Game CPU-oriented model identification method and system

Abstract

The invention discloses a game CPU-oriented model identification method and system, and belongs to the technical field of computer performance optimization, and the method comprises the following steps: S1, collecting a real-time performance vector from a CPU core performance monitoring unit; S2, obtaining a predictive load vector from a game engine; s3, adjusting the dynamic learning rate based on the change degree of the predictive load vectors at the current moment and the previous moment; s4, calculating an instantaneous matching degree for each candidate model in a preset candidate CPU micro-architecture model set based on the real-time performance vector, the predictive load vector and a preset reference performance matrix associated with the candidate model; and S5, updating and generating a dynamic confidence score for each candidate model in combination with the confidence, the instantaneous matching degree and the dynamic learning rate of the previous moment, and winning a precious time window for predictive performance optimization of a game engine.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer performance optimization, in particular to a game CPU model identification method and system. BACKGROUND

[0002] In the game CPU model identification application, the traditional identification method mainly relies on a static and serial execution identification framework. These methods usually regard the identification process as a one-time preprocessing task completed at the program startup stage or a specific time. This status quo results in the inability of the identification result to adapt to the severe dynamic changes of the game runtime load, affecting the accuracy and timeliness of the performance optimization of the game engine.

[0003] The above status and deficiencies are mainly due to the limitations of the identification mechanism and technical strategy: the identification framework is static: the existing technology regards CPU model identification as a one-time task, which is completed at the beginning of the game or at a specific node, and cannot be adjusted according to real-time changes during the continuous running of the game, i.e., static identification cannot adapt to dynamic load; the response mechanism is passive: the traditional method makes judgments based on current or historical performance data, which is essentially a passive response mode, lacking the ability to predict future game loads; the adjustment capability is missing: the parameters of the identification algorithm are usually fixed; this leads to over-sensitivity when the game load is stable, and inability to respond quickly when the load mutates, making it difficult to balance the stability and sensitivity of identification.

[0004] As a result, when the game scene and computing demand mutate, this static and passive identification method is difficult to balance between identification accuracy and response delay; this makes the game engine unable to obtain real-time and accurate CPU model information, thereby hindering its performance of forward-looking performance coordination and resource scheduling, and making it difficult to meet the adaptive optimization needs in a high-dynamic game environment.

[0005] The above information disclosed in the background section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] The purpose of the present application is to provide a game CPU model identification method and system to solve the problems raised in the background technology.

[0007] The technical solution of the present application is as follows:

[0008] S1, collecting real-time performance vectors from a CPU core performance monitoring unit;

[0009] S2, obtaining a predictive load vector from a game engine;

[0010] S3, adjust the dynamic learning rate based on the variation degree of the current and previous predictive load vector

[0011] S4, for each candidate model in the set of pre-set candidate CPU micro-architecture models, calculate an instantaneous matching degree based on the real-time performance vector, the predictive load vector, and a pre-set benchmark performance matrix associated with the candidate model;

[0012] S5, for each candidate model, update a dynamic confidence score by combining the confidence at the previous time, the instantaneous matching degree, and the dynamic learning rate;

[0013] S6, monitor the dynamic confidence scores of all candidate models, and if the dynamic confidence score of a candidate model is continuously the highest and continuously higher than a pre-set confidence threshold within a pre-set time window, determine that the candidate model is the current effective CPU model;

[0014] S7, provide the current effective CPU model to the game engine.

[0015] Preferably, the real-time performance vector collected in S1 is composed of multiple normalized performance indicators, and the performance indicators represent the CPU micro-architecture state; the predictive load vector obtained in S2 is composed of multiple expected load weights, and the load weights represent the proportion of different computing tasks in the game engine.

[0016] Preferably, the step of calculating the instantaneous matching degree in S4 includes:

[0017] S41, multiply the predictive load vector and the benchmark performance matrix associated with the candidate model to generate a predicted performance vector;

[0018] S42, use a modified Gaussian radial basis function and a pre-set diagonal weight matrix to calculate the weighted distance between the real-time performance vector and the predicted performance vector to generate the instantaneous matching degree.

[0019] Preferably, the step of updating the dynamic confidence score in S5 includes: using a recursive formula to weight-average the confidence of the candidate model at the previous time and the instantaneous matching degree calculated at the current time according to the dynamic learning rate to generate the dynamic confidence score at the current time.

[0020] Preferably, the step of adjusting the dynamic learning rate in S3 includes:

[0021] S31, calculate the norm difference between the current and previous predictive load vectors to quantify the variation degree of the game load;

[0022] S32, input the norm difference value into a preset hyperbolic tangent function to generate a dynamic learning rate between a preset lower bound of the learning rate and an upper bound of the learning rate; when the norm difference value tends to zero, the dynamic learning rate tends to the lower bound of the learning rate, and when the norm difference value increases, the dynamic learning rate tends to the upper bound of the learning rate.

[0023] Preferably, the determination manner of the preset parameter comprises:

[0024] By performing offline benchmark testing on the CPU micro-architecture and fitting the test data, a benchmark performance matrix associated with each candidate model is determined; based on analysis of statistical characteristics of each performance indicator in the offline data set, a diagonal weight matrix is determined; by performing offline calibration on the validation data set, a lower bound of the learning rate, an upper bound of the learning rate and a load sensitivity coefficient are determined.

[0025] Preferably, S6 further comprises: if the determination condition is still not met at the end of the preset time window, maintaining monitoring of the dynamic confidence score of all candidate models without determining the current effective CPU model.

[0026] A CPU model identification system for games comprises:

[0027] A data acquisition module is configured to acquire a real-time performance vector from a CPU core performance monitoring unit and to obtain a predictive load vector from a game engine;

[0028] A core processing module is configured to adjust a dynamic learning rate based on a variation degree of the predictive load vector, to calculate an instantaneous matching degree for each candidate model in a preset candidate CPU micro-architecture model set, and to update a dynamic confidence score by combining a confidence score at a previous time, the instantaneous matching degree and the dynamic learning rate.

[0029] An identification decision module is configured to monitor the dynamic confidence scores of all candidate models and to determine a current effective CPU model when a preset duration and a threshold condition are met.

[0030] A feedback and adaptive module is configured to provide the current effective CPU model determined by the identification decision module to the game engine.

[0031] The present application provides a CPU model identification method and system for games by improvement, compared with the prior art, has the following improvements and advantages:

[0032] 1. By introducing the instantaneous matching degree calculation based on the predictive load vector, and combining the use of the diagonal weight matrix to screen the performance indicators, the recognition accuracy in complex working conditions is greatly improved. At the same time, the dynamic learning rate adjustment formula ensures that the recognition system can quickly switch from a stable state to a high response state when the game scene switches from a flat, such as a menu interface, to a severe, such as a large-scale team battle, and the recognition delay is controlled within milliseconds, which wins a valuable time window for the predictive performance optimization of the game engine;

[0033] 2. By recursively updating the dynamic confidence score, the data noise is smoothed, and the result jitter is avoided. Further, by setting a recognition decision logic that must continuously satisfy the highest score within a preset time window and be higher than the preset confidence threshold, the fuzzy state of the model confidence in the competition phase is effectively filtered. This strict decision mechanism ensures that the final output of the CPU model to the game engine has a very high credibility, preventing false recognition results from interfering with the game optimization strategy, thereby ensuring smooth and stable player experience;

[0034] 3. All key preset parameters in the present application, including the reference performance matrix, the diagonal weight matrix, and the learning rate related hyperparameters, have a clear, offline big data test and machine learning fitting based determination method. This makes the present scheme not an abstract theoretical framework, but an engineering solution with high implementability, which can exhibit consistent and robust performance on various hardware platforms and game types. BRIEF DESCRIPTION OF DRAWINGS

[0035] The present application will be further explained in conjunction with the accompanying drawings and examples:

[0036] Figure 1 is a flowchart of the method of the present application. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in conjunction with specific examples.

[0038] Example 1:

[0039] Please refer to Figure 1 The present application provides a CPU model recognition method for games, comprising the following steps: S1, collecting real-time performance vectors from the CPU core performance monitoring unit;

[0040] S2, obtaining a predictive load vector from the game engine;

[0041] S3, adjusting the dynamic learning rate based on the variation degree of the predictive load vector at the current and previous time

[0042] S4, for each candidate model in the preset candidate CPU micro-architecture model set, calculate an instantaneous matching degree based on the real-time performance vector, the predictive load vector and a preset benchmark performance matrix associated with the candidate model;

[0043] S5, for each candidate model, update a dynamic confidence score in combination with the confidence of the last moment, the instantaneous matching degree and a dynamic learning rate;

[0044] S6, monitor the dynamic confidence scores of all candidate models, and if the dynamic confidence score of a certain candidate model is continuously the highest score and continuously higher than a preset confidence threshold within a preset time window, determine that the candidate model is a current effective CPU model;

[0045] S7, provide the current effective CPU model to the game engine;

[0046] The CPU model identification method for games provided in the embodiment aims to improve the traditional static and serial CPU identification framework; the implementation basis of the method is a concurrent processing mechanism, which couples the real-time state sampling of hardware with the future load prediction of software, so that performance adaptive optimization is realized in a high dynamic game environment; by periodically collecting performance data of CPU micro-architecture and combining with the calculation demand prediction of the game engine for future scenes, the method can probabilistically evaluate the candidate CPU models; a high-confidence identification result is provided to the game engine in real time, which is used for performing forward-looking performance coordination and resource scheduling; this closed-loop adaptive design aims to alleviate the problem that the identification accuracy and response delay of the prior art are difficult to balance when the load mutates;

[0047] The present application is from static preprocessing to dynamic runtime service: the core of the prior art is to regard identification as a one-time task, while the present scheme changes it into a dynamic service that is parallel to the game process and is continuously performed; by periodically collecting real-time performance vectors and obtaining predictive load vectors in real-time game running, the present scheme realizes continuous monitoring and dynamic evaluation of CPU models, solving the problem that static identification cannot adapt to dynamic load.

[0048] Embodiment 2

[0049] The real-time performance vector collected in S1 is composed of multiple normalized performance indicators, and the performance indicators represent the CPU micro-architecture state; the predictive load vector obtained in S2 is composed of multiple expected load weights, and the load weights represent the proportion of different calculation tasks in the game engine;

[0050] The step of calculating the instantaneous matching degree in S4 includes:

[0051] S41, multiply the predictive load vector and the benchmark performance matrix associated with the candidate model to generate a predictive performance vector;

[0052] S42. Using a modified Gaussian radial basis function and a preset diagonal weight matrix, calculate the weighted distance between the real-time performance vector and the predicted performance vector to generate the instantaneous matching degree.

[0053] In this implementation scenario, the calculation of the instantaneous matching degree is a fundamental step in the subsequent recognition and inference process; this process begins with the acquisition of two key vectors: one is the real-time performance vector h collected from the CPU core performance monitoring unit at a sub-frame frequency. t It consists of K normalized metrics such as cycle instruction count and cache miss rate, which objectively reflect the microarchitectural operating state of the processor at time point t.

[0054] In a preferred embodiment of the present invention, K = 4, and the real-time performance vector h t Defined as After normalization, each component represents: Number of instructions executed per cycle. L3 cache miss rate; Branch prediction miss rate; The percentage of time the core is in an active C0 state; correspondingly, in this preferred embodiment, M = 5, and the predictive load vector l t+1 Defined as Each component is a weight, and the sum of the weights is 1. They represent the expected load percentage of the following tasks in the next time step: Physics simulation and collision detection; Artificial Intelligence (AI) Logic and Pathfinding; Character animation and skeletal updates; Rendering command generation and submission; Audio processing and mixing;

[0055] It should be emphasized that the values ​​of K and M and the specific indicators mentioned above are only examples. Those skilled in the art can add, delete or replace these indicators according to actual needs and available data without departing from the protection scope of this invention.

[0056] The other is the predictive load vector l obtained from the game engine interface for the next time step t+1. t+1 The expected load weights of M tasks, including physical computing and AI logic, characterize the upcoming computing demands.

[0057] For the real-time performance vector h tcollection can be achieved by accessing the performance monitoring unit of the processor. In practice, one can program the access using cross-platform libraries such as PAPI or operating system native APIs such as the perf_event_open interface of Linux, which may require corresponding driver support or program execution permissions. For the acquisition of the predictive load vector l t+1 , the game engine is usually required to provide a customized API interface. For example, the game engine can expose the internal task scheduler's estimation of the allocation of computing resources for the next frame or several frames, such as the thread scheduling weights or expected execution times of physics, AI, etc. tasks, to the recognition system through the API;

[0058] The generation of the instantaneous matching degree, which is based on the quantification of the strength of the observed evidence, i.e. the degree of agreement between the observed real-time performance h t+1 and the expected behavior of a candidate CPU microarchitecture model j under the known future load l t ; To do this, first multiply the predictive load vector l t+1 by the benchmark performance matrix B j preset by the model to obtain a predicted performance vector p j = l t+1 B j ; then the instantaneous matching degree S t (j) is defined by a mathematical formula.

[0059] To further clarify the formula, it is mathematically derived from the Gaussian radial basis function in statistics, but the distance metric inside it is modified as a weighted norm. The technical consideration of this modification is that the importance and stability of each component of the performance vector h t in distinguishing different CPU models differ, and introducing a diagonal weight matrix W is to give higher computational weight to performance indicators with high discrimination and good signal-to-noise ratio, thereby improving the accuracy of the matching degree calculation.

[0060] The composition and parameter interpretation of the formula are as follows:

[0061] S t (j) = exp(-(h t -l t+1 B j )W(h t -l t+1 B j ) T )

[0062] where S t (j) represents the dimensionless instantaneous matching degree of candidate model j at time point t; exp is the exponential function; h tis a 1xK dimensional real-time performance row vector, t is the current time point index; l t+1 is a 1xM dimensional predictive load row vector, t+1 is the next time point index; B j is an MxK dimensional preset benchmark performance matrix associated with the candidate model j, j is the candidate model index; W is a KxK dimensional diagonal weight matrix; in order to ensure dimensional consistency, if the k-th component of the performance vector h t has a dimension of [P k ], the dimension of the diagonal element w k of the diagonal matrix W is set to 1 / [P k ] 2 to ensure that the parameters of the exponential are dimensionless scalars; T is the transpose operator of the matrix; M is the dimension of the predictive load vector l t+1 , that is, the number of expected load tasks; in a preferred embodiment, M=5;

[0063] At the application level, the instantaneous matching degree function S t (j) is calculated once for each candidate CPU model j in parallel at each identification period; this process converts the original, multi-dimensional performance data into a clear, quantified evidence strength scalar, providing high-quality input for subsequent confidence updates;

[0064] The present application changes from passive response to forward-looking prediction: the prior art is based on current or historical performance data for judgment, which is essentially a passive response; the present scheme introduces the predictive load vector l t+1 of the game engine at the next time step, so that the identification process has forward-looking; it evaluates the behavior pattern of the current CPU under future load, and this predictive coupling makes the subsequent game performance optimization more forward-looking and effective.

[0065] Embodiment 3

[0066] The step of updating the dynamic confidence score in S5 includes: using a recursive formula to weight average the confidence of the candidate model at the last time and the instantaneous matching degree calculated at the current time according to the dynamic learning rate, to generate the dynamic confidence score at the current time;

[0067] The step of adjusting the dynamic learning rate in S3 includes:

[0068] S31, calculate the norm difference between the two predictive load vectors at the current and last time to quantify the degree of change of the game load;

[0069] S32, input the norm difference value into a preset hyperbolic tangent function to generate a dynamic learning rate between a preset lower bound of the learning rate and an upper bound of the learning rate; when the norm difference value tends to zero, the dynamic learning rate tends to the lower bound of the learning rate, and when the norm difference value increases, the dynamic learning rate tends to the upper bound of the learning rate;

[0070] In this embodiment, in order to obtain a stable and reliable identification conclusion, the updating process of the dynamic confidence score integrates a series of observation evidence; this process is realized through a recursive updating confidence evaluation, which balances the historical confidence and new evidence; at the same time, in order to cope with the dynamic changes of the game load, the dynamic learning rate used for weighting is also designed to be self-adaptive adjustment;

[0071] The confidence updating formula adopts a recursive form derived from the exponential moving average filter in signal processing, and its function is to model the identification task as an inference process that continuously updates the prior confidence according to new evidence; this can effectively filter out the instantaneous matching degree S t (j) existing noise; the adjustment logic of the dynamic learning rate, which takes into technical consideration the inherent characteristics of the game scene: the load is not uniformly changed; when the load is stable, the system should remain stable; and when the load changes dramatically, the system must respond quickly;

[0072] The updating of the confidence score is completed by the following formula:

[0073] C t (j)=(1-α t )C t-1 (j)+α t S t (j)

[0074] Wherein, C t (j) is the dimensionless confidence score of model j at time point t; C t-1 (j) is the confidence at the last time; S t (j) is the instantaneous matching degree at the current time; α t is the dynamic learning rate; t is the subscript of the current time point; j is the index of the candidate CPU model; t-1 is the subscript of the last time point; at the initial time t=0, the confidence of all models can be initialized as a uniform distribution C0(j)=1 / N;

[0075] The dynamic learning rate α t itself is determined by the degree of change of the game load, and its calculation formula is:

[0076] α t =α n +(α x -α n )tanh(β||l t+1 -l t ||)

[0077] where, a t is the dimensionless dynamic learning rate at time t; a n and a x are the preset lower and upper bounds of learning rate, respectively; b is a preset load sensitivity coefficient; l t+1 and l t are the current and previous predictive load vectors, respectively; tanh(·) is the hyperbolic tangent function; ||·|| represents the norm of a vector; l t is the previous predictive load row vector;

[0078] These two formulas constitute the adaptive core of the recognition system; at each time step, the system first evaluates the degree of change in the game scene by calculating the norm difference of the load vector, and generates the learning rate a t accordingly; then, a t is used to determine the weight of new evidence and historical confidence; when the game load is stable, a t tends to the lower bound a n , and the system exhibits high stability; when the game load mutates, a t quickly rises to the upper bound a x , and the system responsiveness is enhanced.

[0079] Embodiment 4

[0080] The determination method of the preset parameters includes:

[0081] By offline benchmark testing of CPU microarchitecture and fitting of test data, the benchmark performance matrix associated with each candidate model is determined; based on the analysis of the statistical characteristics of each performance indicator in the offline data set, the diagonal weight matrix is determined; by offline calibration on the validation data set, the lower bound of learning rate, the upper bound of learning rate and the load sensitivity coefficient are determined;

[0082] The source of all preset parameters in this method is based on offline data analysis and calibration to ensure the implementability of the method; the determination of the benchmark performance matrix B j is obtained by large-scale offline benchmark testing of each known CPU microarchitecture; run the standard load in the controlled environment, and collect the corresponding load and performance data pairs, and use multivariate linear regression method to fit the performance matrix B j for each CPU model j; the determination of the diagonal weight matrix W depends on the analysis of the statistical characteristics of each performance indicator in the offline test data set; each diagonal element w kThe value of is set to be inversely proportional to the within-class variance of the kth performance indicator under all CPU models and loads, so that the indicators that contribute more to distinguishing different CPU models and have more stable numerical values are given higher weights; such as learning rate boundary n , x and load sensitivity coefficient β, are ultimately determined by systematic offline calibration on an independent validation dataset containing multiple CPU models and game scenarios; the calibration process uses algorithms such as grid search, and the optimization goal is to maximize the accuracy and convergence speed of identification;

[0083] As an example, for a specific candidate CPU model j, for example, a model corresponding to AMD Zen4 microarchitecture, a set of parameters that may be obtained after offline calibration are as follows:

[0084] Benchmark performance matrix B j : The matrix is M x K, in this example, a 5 x 4 matrix, each row corresponds to a load task, and each column corresponds to a performance indicator; for example, the row vector corresponding to the physical simulation task may be [0.85, 0.15, 0.10, 0.90], which indicates that under this load, the expected IPC of this CPU is 85% of the nominal value, the L3 miss rate is high, the branch miss rate is low, and the core activity is high;

[0085] Diagonal weight matrix W: According to the aforementioned K = 4 performance indicators, a possible weight matrix is W = diag(12.0, 8.5, 9.5, 4.0). The value reflects that in this application scenario, IPC, weight 12.0 and branch prediction miss rate, weight 9.5 are the most critical indicators for distinguishing different CPU microarchitectures;

[0086] Hyperparameters: learning rate lower bound α n may be set to 0.05; learning rate upper bound α x may be set to 0.4; load sensitivity coefficient β may be set to 0.1. For parameters in decision logic, duration window τ may be set to 150 milliseconds, and confidence threshold C th may be set to 0.95;

[0087] These specific numerical values provide effective initial references for those skilled in the art when implementing and debugging the system.

[0088] Embodiment 5

[0089] S6 further comprises: if the conditions are still not met at the end of the preset time window, maintaining monitoring of the dynamic confidence scores of all candidate models without determining the current effective CPU model;

[0090] In this embodiment, the logic of the identification decision is designed to ensure the reliability of the output result; the system adopts a verification rule with time delay: the system will output a high-confidence identification result only when and if the confidence score C k (μ t ) of a certain candidate model μ k is not only always the highest among all models, but also consistently higher than a preset confidence threshold C th within a preset duration window τ; this double constraint aims to ensure the stability of the identification result, effectively preventing frequent changes in the output result caused by temporary signal fluctuations or confidence competition among models;

[0091] The decision logic contains a fault-tolerant mechanism; if the above determination conditions are not met at the end of the preset time window τ, the system will not make any identification decision and will maintain continuous monitoring of the confidence scores of all candidate models, entering the next evaluation cycle; the design of this decision mechanism aims to ensure the certainty and reliability of the output result, ensuring that the CPU model information is only passed to the game engine when the evidence is sufficient; the specific values of the time window τ and the confidence threshold C th , like other hyperparameters, are determined through offline calibration on the validation dataset to achieve the best balance between stability and responsiveness.

[0092] Embodiment 6

[0093] A game CPU model identification system, comprising: a data acquisition module for acquiring real-time performance vectors from a CPU core performance monitoring unit and obtaining predictive load vectors from a game engine;

[0094] a core processing module for calculating the instantaneous matching degree for each candidate model in a preset set of candidate CPU microarchitecture models, wherein the preset set of candidate CPU microarchitecture models is established in the offline stage, and the selection criteria are usually the mainstream CPU microarchitectures with high market share in the target game market; for example, the set can include microarchitecture models such as GoldenCove and RaptorCove of Intel Corporation, and microarchitecture models such as Zen3 and Zen4 of AMD Corporation; the set is extensible, and as new CPU models are launched, their corresponding microarchitecture models can be added to the set after offline benchmarking to ensure the timeliness of identification, and the dynamic confidence score is updated and generated in combination with the confidence score, instantaneous matching degree, and dynamic learning rate at the previous time;

[0095] an identification decision module for monitoring the dynamic confidence scores of all candidate models and determining the current effective CPU model when the preset duration and threshold conditions are met;

[0096] a feedback and adaptive module for providing the current effective CPU model determined by the recognition decision module to the game engine;

[0097] The embodiment also provides a CPU model recognition system for games, which is a physical carrier of the method and internally structured into four closely coordinated functional modules to form a closed-loop adaptive recognition system; the data acquisition module is responsible for concurrent input of double-path data, one path of which is used to capture real-time performance vectors from a CPU core performance monitoring unit and the other path is used to obtain predictive load vectors from a game engine; the core processing module receives the collected data, calculates an instantaneous matching degree for each candidate CPU model, and updates a dynamic confidence score; the recognition decision module continuously monitors confidence score streams of all models and applies a double-verification rule based on a time window and a confidence threshold to determine a current effective CPU model when the conditions are met; the feedback and adaptive module feeds back the CPU model determined by the recognition decision module to the game engine in real time, monitors continuous changes of the predictive load vectors, and dynamically adjusts a learning rate used by the core processing module according to the degree of change; the four modules work together to enable the whole system to serve as a dynamic runtime service and realize accurate, stable and sensitive CPU model recognition.

[0098] The present application adjusts from a fixed strategy to an adaptive adjustment: once the recognition algorithm of the prior art is set, the parameters thereof are usually fixed; the present scheme introduces a dynamic learning rate α t adjustment mechanism, so that the system can autonomously adjust the speed of acceptance of new evidence according to the degree of change of game load; this adaptive ability is the key to the present scheme to simultaneously maintain high stability and high responsiveness in a high dynamic range of game scenarios, which is not possessed by the fixed strategy of the prior art.

[0099] In the present application, the instantaneous matching degree formula S t (j) = exp(-(h t -l t+1 B j )W(h t -l t+1 B j ) T ), wherein S t (j) is the instantaneous matching degree, h t is the real-time performance vector, l t+1 is the predictive load vector, and B jis the benchmark performance matrix, W is the diagonal weight matrix, and its core practical significance lies in providing a quantitative evidence snapshot; it answers the question of how similar the current hardware behavior is to the expected behavior of model j under the upcoming load by calculating the weighted distance between the real-time observed performance and the model's theoretical predicted performance; the diagonal weight matrix W here is particularly critical, as it enables the formula to intelligently amplify the influence of key performance indicators and suppress the interference of noise indicators, thereby improving the quality and reliability of single evidence;

[0100] Dynamic confidence update formula C t (j) = (1 - a t )C t-1 (j) + a t S t (j); where C t (j) is the current confidence, C t-1 (j) is the last time confidence, S t (j) is the instantaneous matching degree, a t is the dynamic learning rate, which realizes a belief smoothing and accumulation process in the time domain; it recursively averages a series of noisy instantaneous evidence snapshots S t (j) to form a stable growing or decaying confidence score C t (j); this effectively avoids the system from making false judgments due to single abnormal performance readings, and gives the recognition process robustness in the time dimension;

[0101] Dynamic learning rate adjustment formula a t = a n + (a x - a n ) tanh (b || l t+1 - l t ||); where a t is the dynamic learning rate, a n , a x are the learning rate boundaries, b is the load sensitivity coefficient, l t+1 , l t are the load vectors; its practical significance is to equip the entire recognition system with an adaptive governor; it directly links the sensitivity of recognition to the dynamics of the game scenario; when the game load is stable, the learning rate a t automatically decreases, the system tends to be conservative, and relies more on the historically accumulated confidence for stability; when the game load changes dramatically, the learning rate a t then automatically increases, the system becomes aggressive, giving new evidence a higher weight to achieve rapid response;

[0102] These three are logically tightly coupled: S t(j) provides high-quality transient evidence, alpha t According to the scene dynamics, the credibility weight of the evidence is determined, and finally acts on the C t The update of (j) realizes the dynamic identification considering stability, accuracy and agility.

[0103] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A method for identifying the model of a gaming CPU, characterized in that, Includes the following steps: S1. Collect real-time performance vectors from the CPU core performance monitoring unit; S2. Obtain the predictive load vector from the game engine; S3. Adjust the dynamic learning rate based on the degree of change of the predictive load vector between the current time and the previous time step; S4. For each candidate model in the preset set of candidate CPU microarchitecture models, calculate the instantaneous matching degree based on the real-time performance vector, the predictive load vector, and the preset baseline performance matrix associated with the candidate model. S5. For each candidate model, combine the confidence score, instantaneous matching score and dynamic learning rate from the previous time step to update and generate a dynamic confidence score. S6. Monitor the dynamic confidence scores of all candidate models. If the dynamic confidence score of a candidate model remains the highest and remains higher than the preset confidence threshold within a preset time window, then determine that the candidate model is the currently valid CPU model. S7. Provide the game engine with the currently available CPU model.

2. The method for model identification of gaming CPUs according to claim 1, characterized in that, The real-time performance vector collected by S1 consists of multiple normalized performance metrics, which represent the state of the CPU microarchitecture; the predictive load vector obtained by S2 consists of multiple expected load weights, which represent the proportion of different computing tasks in the game engine.

3. The method for model identification of gaming CPUs according to claim 1, characterized in that, The steps for calculating the instantaneous matching degree in S4 include: S41. Multiply the predictive load vector with the baseline performance matrix associated with the candidate model to generate the predictive performance vector; S42. Using a modified Gaussian radial basis function and a preset diagonal weight matrix, the weighted distance between the real-time performance vector and the predicted performance vector is calculated to generate the instantaneous matching degree.

4. The method for model identification of gaming CPUs according to claim 1, characterized in that, The steps for updating the dynamic confidence score in S5 include: using a recursive formula, weighting the candidate model's confidence score at the previous time step with the currently calculated instantaneous matching degree based on the dynamic learning rate to generate the dynamic confidence score at the current time step.

5. The method for model identification of gaming CPUs according to claim 1, characterized in that, The steps for adjusting the dynamic learning rate in S3 include: S31. Calculate the norm difference between the two predictive load vectors at the current time and the previous time to quantify the drastic changes in game load. S32. Input the norm difference value into the preset hyperbolic tangent function to generate a dynamic learning rate between the preset lower bound and upper bound of the learning rate; when the norm difference value approaches zero, the dynamic learning rate approaches the lower bound of the learning rate, and when the norm difference value increases, the dynamic learning rate approaches the upper bound of the learning rate.

6. The method for model identification of gaming CPUs according to claim 1, characterized in that, The methods for determining preset parameters include: By conducting offline benchmark tests on the CPU microarchitecture and fitting the test data, the benchmark performance matrix associated with each candidate model is determined; based on the analysis of the statistical characteristics of each performance index in the offline dataset, the diagonal weight matrix is ​​determined; and by performing offline calibration on the validation dataset, the lower bound of the learning rate, the upper bound of the learning rate, and the load sensitivity coefficient are determined.

7. The method for model identification of gaming CPUs according to claim 1, characterized in that, S6 also includes: if the determination condition is not met by the end of the preset time window, then the dynamic confidence scores of all candidate models are monitored, but the currently effective CPU model is not determined.

8. A model identification system for gaming CPUs, applied to the model identification method for gaming CPUs as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect real-time performance vectors from the CPU core performance monitoring unit and obtain predictive load vectors from the game engine. The core processing module is used to adjust the dynamic learning rate based on the degree of change of the predictive load vector, calculate the instantaneous matching degree for each candidate model in the preset candidate CPU microarchitecture model set, and update and generate the dynamic confidence score by combining the confidence score, instantaneous matching degree and dynamic learning rate of the previous time step. The identification and decision-making module is used to monitor the dynamic confidence scores of all candidate models and determine the currently effective CPU model when the preset duration and threshold conditions are met. The feedback and adaptation module is used to provide the game engine with the currently valid CPU model determined by the identification and decision-making module.

Images