters when the current operating parameters are the operating parameters of the first element; and to adjust the performance of the second element according to the current operating parameters when the current operating parameters are the operating parameters of the second element.

[0175] In an alternative design, such as Fig.16 As shown, the device also includes a processing module 1504, which is used to start executing the steps of obtaining the historical operating parameters, predicting the current operating parameters based on the historical operating parameters, and adjusting the performance of the component based on the current operating parameters when the server board receives the first instruction; and stop executing the steps of obtaining the historical operating parameters, predicting the current operating parameters based on the historical operating parameters, and adjusting the performance of the component based on the current operating parameters when the server board receives the second instruction.

[0176] It should be noted that the performance adjustment device for the server board provided in the above embodiment is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the performance adjustment device for the server board provided in the above embodiment and the performance adjustment method embodiment of the server board belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0177] The embodiment of the present application further provides a computer device, which includes the server board provided in the above embodiment. Optionally, the computer device is a server. For example, Fig.17 It is a structural diagram of a computer device provided by an exemplary embodiment of the present application.

[0178] The computer device 1700 includes a central processing unit (CPU) 1701, a system memory 1704 including a random access memory (RAM) 1702 and a read-only memory (ROM) 1703, and a system bus 1705 connecting the system memory 1704 and the central processing unit 1701. The computer device 1700 also includes a basic input / output system (I / O system) 1706 for facilitating information transmission between various components in the computer device, and a large-capacity storage device 1707 for storing an operating system 1713, application programs 1714, and other program modules 1715.

[0179] The basic input / output system 1706 includes a display 1708 for displaying information and an input device 1709 such as a mouse and a keyboard for user inputting information. The display 1708 and the input device 1709 are connected to the central processing unit 1701 through an input / output controller 1710 connected to the system bus 1705. The basic input / output system 1706 may also include an input / output controller 1710 for receiving and processing inputs from a plurality of other devices such as a keyboard, a mouse, or an electronic stylus. Similarly, the input / output controller 1710 also provides output to a display screen, a printer, or other types of output devices.

[0180] The mass storage device 1707 is connected to the central processing unit 1701 via a mass storage controller (not shown) connected to the system bus 1705. The mass storage device 1707 and its associated computer readable storage medium provide non-volatile storage for the computer device 1700. That is, the mass storage device 1707 may include a computer readable storage medium (not shown) such as a hard disk or a compact disc read-only memory (CD-ROM) drive.

[0181] Without loss of generality, the computer-readable storage medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable storage instructions, data structures, program modules or other data. Computer storage media include RAM, ROM, Erasable Programmable Read Only Memory (EPROM), Electronically Erasable Programmable Read-Only Memory (EEPROM), flash memory or other solid-state storage devices, CD-ROM, Digital Versatile Disc (DVD) or other optical storage, cassettes, tapes, disk storage or other magnetic storage devices. Of course, those skilled in the art will know that the computer storage medium is not limited to the above. The above-mentioned system memory 1704 and mass storage device 1707 can be collectively referred to as memory.

[0182] The memory stores one or more programs, and the one or more programs are configured to be executed by one or more central processing units 1701. The one or more programs contain instructions for implementing the above-mentioned method embodiments. The central processing unit 1701 executes the one or more programs to implement the methods provided by the above-mentioned various method embodiments.

[0183] According to various embodiments of the present application, the computer device 1700 can also be connected to a remote computer device on a network through a network such as the Internet. That is, the computer device 1700 can be connected to a network 1712 through a network interface unit 1711 connected to the system bus 1705, or the network interface unit 1711 can be used to connect to other types of networks or remote computer device systems (not shown).

[0184] The memory also includes one or more programs, which are stored in the memory and include steps executed by a computer device in the method provided in the embodiment of the present application.

[0185] The embodiment of the present application also provides a chip, the chip includes a management processor and a computing chip, the management processor is used to manage the components in the server board;

[0186] The management processor is used to obtain historical operating parameters of the component, wherein the historical operating parameters include parameters used to reflect the operating performance of the component;

[0187] The computing chip is used to predict the current operating parameters of the component according to the historical operating parameters of the component, and the current operating parameters are used to characterize the operating parameters corresponding to the current moment in the operating parameter change trend corresponding to the historical operating parameters;

[0188] The management processor is further configured to adjust the performance of the component according to the current operating parameters of the component.

[0189] A server is also provided in an embodiment of the present application, and the server includes the server board provided in the above embodiment.

[0190] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described device and each module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0191] The above description is only an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent switching, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A server card, characterized in that: The server board includes a management processor and a computing chip, and the management processor is used to manage the components in the server board; The management processor is used to obtain historical operating parameters of the component, wherein the historical operating parameters include parameters used to reflect the operating performance of the component; The computing chip is used to predict the current operating parameters of the component according to the historical operating parameters of the component, and the current operating parameters are used to characterize the operating parameters corresponding to the current moment in the operating parameter change trend corresponding to the historical operating parameters; The management processor is further configured to adjust the performance of the component according to the current operating parameters of the component.

2. The server board according to claim 1, characterized in that: The computing chip includes multiple computing cores, and a first computing core among the multiple computing cores is used to run a prediction model, and the prediction model is an artificial intelligence AI algorithm model; The first computing core is used to run the prediction model to predict the current operating parameters based on the historical operating parameters.

3. The server board according to claim 2, characterized in that: The plurality of computing cores includes a second computing core, the second computing core being different from the first computing core; In a case where the prediction model is deployed in the first computing core, the first computing core is used to run the prediction model to predict the current operating parameter according to the historical operating parameter; In a case where the prediction model is not deployed in the first computing core, the second computing core is used to deploy the prediction model in the first computing core; The first computing core is used to run the prediction model to predict the current operating parameters based on the historical operating parameters.

4. The server board according to claim 3, characterized in that: The prediction model to be trained is deployed in the second computing core; The second computing core is used to train the prediction model to be trained through training samples and training labels, wherein the training samples and the training labels are obtained by preprocessing the historical operating parameters; The second computing core is also used to transmit the trained prediction model to the first computing core.

5. The server board according to claim 4, characterized in that: The server board also includes a system scheduling processor and a core scheduling processor, wherein the system scheduling processor is used to schedule the components in the computing chip, and the core scheduling processor is used to schedule the multiple computing cores; The system scheduling processor is used to pre-process the historical operating parameters to obtain the training samples and the training labels; The system scheduling processor is further used to transmit the training samples and the training labels to the core scheduling processor; The core scheduling processor is used to transmit the training samples and the training labels to the second computing core; The second computing core is used to transmit the trained prediction model to the first computing core through the core scheduling processor.

6. The server board according to claim 4, characterized in that: The system scheduling processor is used to perform time series feature extraction on the historical operating parameters to obtain the training samples; The system scheduling processor is also used to perform linear regression modeling on the historical operating parameters to obtain the training labels.

7. The server board according to any one of claims 1 to 6, characterized in that: The server board includes a first component that is inside the computing chip and a second component that is outside the computing chip; The management processor is configured to obtain at least one of a first historical operating parameter of the first element and a second historical operating parameter of the second element; The computing chip is used for predicting at least one of the current operating parameters of the first component based on the first historical operating parameters and the current operating parameters of the second component based on the second historical operating parameters.

8. The server board according to claim 7, characterized in that: The management processor includes a first management processor located inside the computing chip and a second management processor located outside the computing chip, the first management processor is used to manage the first element, and the second management processor is used to manage the second element; The first management processor is used to obtain a first historical operating parameter of the first component; The second management processor is used to obtain a second historical operating parameter of the second component.

9. The server board according to claim 8, characterized in that: the second management processor being configured to transmit a second historical operating parameter of the second element to the first management processor; The first management processor is used to transmit the first historical operating parameter to the computing chip so that the computing chip predicts the current operating parameter of the first component; The first management processor is further configured to transmit the second historical operating parameter to the computing chip so that the computing chip can predict the current operating parameter of the second component.

10. The server board according to claim 8, characterized in that: In a case where the current operating parameter is an operating parameter of the first component, the first management processor is configured to adjust the performance of the first component according to the current operating parameter; In the case where the current operating parameter is an operating parameter of the second element, the first management processor is used to transmit the current operating parameter to the second management processor; The second management processor is used to adjust the performance of the second component according to the current operating parameters.

11. A method for adjusting the performance of a server board, characterized in that: The method is applied to the server board according to any one of claims 1 to 10, and the method comprises: Acquire historical operating parameters of components in the server board, wherein the historical operating parameters include parameters used to reflect operating performance of the components; Predicting the current operating parameters of the component according to the historical operating parameters of the component, the current operating parameters being used to characterize the operating parameters corresponding to the current moment in the operating parameter change trend corresponding to the historical operating parameters; The performance of the component is adjusted according to the current operating parameters of the component.

12. The method according to claim 11, characterized in that The server board is provided with a computing chip, the computing chip includes a plurality of computing cores, a first computing core among the plurality of computing cores is used to run a prediction model, and the prediction model is an AI algorithm model; The predicting the current operating parameters of the component according to the historical operating parameters of the component includes: The prediction model is run through the first computing core to predict the current operating parameters based on the historical operating parameters.

13. The method according to claim 12, characterized in that The plurality of computing cores includes a second computing core, the second computing core being different from the first computing core; The running the prediction model by the first computing core to predict the current operating parameter according to the historical operating parameter includes: In a case where the prediction model is deployed in the first computing core, the prediction model is run by the first computing core to predict the current operating parameter according to the historical operating parameter; In the case where the prediction model is not deployed in the first computing core, the prediction model is deployed in the first computing core through the second computing core; and the prediction model is run through the first computing core to predict the current operating parameters based on the historical operating parameters.

14. The method according to claim 13, characterized in that The prediction model to be trained is deployed in the second computing core; The deploying the prediction model in the first computing core through the second computing core includes: Preprocessing the historical operating parameters to obtain training samples and training labels; Training the prediction model to be trained according to the training samples and the training labels by the second computing core; The trained prediction model is transmitted to the first computing core.

15. The method according to claim 14, characterized in that The preprocessing of the historical operating parameters to obtain training samples and training labels includes: Performing time series feature extraction on the historical operating parameters to obtain the training samples; Linear regression modeling is performed on the historical operating parameters to obtain the training labels.

16. The method according to any one of claims 11 to 15, characterized in that: The server board is provided with a computing chip, and the server board includes a first component inside the computing chip and a second component outside the computing chip; The obtaining of historical operating parameters of the components in the server board includes at least one of the following: Acquire a first historical operating parameter of the first element; acquire a second historical operating parameter of the second element; The predicting the current operating parameters of the component according to the historical operating parameters of the component includes at least one of the following: The current operating parameters of the first component are predicted based on the first historical operating parameters; and the current operating parameters of the second component are predicted based on the second historical operating parameters.

17. The method according to claim 16, characterized in that The step of adjusting the performance of the component according to the current operating parameters of the component comprises: In a case where the current operating parameter is an operating parameter of the first element, adjusting the performance of the first element according to the current operating parameter; In a case where the current operating parameter is an operating parameter of the second element, the performance of the second element is adjusted according to the current operating parameter.

18. The method according to any one of claims 11 to 15, characterized in that: The method further comprises: When the server board receives the first instruction, the steps of acquiring the historical operating parameters, predicting the current operating parameters according to the historical operating parameters, and adjusting the performance of the component according to the current operating parameters are started; When the server board receives the second instruction, the steps of obtaining the historical operating parameters, predicting the current operating parameters according to the historical operating parameters, and adjusting the performance of the component according to the current operating parameters are stopped.

19. A chip, characterized in that: The chip includes a management processor and a computing chip, and the management processor is used to manage the components in the server board; The management processor is used to obtain historical operating parameters of the component, wherein the historical operating parameters include parameters used to reflect the operating performance of the component; The computing chip is used to predict the current operating parameters of the component according to the historical operating parameters of the component, and the current operating parameters are used to characterize the operating parameters corresponding to the current moment in the operating parameter change trend corresponding to the historical operating parameters; The management processor is further configured to adjust the performance of the component according to the current operating parameters of the component.

20. A server, characterized in that: The server comprises a server board as claimed in any one of claims 1 to 10.

Images