Method, device and equipment for calculating computing power demand of intelligent computing center based on business

Abstract

The invention provides a business-based intelligent computing center computing power demand measuring and calculating method, device and equipment, and effectively solves the problem that an existing intelligent computing power total demand measuring and calculating method cannot be associated with a business application scene and cannot meet the accuracy demand of an intelligent computing center. The method comprises the following steps: pre-mapping an AI technology of an intelligent calculation center as a processing model, and dividing the processing model into different types according to a business scene of the intelligent calculation center; according to the stage of the processing model, respectively configuring corresponding computing power measuring and calculating networks for different types of processing models; the computing power calculation network is obtained by optimizing an original computing power network; the processing model has different stages; and measuring and calculating type measurement and calculation results of the processing model under different types based on the computing power measurement and calculation network, and fusing the type measurement and calculation results under different stages to obtain a target measurement and calculation result, so as to measure and calculate the computing power demand of the intelligent computing center based on the target measurement and calculation result.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more specifically, to a method, apparatus, and equipment for measuring the computing power requirements of a business-based intelligent computing center. Background Technology

[0002] According to the "Scaling Law," the main factors affecting the computing power consumption of large-scale intelligent models include the number of model parameters, the size of the training dataset, and the computational cost per token per unit parameter. Its computing power requirements can be expressed as the following formula: . While the computational power consumption of a single model can be estimated using Formula 1, at the regional level, due to the diverse types and significant differences in scale of the deployed and applied large models, uniform calculations based solely on a single general model may result in substantial deviations. To improve the scientific rigor and accuracy of the calculations, the intelligent computing center categorizes models based on parameters and their intended use, conducts refined calculations, and ultimately derives the total future intelligent computing power requirement for the region by summing the demands of each category. Therefore, existing technologies provide methods for separately calculating the intelligent computing power consumed in the training and inference phases, and then summing these calculations to obtain the overall intelligent computing power requirement for that type of model. Figure 1 As shown. The intelligent computing power consumption during the training phase can be calculated using the following formula (2); the intelligent computing power consumption during the inference phase is calculated as shown in formula (3): ; ; By summing up the computational power consumption of the training and inference phases, the total intelligent computing power requirement of a certain type of large model throughout its entire lifecycle can be systematically evaluated, as shown in formula (4): This formula comprehensively reflects the total amount of computing resources required for the entire process from model training to deployment and application, providing a complete basis for the system to evaluate the demand for intelligent computing power. However, the above methods do not establish a connection between business application scenarios and AI technology capabilities. The general model cannot reflect business application scenarios and cannot complete the conversion of business scenarios into technical requirements. This results in a large error in the total demand for intelligent computing power calculated based on the above methods, which cannot meet the accuracy requirements of the intelligent computing center. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a business-based method, apparatus and equipment for calculating the computing power requirements of intelligent computing centers, which effectively solves the problem that existing methods for calculating the total intelligent computing power requirements cannot be linked to business application scenarios and cannot meet the accuracy requirements of intelligent computing centers.

[0004] Firstly, embodiments of this application provide a method for calculating the computing power requirements of a smart computing center based on business needs, the method comprising: The AI ​​technology of the intelligent computing center is pre-mapped as a processing model, and the processing model is divided into different types according to the business scenarios of the intelligent computing center; Based on the stage of the processing model, a corresponding computing power measurement network is configured for different types of processing models; the computing power measurement network is obtained by optimizing the original computing power network; the processing model has different stages; Based on the computing power measurement network, the type measurement results of the processing model under different types are calculated, and the type measurement results under different stages are integrated to obtain the target measurement result, so as to calculate the computing power requirement of the intelligent computing center based on the target measurement result.

[0005] In conjunction with the first aspect, this application provides a first possible implementation of the first aspect, wherein configuring corresponding computing power measurement networks for different types of processing models according to the stage in which the processing model is located includes: Based on the lifecycle of the processing model, the processing model is divided into a training phase and an inference phase; Configure corresponding computing power measurement networks for the training and inference phases respectively to measure computing power requirements.

[0006] In conjunction with the first aspect, this application provides a second possible implementation of the first aspect, wherein configuring corresponding computing power measurement networks for the training phase and the inference phase respectively includes: Extract the data features of the processing model in each type, and configure the corresponding type computing power network according to the data features; Based on the aforementioned type of computing network, computing power is calculated for the corresponding processing model.

[0007] In conjunction with the first aspect, this application provides a third possible implementation of the first aspect, wherein the type includes at least an image type; The step of configuring the corresponding type of computing power network based on the data characteristics includes: Set the image pixels as the data feature of the image type processing model, and configure the resolution influence factor; Based on the resolution influence factor, the original computing power network for the image type is optimized to obtain the computing power network for that type.

[0008] In conjunction with the first aspect, this application provides a fourth possible implementation of the first aspect, wherein the type includes at least a voice type; The step of configuring the corresponding type of computing power network based on the data characteristics includes: Set the temporal features of speech as the data features of the speech type processing model, and configure the temporal length influence factor; Based on the time-series length influence factor, the original computing power network for the speech type is optimized to obtain the computing power network for that type.

[0009] In conjunction with the first aspect, this application provides a fifth possible implementation of the first aspect, wherein configuring corresponding computing power measurement networks for the training phase and the inference phase respectively includes: The number of inferences performed by the processing model in the statistical inference stage, in different types of models; The computing power measurement network for the inference stage is obtained by performing calculations based on the number of inferences in the model and the initial computing power results of the corresponding type of processing model.

[0010] In conjunction with the first aspect, this application provides a sixth possible implementation of the first aspect, wherein the step of fusing the type calculation results from different stages to obtain the target calculation result includes: Collect and calculate the type of computing power measurement network output results of different types of processing models at the same stage to obtain the stage measurement results; By integrating the stage measurement results from different stages, the target measurement result of the processing model is obtained.

[0011] Secondly, embodiments of this application provide a business-based intelligent computing center computing power demand measurement device, the device comprising: The mapping module is used to pre-map the AI ​​technology of the intelligent computing center into processing models, and to classify the processing models into different types according to the business scenarios of the intelligent computing center; The configuration module is used to configure corresponding computing power measurement networks for different types of processing models according to the stage in which the processing model is located; the computing power measurement network is obtained by optimizing the original computing power network; the processing model has different stages; The calculation module is used to calculate the type calculation results of the processing model under different types based on the computing power calculation network, and to integrate the type calculation results under different stages to obtain the target calculation result, so as to calculate the computing power requirement of the intelligent computing center based on the target calculation result.

[0012] Thirdly, embodiments of this application provide an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of any one of the methods for calculating the computing power requirements of a business-based intelligent computing center are performed.

[0013] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of any of the methods described in the "Business-Based Intelligent Computing Center Computing Power Demand Calculation Method".

[0014] This application provides a business-based method for calculating the computing power requirements of an intelligent computing center. The method pre-maps the AI ​​technology of the intelligent computing center as a processing model, and categorizes these models into different types based on the center's business scenarios. Then, based on the stage of each processing model, a corresponding computing power calculation network is configured for each type. This network is obtained by optimizing an existing computing power network. The processing models have different stages. Finally, the method calculates the type calculation results of the processing models under different stages based on the computing power calculation network, and integrates these results to obtain a target calculation result. The computing power requirements of the intelligent computing center are then calculated based on this target result. This method links business application scenarios with the computing power calculation of AI technology-corresponding models, improving the accuracy of computing power calculations for processing models, meeting the accuracy requirements of intelligent computing centers, and avoiding errors. This transforms intelligent computing center planning from an art based on empirical speculation into a data-driven science. Only by adhering to business application scenarios and conducting accurate calculations can this significant infrastructure investment in intelligent computing centers truly and effectively empower business innovation and drive the intelligent transformation of industries. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic diagram illustrating existing methods for measuring overall intelligent computing power requirements is shown; Figure 2 A flowchart illustrating a business-based intelligent computing center computing power demand measurement method provided in an embodiment of this application is shown. Figure 3 This paper illustrates another flowchart of a business-based intelligent computing center computing power demand measurement method provided in an embodiment of this application. Figure 4 This paper illustrates a structural block diagram of a business-based intelligent computing center computing power demand measurement device provided in an embodiment of this application. Figure 5 A structural block diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0018] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0019] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0020] Existing methods for calculating total intelligent computing power requirements do not link business application scenarios with AI technology capabilities. General models cannot reflect business application scenarios and cannot convert business scenarios into technical requirements. This results in significant errors in the total intelligent computing power requirements calculated using these methods, failing to meet the accuracy requirements of intelligent computing centers.

[0021] Based on this, this application provides a method, apparatus, and equipment for measuring the computing power requirements of a business-based intelligent computing center, which will be described below through embodiments.

[0022] Example 1 To facilitate understanding of this embodiment, a method for calculating the computing power requirements of a business-based intelligent computing center, disclosed in this application embodiment, will first be described in detail. For example... Figure 2 The diagram shows a flowchart of a business-based intelligent computing center computing power requirement measurement method, as follows: Figure 3 The diagram shows another flowchart of a business-based intelligent computing center computing power requirement calculation method. This application provides a business-based intelligent computing center computing power requirement calculation method, the method comprising: S101. The AI ​​technology of the intelligent computing center is pre-mapped as the processing model, and the processing model is divided into different types according to the business scenarios of the intelligent computing center; S102. Based on the stage of the processing model, configure corresponding computing power measurement networks for different types of processing models; the computing power measurement network is obtained by optimizing the original computing power network; the processing model has different stages; S103. Calculate the type calculation results of the processing model under different types based on the computing power calculation network, and integrate the type calculation results under different stages to obtain the target calculation result, so as to calculate the computing power requirement of the intelligent computing center based on the target calculation result.

[0023] In step S101, this application analyzes the implementation of each business scenario based on the majority of business scenarios currently supported by the intelligent computing center. These scenarios are all based on the intelligent analysis and processing of four types of data: video, image, voice, and text, providing support for single or integrated application scenarios across various industries. Therefore, based on the data processing and analysis capabilities of video, image, voice, and text data provided by the intelligent computing center using AI technology, these capabilities are mapped into processing models. This transforms the vague AI technology capabilities into clear measurement units. Simultaneously, a three-layered fixed correspondence is established between business scenarios, AI technology, and processing models. The measurement rules for each model are clearly defined, forming a mapping table. Based on AI technology, these scenarios are categorized into three types: image, voice, and natural language. This establishes a more structured and accurate framework for measuring computing power requirements. Therefore, the processing models can also be divided into three types: image, voice, and natural language.

[0024] In step S102, based on the training phase and inference phase that are common to most models, the processing model is also divided into a training phase and an inference phase. According to the phase in which the processing model is located, a corresponding computing power measurement network is configured for each type of processing model. The computing power measurement network is obtained by optimizing the original computing power network. The processing model has different phases. The original computing power network for the training phase is the formula (2) in the background technology. The original computing power network for the inference phase is the formula (3) in the background technology. That is, this application obtains a computing power measurement network with higher matching and accuracy by optimizing the original computing power network. The computing power measurement network includes type computing power networks corresponding to three types: image, speech, and natural language.

[0025] In a specific implementation of step S102, one embodiment is as follows: configuring corresponding computing power measurement networks for different types of processing models according to the stage in which the processing model is located includes: S1021. Based on the lifecycle of the processing model, the processing model is divided into a training phase and an inference phase; S1022. Configure corresponding computing power measurement networks for the training phase and the inference phase respectively, so as to perform computing power requirement measurement.

[0026] In steps S1021-S1022, the processing model described in this application has the lifecycle of a general model, namely, a training phase and an inference phase. Therefore, the processing model is divided into a training phase and an inference phase. Corresponding computing power measurement networks are configured for the training phase and the inference phase respectively to perform computing power demand measurement. That is, different phases correspond to different computing power measurement networks. For example, the computing power measurement network for the training phase is represented by formula (5): (5); Based on formula (5), a fundamental shift from technology-capability-oriented to business-value-oriented planning of intelligent computing centers has been achieved for the first time, ensuring that investment and operation of computing infrastructure are always aligned with clear business objectives. The computing power measurement network for the inference phase is represented by formula (6): (6); Where A, B, and C represent the number of processing models for the three types of images, natural language, and speech, respectively. The computing power (inference) of the image processing model is the computing power consumed by the image type computing power network during the inference stage, and so on for the others.

[0027] In a specific implementation of step S1022, one embodiment is as follows: configuring corresponding computing power measurement networks for the training and inference phases respectively includes: A11. Extract the data features of the processing model in each type, and configure the corresponding type computing power network according to the data features; A12. Based on the computing power network of the aforementioned type, computing power is calculated for the corresponding processing model.

[0028] In steps A11-A12, during the training phase, it is necessary to extract the data features of the processing model in each type. Different types have different data features. For example, in the image type, it is necessary to reflect the spatial locality and high-resolution sensitivity of the image based on pixels. In the speech type, it is necessary to reflect the high frame rate and long-range dependency characteristics of the speech based on time sequence. Thus, pixels and time sequence are the data features in the image and speech types, respectively. The corresponding type computing power network is configured according to the data features. Based on the type computing power network, the computing power is measured for the processing model of the corresponding type, thereby calculating the computing power required or consumed by the processing model under each type.

[0029] In the current computing power demand assessment system of intelligent computing centers, since the core calculation formula is mainly designed based on the architectural features of natural language processing models, we continue to use formula (2) as the basic framework to assess the computing power demand in the direction of natural language processing. This formula is built on the basis of deep learning theory. The number of model parameters reflects the structural complexity of the neural network, the size of the training dataset reflects the total amount of information that the model needs to learn, the coefficient 3 corresponds to the basic calculation process of forward propagation and backward propagation, and the computing power demand per unit parameter per token includes the influence factors of the actual operating environment such as hardware architecture and calculation accuracy. Although the formula is concise, it captures the core elements of computing power consumption of natural language processing models and can provide a basic reference for resource planning of intelligent computing centers.

[0030] In a specific implementation of step A1, one embodiment is as follows: the type includes at least an image type; The step of configuring the corresponding type of computing power network based on the data characteristics includes: B1. Set the image pixels as the data feature of the image type processing model, and configure the resolution influence factor; B2. Based on the resolution influence factor, optimize the original computing power network of the image type to obtain the computing power network of the image type.

[0031] In steps B1-B2, the direct application of the NLP token concept in the computational power estimation of the image processing model is a fundamental misuse. Text tokens are discrete semantic units, and their computational cost increases linearly; while image pixels have strong spatial correlations, with adjacent pixels sharing information and establishing complex relationships through convolution or attention mechanisms. The basic formula completely ignores this spatial dependence, incorrectly assuming that pixels are processed independently, leading to serious distortion in computational power estimation at high resolutions. For example, if the resolution is increased from 224p to 4K, the number of pixels increases by 300 times, and the actual computational power requirement may increase by thousands of times, while the original computational power network corresponding to the image type, i.e., formula (2) in the background technology, shows no change, reducing the accuracy of the calculated type measurement results. Therefore, this application uses pixels to reflect the spatial locality and high-resolution sensitivity of images, sets the pixels of the image as the data feature of the image type processing model, and calculates the resolution influence factor based on the pixels. The resolution influence factor is calculated as (H*W)^α, where H*W is the image resolution (number of pixels), and α is the resolution index. Different architectures have different efficiencies in processing spatial relationships. For example, CNN has locality, so it is efficient with a small α; ViT is global attention, so it is inefficient with a large α. α is the tuning knob of this dedicated network, and it is assigned a value according to different models. Table 1 shows the resolution index assigned to different models. The original computing power network of formula (2) is optimized according to the resolution influence factor to obtain the type computing power network corresponding to the image type processing model as shown in formula (7): (7); The unit parameter, computing power requirement per Token, is set to 1.

[0032] Table 1. Resolution index assignment table for different models.

[0033] In the specific implementation of step A1, another embodiment exists in which the type includes at least a voice type; The step of configuring the corresponding type of computing power network based on the data characteristics includes: C1. Set the temporal features of speech as the data features of the speech type processing model, and configure the temporal length influence factor; C2. Based on the time-series length influence factor, optimize the original computing power network for the speech type to obtain the computing power network for that type.

[0034] In steps C1-C2, the original computing network corresponding to formula (2) completely ignores the fundamental impact of speech temporal characteristics on computing power. Speech, as a continuous time series, has extremely high data density—even when converted to Mel spectrograms, it still requires processing 50-100 time steps per second, with sequence lengths often reaching thousands. In contrast, NLP processes sparse discrete symbols with fixed and shorter sequence lengths (e.g., 512 tokens). This difference in magnitude means that to process the same amount of data, speech requires calculating sequence lengths tens of times longer than text, and it must model millisecond-level fine dependencies to capture features such as phonemes and prosody, leading to a severe underestimation of computing power requirements. Therefore, this application sets the temporal features of speech as the data features of the speech type processing model, and configures the temporal length influence factor. The calculation method of the temporal length influence factor is: temporal length influence factor = frame rate * average audio duration * architecture temporal factor, where β is the architecture temporal factor. Different architectures have different capabilities for processing long sequence dependencies. For example, RNN has higher efficiency and smaller β; Transformer has strong capabilities but large computational load and larger β. β is the complexity amplifier of this dedicated network. The architecture temporal factor is assigned according to different models, as shown in Table 2. Based on the temporal length influence factor, the original computing power network of the image type is optimized, i.e., formula (2) to obtain the computing power network of the type, as shown in formula (8). (8); The unit parameter, computing power requirement per Token, is set to 1.

[0035] Table 2 shows the architecture time factor assignment table.

[0036] In the specific implementation of step S1022, another embodiment exists: configuring corresponding computing power measurement networks for the training phase and the inference phase respectively includes: A21. The processing model in the statistical reasoning stage, and the number of inferences performed by the model in different types; A22. Based on the number of inferences in the model and the initial computing power results of the corresponding type of processing model, calculations are performed to obtain the computing power measurement network for the inference stage.

[0037] In steps A21-A22, since the processing models of each type have been configured with type computing power networks in the inference stage, it is necessary to count the number of inference models in the inference stage and the number of inference models in different types. The computing power measurement network for the inference stage is obtained by calculating the number of inference models and the initial computing power result of the corresponding type of processing model. This is shown in formula (6). The initial computing power result is the computing power required or consumed by each processing model for one inference. Compared with formula (3), formula (6) of this application binds the computing power requirement with the specific business value stream, so that the resource allocation of the processing model has a clear service object. If there is only formula (3), the value of classification optimization cannot be reflected, and the measurement result will fall back to the general stage estimation.

[0038] In step S103, the computing power calculation network of the image, speech, and natural language type computing power networks in the training phase of this application calculates the type calculation results of the processing model under different types, thereby obtaining the sum of computing power required by the above three types of computing power networks in the training phase, and the type calculation result of the sum of computing power required by the above three types of computing power networks in the inference phase. Based on the type calculation results in the training phase and the type calculation results in the inference phase, the total computing power required by the processing model is fused to obtain the target calculation result, so as to calculate the computing power requirement of the intelligent computing center based on the target calculation result. Based on this method, this application ensures the accuracy of the calculated computing power required, and avoids the inaccuracies that exist in the original computing power networks corresponding to formulas (2) and (3). In a specific implementation of step S103, one embodiment is as follows: the fusion of type calculation results from different stages to obtain the target calculation result includes: S1031. Collect and calculate the type of calculation results of the computing power measurement network output of different types of processing models in the same stage to obtain the stage calculation results; S1032. Integrate the stage measurement results from different stages to obtain the target measurement result of the processing model.

[0039] In steps S1031-S1032, this application collects and calculates the type calculation results of the computing power calculation network output of different types of processing models at the same stage to obtain the stage calculation results, such as formula (5) and formula (6). The stage calculation results of the training stage and the stage calculation results of the inference stage can be obtained, and the stage calculation results of different stages are fused together. Specifically, the stage calculation results of different stages are added to obtain the target calculation result of the processing model, such as formula (4), thereby obtaining the total computing power required by the processing model.

[0040] This application selected image models such as Swin, ConvNeXt, and DiT X, which have clear data sources, for verification. The relevant models and computing power evaluation results are detailed in Table 3. Among them, the number of parameters is a preset scale of tens of millions or hundreds of millions of parameters. The resolution influence factor is composed of the length, width, and resolution index of the input image. The computing power requirements of the old formula and the new formula are calculated according to formula (2) and formula (7), respectively, with the unit being GFLOPS. The number of parameters, H and W in the resolution influence factor, and the actual computing power requirements used in the calculation are all from the original papers of the corresponding models. Since the original data is generally the computing power of processing a single image, in order to unify the comparison benchmark, the computing power of each item in Table 3 is converted according to the scale of the training set used by the model, and the accuracy difference between the old and new formulas is presented as a percentage error improvement index to verify the effectiveness of the method provided in this application.

[0041] Table 3. Relevant Models and Computing Power Evaluation Results

[0042] This application selected Conformer-CTC, QuartzNet, and Squeezeformer models with clear data sources for verification. The relevant models and computational power evaluation results are detailed in Table 4. Among them, the number of parameters is the model's preset scale of millions, tens of millions, or hundreds of millions of parameters, and the resolution influence factor is composed of the sampling rate of the input audio, the average audio duration, and the architecture temporal factor. The computational power requirements of the old formula and the new formula are calculated according to formula (2) and formula (8), respectively, and the unit is GFLOPS. The number of parameters and the actual computational power requirements used in the calculation are all from the original papers of the corresponding models, and the sampling rate and average audio duration values ​​are taken from the articles corresponding to the LibriSpeech dataset used. Since the original data is generally the computational power of processing a single audio segment, in order to unify the comparison benchmark, the computational power of each item in Table 4 is converted according to the training set size used by the model, and the accuracy difference between the old and new formulas is presented by the "percentage error improvement" index to verify the effectiveness of the formula proposed in this application.

[0043] Table 4. Speech Model and Computing Power Evaluation Results

[0044] Example 2 This application also provides a business-based intelligent computing center computing power demand measurement device, such as... Figure 4The diagram shows a block diagram of a business-based intelligent computing center computing power demand calculation device. The functions implemented by this device correspond to the steps described above in executing a business-based intelligent computing center computing power demand calculation method on a terminal device. This device can be understood as a server component including a processor. The business-based intelligent computing center computing power demand calculation device described in this application includes: The mapping module 401 is used to pre-map the AI ​​technology of the intelligent computing center into a processing model, and to classify the processing model into different types according to the business scenario of the intelligent computing center. Configuration module 402 is used to configure corresponding computing power measurement networks for different types of processing models according to the stage in which the processing model is located; the computing power measurement network is obtained by optimizing the original computing power network; the processing model has different stages; The calculation module 403 is used to calculate the type calculation results of the processing model under different types based on the computing power calculation network, and to integrate the type calculation results under different stages to obtain the target calculation result, so as to calculate the computing power requirement of the intelligent computing center based on the target calculation result.

[0045] In one feasible implementation, the configuration module includes: A partitioning module is used to partition the processing model into a training phase and an inference phase based on the lifecycle of the processing model; The first configuration module is used to configure corresponding computing power measurement networks for the training phase and the inference phase respectively, so as to perform computing power requirement measurement.

[0046] In one feasible implementation, the configuration module further includes: An extraction module is used to extract the data features of the processing model in each type, and configure the corresponding type computing power network according to the data features; The first calculation module is used to calculate the computing power of the corresponding processing model based on the type of computing power network.

[0047] In one feasible implementation, the configuration module also includes: The settings module is used to set the image pixels as data features of the image type processing model and configure the resolution influence factor. The first optimization module is used to optimize the original computing power network of the image type based on the resolution influence factor to obtain the computing power network of the image type.

[0048] In one feasible implementation, the configuration module further includes: The settings module is used to set the temporal features of speech as the data features of the speech type processing model, and to configure the temporal length influence factor. The second optimization module is used to optimize the original computing power network of the speech type based on the time-series length influence factor to obtain the computing power network of the speech type.

[0049] In one feasible implementation, the configuration module further includes: The statistics module is used to process the model during the statistical inference stage, and to perform inference on the model in different types. The computation module is used to perform calculations based on the number of inferences in the model and the initial computing power results of the corresponding type of processing model to obtain the computing power measurement network for the inference stage.

[0050] In one feasible implementation, the measurement module includes: The collection module is used to collect and calculate the type of calculation results of the computing power measurement network output of different types of processing models at the same stage to obtain the stage calculation results. The fusion module is used to fuse the stage measurement results from different stages to obtain the target measurement result of the processing model.

[0051] Example 3 This application also provides an electronic device, such as Figure 5 As shown, it includes: a processor 501, a memory 502, and a bus 503. The memory 502 stores machine-readable instructions that can be executed by the processor 501. When the electronic device is running, the processor 501 and the memory 502 communicate through the bus 503. When the machine-readable instructions are executed by the processor 501, the steps of any one of the business-based intelligent computing center computing power demand measurement methods described above are executed.

[0052] Example 4 This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of any of the methods described in the application for calculating the computing power requirements of a business-based intelligent computing center.

[0053] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some communication interfaces; the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.

[0054] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0055] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0056] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a platform server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0057] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for measuring the computing power requirements of an intelligent computing center based on business needs, characterized in that, The method includes: The AI ​​technology of the intelligent computing center is pre-mapped as a processing model, and the processing model is divided into different types according to the business scenarios of the intelligent computing center; Based on the stage of the processing model, a corresponding computing power measurement network is configured for different types of processing models; the computing power measurement network is obtained by optimizing the original computing power network; the processing model has different stages; Based on the computing power measurement network, the type measurement results of the processing model under different types are calculated, and the type measurement results under different stages are integrated to obtain the target measurement result, so as to calculate the computing power requirement of the intelligent computing center based on the target measurement result.

2. The method according to claim 1, characterized in that, The step of configuring corresponding computing power measurement networks for different types of processing models according to the stage in which the processing model is located includes: Based on the lifecycle of the processing model, the processing model is divided into a training phase and an inference phase; Configure corresponding computing power measurement networks for the training and inference phases respectively to measure computing power requirements.

3. The method according to claim 2, characterized in that, The configuration of corresponding computing power measurement networks for the training and inference phases includes: Extract the data features of the processing model in each type, and configure the corresponding type computing power network according to the data features; Based on the aforementioned type of computing network, computing power is calculated for the corresponding processing model.

4. The method according to claim 3, characterized in that, The types include at least image types; The step of configuring the corresponding type of computing power network based on the data characteristics includes: Set the image pixels as the data feature of the image type processing model, and configure the resolution influence factor; Based on the resolution influence factor, the original computing power network for the image type is optimized to obtain the computing power network for that type.

5. The method according to claim 3, characterized in that, The types include at least voice types; The step of configuring the corresponding type of computing power network based on the data characteristics includes: Set the temporal features of speech as the data features of the speech type processing model, and configure the temporal length influence factor; Based on the time-series length influence factor, the original computing power network for the speech type is optimized to obtain the computing power network for that type.

6. The method according to claim 2, characterized in that, The configuration of corresponding computing power measurement networks for the training and inference phases includes: The number of inferences performed by the processing model in the statistical inference stage, in different types of models; The computing power measurement network for the inference stage is obtained by performing calculations based on the number of inferences in the model and the initial computing power results of the corresponding type of processing model.

7. The method according to claim 1, characterized in that, The target calculation result is obtained by integrating the type calculation results from different stages, including: Collect and calculate the type of computing power measurement network output results of different types of processing models at the same stage to obtain the stage measurement results; By integrating the stage measurement results from different stages, the target measurement result of the processing model is obtained.

8. A business-based intelligent computing center computing power demand measurement device, characterized in that, The device includes: The mapping module is used to pre-map the AI ​​technology of the intelligent computing center into processing models, and to classify the processing models into different types according to the business scenarios of the intelligent computing center; The configuration module is used to configure corresponding computing power measurement networks for different types of processing models according to the stage in which the processing model is located; the computing power measurement network is obtained by optimizing the original computing power network; the processing model has different stages; The calculation module is used to calculate the type calculation results of the processing model under different types based on the computing power calculation network, and to integrate the type calculation results under different stages to obtain the target calculation result, so as to calculate the computing power requirement of the intelligent computing center based on the target calculation result.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of a business-based intelligent computing center computing power demand measurement method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of a business-based intelligent computing center computing power demand calculation method as described in any one of claims 1 to 7.

Images