AI computing power server asymmetric air flow channel heat dissipation system

Abstract

This invention discloses an asymmetric airflow cooling system for AI computing servers, relating to the field of server cooling technology. It collects the temperature of heat-generating elements using a thin-film thermocouple array and calculates heat dissipation using a thermal resistance network model, constructing a heat load distribution matrix. A graph convolutional neural network and density peak clustering algorithm are used to identify hot and cold areas, generating an asymmetric airflow layout strategy. The layout is optimized based on a deep Q-network, and independently controllable airflow units are set. Computational fluid dynamics simulations are used to obtain velocity and temperature fields and determine if the cooling threshold is met. A Bayesian optimization algorithm is used to collaboratively adjust the airflow cross-sectional shape, inlet angle, and fan speed, iteratively optimizing until the requirements are met. The airflow is then controlled to perform cooling operations according to the target model. This invention achieves precise matching between airflow layout and heat load distribution, significantly improving cooling efficiency, reducing fan energy consumption, and ensuring stable operation of AI computing servers under high load conditions.

Description

Technical Field

[0001] This invention relates to the field of server heat dissipation technology, specifically to an asymmetric airflow heat dissipation system for AI computing power servers. Background Technology

[0002] With the rapid development of artificial intelligence technology, the computing power density of AI computing servers continues to increase, especially with the large-scale deployment of high-performance computing chips such as GPUs / NPUs. This has resulted in a highly non-uniform distribution of heat flux density within the servers. Traditional server cooling duct designs often employ a symmetrical layout, where air inlets, outlets, and fan modules are arranged uniformly and symmetrically, with airflow distribution relying on experience or fixed proportions. However, under high-load AI training scenarios, the heat generation varies significantly across different areas within the server, with some areas (such as around the GPU and power modules) becoming localized hotspots. Symmetrical airflow ducts cannot dynamically match airflow according to the actual heat load, leading to low cooling efficiency, increased energy consumption, and even frequency throttling or system crashes, severely impacting the stable operation and computing power output efficiency of AI computing servers. Therefore, designing a cooling method that can adapt to asymmetrical heat load distribution and achieve precise airflow matching has become a key technical challenge for improving the cooling performance of AI computing servers. Summary of the Invention

[0003] The purpose of this invention is to provide an asymmetric airflow cooling system for AI computing servers to address the shortcomings of the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an asymmetric airflow cooling system for AI computing power servers, comprising:

[0005] Thermal power consumption data acquisition module: Acquires thermal power consumption data of key heat-generating components of AI computing server under typical load conditions, and constructs thermal load distribution matrix inside the server;

[0006] Layout strategy generation module: Based on the heat load distribution matrix, identify hot and cold areas inside the server and generate an asymmetric airflow layout strategy, including the air volume distribution ratio of each airflow inlet and the airflow direction.

[0007] Air duct unit construction module: Based on the air volume distribution ratio, multiple independently controllable air duct units are constructed, each corresponding to a different heat generation area;

[0008] Simulation analysis module: Performs fluid dynamics simulation on each air duct unit, obtains the airflow velocity field and temperature field distribution in each air duct unit, and determines whether the preset heat dissipation threshold is met;

[0009] Adjust and update the module: If there are areas that do not meet the heat dissipation threshold, adjust the cross-sectional shape, inlet angle or fan speed of the corresponding air duct unit, update the air volume distribution ratio, and re-simulate until all areas meet the heat dissipation requirements to obtain the target asymmetric airflow duct model.

[0010] Execution module: Controls the internal airflow of the server to perform heat dissipation operations according to the target asymmetric airflow model.

[0011] Preferably, the thermal power consumption data acquisition module specifically includes the following steps:

[0012] Under typical load conditions of AI computing servers, temperature data is collected in real time by thin-film thermocouple arrays deployed on the surface of each key heat-generating element;

[0013] The collected temperature data is input into the pre-built thermal resistance network model. By combining the thermal resistance parameters of each heating element with the inlet temperature of the cooling medium, the real-time heat power consumption of each heating element is derived in reverse.

[0014] The real-time thermal power consumption value is normalized according to the component type and spatial location to generate an m-row n-column heat load distribution matrix, where the rows of the matrix correspond to different types of heat-generating components and the columns of the matrix correspond to the spatial partitions inside the server.

[0015] Preferably, identifying hot and cold areas within the server based on the heat load distribution matrix specifically includes the following steps:

[0016] The heat load distribution matrix is ​​input into a spatial heat distribution feature extraction model based on graph convolutional neural network to extract the heat load coupling features and spatial correlation features between each spatial partition; the extracted feature vectors are input into the density peak clustering algorithm for cluster analysis to divide adjacent spatial partitions with similar heat load features into the same heat load cluster.

[0017] Calculate the average heat flux density value of each heat load cluster and compare the average heat flux density value with a preset heat flux density threshold. If the average heat flux density value is greater than or equal to the preset heat flux density threshold, the cluster is marked as a hot spot area. If the average heat flux density value is less than the preset heat flux density threshold, the cluster is marked as a cold spot area.

[0018] Preferably, constructing multiple independently controllable duct units based on airflow distribution ratios specifically includes the following steps:

[0019] The air volume distribution ratio is input into the air duct unit layout optimization model based on deep Q network. Taking the internal space partition of the server as nodes, and taking the maximization of heat dissipation efficiency and the minimization of flow resistance as dual objectives, the optimal spatial layout scheme of the air duct unit is generated.

[0020] According to the optimal spatial layout scheme, multiple independent air duct unit areas are divided inside the server. Each air duct unit area corresponds to a spatial partition cluster, and an independent air volume control valve is set at the entrance of each air duct unit area.

[0021] Preferably, each air volume ratio value in the air volume distribution ratio is assigned to the air volume control valve at the entrance of the corresponding air duct unit area as the initial air volume setting value of the air duct unit.

[0022] Preferably, performing fluid dynamics simulation on each air duct unit and determining whether the preset heat dissipation threshold is met specifically includes the following steps:

[0023] Based on the spatial location information and initial air volume setting of each air duct unit area, an initial air duct model M0 containing multiple independent air duct units is constructed, and the initial air duct model M0 is subjected to mesh division processing.

[0024] The mesh file of the initial duct model M0 is input into the computational fluid dynamics solver for iterative solution to obtain the three-dimensional airflow velocity field distribution data and the three-dimensional temperature field distribution data inside each duct unit;

[0025] Extract the temperature values ​​at the surface locations of all key heat-generating elements within each air duct unit, compare the extracted temperature values ​​with a pre-set heat dissipation temperature threshold, and determine whether there are any heat-generating elements that exceed the heat dissipation temperature threshold.

[0026] If there are heat-generating elements that exceed the heat dissipation temperature threshold, the air duct unit containing the heat-generating element will be marked as an air duct unit to be optimized.

[0027] Preferably, adjusting the cross-sectional shape, inlet angle, or fan speed of the corresponding duct unit to update the airflow distribution ratio specifically includes the following steps:

[0028] Based on the labeling results of the duct unit to be optimized, a multi-parameter collaborative adjustment algorithm based on Bayesian optimization is used to generate the adjustment amount of the cross-sectional shape parameter, the adjustment amount of the inlet angle parameter, and the adjustment amount of the corresponding fan speed parameter of the duct unit to be optimized.

[0029] The generated parameter adjustments are applied to the corresponding duct unit to be optimized, the geometric model and fan operating parameters of the duct unit are updated, and the air volume distribution ratio RS at the inlet of each duct unit is recalculated.

[0030] Preferably, the updated duct unit geometric model and air volume distribution ratio RS are re-input into the computational fluid dynamics solver for iterative solution to obtain new airflow velocity field distribution data and temperature field distribution data; until the surface temperature values ​​of the key heat-generating elements inside all duct units are less than or equal to the corresponding heat dissipation temperature threshold, the duct unit geometric model and fan operating parameters at this time are determined as the final target asymmetric airflow duct model.

[0031] Preferably, the key heat-generating components include a graphics processor, a central processing unit, a power supply module, and a memory module.

[0032] Preferably, the calculation formula of the reverse thermal deduction algorithm is that the thermal power consumption equals the surface temperature minus the inlet temperature of the cooling medium, and then divided by the total thermal resistance.

[0033] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0034] 1. This invention achieves accurate identification and adaptive hotspot region division of non-uniform heat load distribution within AI computing servers by constructing a spatial thermal distribution feature extraction model based on graph convolutional neural networks and a density peak clustering algorithm. Compared to traditional airflow design methods that rely on experience or symmetry assumptions, this invention can accurately capture the spatial thermal coupling characteristics of high-power components such as graphics processing units (GPUs) and central processing units (CPUs), avoiding heat dissipation blind spots caused by local hotspots, thereby significantly improving the targeting and effectiveness of the heat dissipation system. Simultaneously, an airflow unit layout optimization model based on deep Q-networks is introduced, with synergistic optimization of maximizing heat dissipation efficiency and minimizing flow resistance as dual objectives. This enables dynamic matching between the airflow layout and heat load distribution, minimizing airflow resistance and reducing fan energy consumption while ensuring controllable temperatures for key heat-generating components.

[0035] 2. This invention employs a multi-parameter collaborative adjustment algorithm based on Bayesian optimization to jointly iteratively optimize the cross-sectional shape, inlet angle, and fan speed of the air duct unit. Closed-loop verification is then performed through computational fluid dynamics simulation until all heat-generating components meet the heat dissipation threshold requirements. This mechanism solves the technical challenge of traditional heat dissipation design, where parameter adjustments rely on empirical trial and error and are difficult to converge to the global optimum. It can automatically optimize and generate a target asymmetric airflow duct model under complex operating conditions. Finally, the optimized model parameters are precisely mapped to the actual physical air duct and fan control through the execution module, achieving complete closed-loop control from heat dissipation perception and air duct layout optimization to precise execution. This significantly improves the heat dissipation efficiency and operational stability of AI computing servers in high-density deployment scenarios, extends server lifespan, and reduces maintenance costs. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0037] Figure 1 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] For examples, please refer to Figure 1 As shown, the asymmetric airflow cooling system for AI computing servers described in this embodiment includes:

[0040] Thermal power consumption data acquisition module: Acquire thermal power consumption data of key heat-generating components of AI computing server under typical load conditions, and construct thermal load distribution matrix inside the server.

[0041] Under typical load conditions of the AI ​​computing server, thermal power consumption data acquisition is performed. Typical load conditions refer to running representative AI training or inference tasks, causing key heat-generating components such as the graphics processor, central processing unit, power supply module, and memory module to reach their expected thermal design power levels. To achieve accurate measurement of the surface temperature of each component, thin-film thermocouple arrays are deployed in the central and edge regions of each key heat-generating component. These arrays consist of multiple K-type thin-film thermocouples with a thickness of less than 0.1 mm, adhered to the component surface using thermally conductive adhesive to ensure close contact between the measurement point and the component surface, minimizing contact thermal resistance. The outputs of all thin-film thermocouples are connected to a data acquisition card, which continuously acquires temperature signals at a sampling frequency of 10 times per second. The analog signals are converted to digital signals and transmitted to the server substrate management controller for storage, thus forming a real-time temperature dataset corresponding to each key heat-generating component.

[0042] The collected real-time temperature data is input into a pre-constructed thermal resistance network model. Combined with pre-measured junction-to-shell thermal resistance parameters of each heat-generating element and the inlet temperature of the cooling medium, the real-time thermal power consumption of each heat-generating element is calculated using a reverse thermal derivation algorithm. The construction process of the thermal resistance network model is as follows: First, for each key heat-generating element, a one-dimensional steady-state heat conduction path is established from the chip junction temperature to the surface temperature measurement point, and then from the surface temperature measurement point to the cooling air. Second, based on the material properties and geometric dimensions of the heat conduction path, the thermal resistance value of each path segment is determined. The internal thermal resistance of the chip is provided by the component datasheet, and the contact thermal resistance between the thermal conductive adhesive and the heat sink is obtained through experimental calibration. Finally, all thermal resistance values ​​are combined according to a series-parallel relationship to form a complete thermal resistance network topology. The reverse thermal inference algorithm is executed as follows: First, the surface temperature measured by the thin-film thermocouple is obtained as a boundary condition; second, the inlet temperature of the cooling medium measured by the temperature sensor deployed at the server air inlet is obtained as another boundary condition; third, based on the thermal resistance network model, the thermal power consumption is calculated by using the formula: thermal power consumption equals surface temperature minus cooling medium inlet temperature divided by total thermal resistance, where the total thermal resistance is obtained by summing all thermal resistances from the surface temperature measurement point to the cooling air in the thermal resistance network model; fourth, the calculated value is determined as the real-time thermal power consumption value of the heating element.

[0043] The calculated real-time thermal power consumption values ​​of each key heat-generating component are normalized according to component type and spatial location to generate a heat load distribution matrix with m rows and n columns. The normalization process is implemented as follows: First, the internal space of the server is divided into n consecutive spatial partitions along the airflow direction, each partition having the same vertical span. Second, all key heat-generating components are divided into m categories according to component type, including graphics processors, central processing units, power modules, and memory modules. Then, for each component category, the sum of the real-time thermal power consumption values ​​of all heat-generating components within each spatial partition is calculated. Finally, the above statistical results are arranged with categories as rows and spatial partitions as columns to form a heat load distribution matrix. The element value in the i-th row and j-th column of the matrix represents the total thermal power consumption value of the i-th type of heat-generating component in the j-th spatial partition.

[0044] Layout strategy generation module: Based on the heat load distribution matrix, identify hot and cold areas inside the server and generate an asymmetric airflow layout strategy, including the air volume distribution ratio of each airflow inlet and the airflow direction.

[0045] The heat load distribution matrix is ​​input into a pre-built spatial heat distribution feature extraction model based on a graph convolutional neural network. Through the calculation and processing of this model, the heat load coupling features and spatial correlation features between different spatial partitions are extracted. The construction process and specific implementation method of the spatial heat distribution feature extraction model based on the graph convolutional neural network are as follows:

[0046] A spatial topology graph is constructed, defining the n spatial partitions within the server as nodes. Edge connections are established based on whether adjacent spatial partitions share physical boundaries. Specifically, if two spatial partitions are adjacent along or perpendicular to the airflow direction within the server, an edge is created between these two nodes, with the initial weight of the edge set to 1. This forms an undirected graph G = (V, E), where V is the set of nodes and E is the set of edges.

[0047] Construct a node feature matrix, using each column of the heat load distribution matrix as the initial feature vector for the corresponding spatial partition. Since the heat load distribution matrix has m rows, containing heat dissipation information for m types of heating elements, the initial feature vector for each spatial partition has a dimension of m. Arrange the initial feature vectors of all n spatial partitions in partition order to form a node feature matrix X, with a dimension of n rows multiplied by m columns.

[0048] A graph convolutional neural network (GCNN) structure is constructed, consisting of an input layer, two graph convolutional layers, a global average pooling layer, and an output layer connected sequentially. The first graph convolutional layer has 64 kernels and uses a rectified linear function (CLM) as its activation function; the second graph convolutional layer has 32 kernels and also uses a CLM as its activation function. The graph convolutional layers employ a Chebyshev polynomial approximation method to implement spectral graph convolution, with the core calculation formula being: Among them, H (1) H represents the feature matrix of the l-th layer node. (1+1) W represents the feature matrix of the (l+1)th layer node. (1) Let be the trainable weight matrix of the l-th layer, and σ represent the linear rectified activation function. The normalized Laplace matrix is ​​calculated using the following formula: Where A is the adjacency matrix of the graph, A{ij}=1 indicates that there is an edge connection between node i and node j, otherwise A{ij}=0; D is the degree matrix, which is a diagonal matrix, and its diagonal element D{ii} is equal to the sum of all elements in the i-th row of the adjacency matrix A, that is, the number of edges connected to node i.

[0049] A graph convolutional neural network is trained by collecting thermal load distribution matrix data of AI computing servers under historical operating conditions as training samples, using manually labeled spatial thermal distribution coupling feature labels as supervision signals, and employing a backpropagation algorithm to iteratively optimize the weight parameters of the graph convolutional neural network until the loss function converges. The loss function is the mean squared error function.

[0050] The feature extraction operation is performed, and the heat load distribution matrix at the current moment is constructed into a node feature matrix X in the same way as the node feature matrix. The X is then input into the trained graph convolutional neural network. After feature transformation of two graph convolutional layers and dimensionality reduction processing of the global average pooling layer, a feature vector F with a dimension of 32 is finally output. The feature vector F is the numerical representation of the heat load coupling features and spatial correlation features between the extracted spatial partitions.

[0051] The extracted feature vector F is input into the density peak clustering algorithm. Through the clustering analysis of this algorithm, adjacent spatial partitions with similar heat load characteristics are divided into the same heat load cluster. The specific implementation steps of the density peak clustering algorithm are as follows:

[0052] For each feature vector point i corresponding to a spatial partition, calculate its Euclidean distance d with all other feature vector points j. ij The Euclidean distance d ij The calculation formula is: Among them, F ik Let F represent the k-th dimension value of the feature vector F corresponding to the i-th spatial partition. jk Let represent the k-th dimension value of the feature vector F corresponding to the j-th spatial partition. Then, the local density ρ of each feature vector point i is calculated using the truncation kernel method. i Its calculation formula is: ρ i =∑ j≠i χ(d ij -d c The function χ(x) is defined as follows: when x < 0, χ(x) = 1; when x ≥ 0, χ(x) = 0. That is, the local density ρ i Equals all conditions d ij -d c The number of eigenvector points j with a value less than 0, which is the number of eigenvector points j whose distance to point i is less than the cutoff distance d. c The number of feature vector points. The cutoff distance parameter d. c The value of is determined by the distance d between all pairs of feature vector points. ij Arrange the distances in ascending order to form a distance sequence, and take the distance value of the first 2% of the positions in this distance sequence as d. c The value.

[0053] For each feature vector point i, find all local densities ρ j Greater than ρ i For feature vector points j, calculate the Euclidean distance d between these points and point i. ij And take the minimum value from these distance values, and define the minimum value as the distance value δ. i Its mathematical expression is: For the eigenvector point with the maximum local density, i.e., satisfying ρ i For a point whose value is max(ρ), its distance δi is defined as the maximum distance between all pairs of points.

[0054] Using the calculated local density ρ as the abscissa and the calculated distance value δ as the ordinate, a scatter plot of all feature vector points is drawn in a two-dimensional coordinate system. This scatter plot is called a decision plot. In the decision plot, feature vector points located in the upper right region have both large local density ρ values ​​and large distance values ​​δ; these points are determined as cluster centers. The specific selection method is as follows: calculate γ for each feature vector point. i =ρ i ×δ i All feature vector points are sorted from largest to smallest according to their γ values. The k points with the highest γ values ​​and significant jumps in γ values ​​are selected as cluster centers. The number of cluster centers, k, is automatically determined based on the actual data distribution.

[0055] Each non-cluster center feature vector point is assigned to the cluster to which the nearest feature vector point with a local density greater than that point belongs. The specific assignment rule is: for a non-cluster center point i, find the cluster that satisfies condition ρ. j >ρ i From all feature vector points j, select those that make d ij The point i with the smallest eigenvector j is assigned to the cluster to which point j belongs. During the assignment process, a spatial adjacency constraint is introduced: point i is only allowed to be assigned to the cluster to which point j belongs if the spatial partitions corresponding to two eigenvector points i and j have an edge connection in the constructed spatial topology graph G, i.e., Aij = 1; if there is no edge connection between the two eigenvector points, i.e., Aij = 0, then even if the features are similar and dij is the smallest, they will not be assigned to the same cluster. This constraint ensures that the clustering results are spatially continuous, meaning that spatial partitions within the same heat load cluster are physically adjacent.

[0056] Spatial partitions with the same cluster identifier after allocation are grouped into the same heat load cluster, and the set of spatial partition numbers of all heat load clusters is output. The output result is represented as C={C1, C2, ..., Cr}, where Cr represents the list of spatial partition numbers contained in the r-th heat load cluster, and r is the total number of clusters.

[0057] For each heat load cluster obtained, the average heat flux density value corresponding to the cluster is calculated, and the calculated average heat flux density value is compared with a pre-set heat flux density threshold. Based on the comparison result, each heat load cluster is marked as a hot spot or cold spot region. The specific implementation steps are as follows:

[0058] For each spatial partition, the total heat dissipation value of that spatial partition is calculated based on the column vectors in its corresponding heat load distribution matrix Q, i.e., by summing all m elements in that column vector. Then, the total heat dissipation value of the spatial partition is divided by its base area to obtain the heat flux density value of that spatial partition. The base area of ​​the spatial partition refers to its projected area in the horizontal direction, in square centimeters.

[0059] For a heat load cluster containing r spatial partitions, the heat flux density values ​​of all r spatial partitions in the cluster are summed to obtain the total heat flux density sum. Then, the total heat flux density sum is divided by r to obtain the average heat flux density value of the heat load cluster.

[0060] Based on the AI ​​computing server heat dissipation design specifications and the heat dissipation capacity of the cooling medium, a heat flux density threshold is predetermined. The specific method for determining the heat flux density threshold is as follows: First, obtain the maximum cooling airflow value that the cooling fans deployed inside the server can provide at full speed; second, calculate the maximum heat that can be removed per unit time based on the cooling airflow value, air specific heat capacity, and allowable temperature rise; finally, divide this maximum heat value by the total bottom area of ​​all space partitions inside the server to obtain the critical heat flux density value, and multiply this critical heat flux density value by a safety factor of 0.85 to obtain the preset heat flux density threshold.

[0061] The average heat flux density value of each heat load cluster is compared with a preset heat flux density threshold. If the average heat flux density value of the heat load cluster is greater than or equal to the preset heat flux density threshold, the heat load cluster is marked as a hot spot area; if the average heat flux density value of the heat load cluster is less than the preset heat flux density threshold, the heat load cluster is marked as a cold spot area.

[0062] The spatial location information corresponding to each hot spot and cold spot region after the marking process is completed is summarized and organized to form a hot spot region spatial location dataset and a cold spot region spatial location dataset. The spatial location information includes the spatial partition number contained in each region, the start and end coordinates of each partition, and the sequential relationship of the regions along the airflow direction inside the server.

[0063] Air duct unit construction module: Based on the air volume distribution ratio, multiple independently controllable air duct units are constructed, each corresponding to a different heat generation area.

[0064] The airflow distribution ratio is input into a pre-built air duct unit layout optimization model based on a deep Q-network. Through reinforcement learning optimization of this model, using the internal space partitions of the server as nodes, and with the dual objectives of maximizing heat dissipation efficiency and minimizing flow resistance, the optimal spatial layout scheme of the air duct units is generated. The construction process and specific implementation method of the air duct unit layout optimization model based on the deep Q-network are as follows:

[0065] The server is divided into n spatial partitions, which are used as the basic units of the state space. Each state is represented by a binary vector s of length n. Each element in the vector takes the value 0 or 1. A value of 1 indicates that the spatial partition has been selected as the boundary position of the air duct unit, and a value of 0 indicates that the spatial partition is not the boundary of the air duct unit. The state space S contains all possible combinations of boundaries.

[0066] Define an action space. The actions performed include setting or removing air duct unit partitions between any two adjacent spatial partitions. The size of the action space is equal to the total number of edges in the spatial topology graph G, and each action corresponds to the state transition of one edge.

[0067] A reward function is defined, which comprehensively considers both heat dissipation efficiency and flow resistance. The heat dissipation efficiency index is defined as the difference between the average airflow velocity in all hot spots and the average airflow velocity in all cold spots; a larger difference indicates higher heat dissipation efficiency. The flow resistance index is defined as the total pressure drop of the entire duct system; a smaller pressure drop indicates lower flow resistance. The reward function equals the heat dissipation efficiency index multiplied by a weighting coefficient α minus the flow resistance index multiplied by a weighting coefficient β. The weighting coefficient α is set to 0.6, and the weighting coefficient β is set to 0.4 to reflect the design principle of prioritizing heat dissipation efficiency.

[0068] A deep Q-network is constructed, employing a fully connected neural network structure consisting of an input layer, three hidden layers, and an output layer. The input layer has n neurons and receives the current state s as input. The first hidden layer contains 128 neurons with a rectified linear function (CLM) activation function; the second hidden layer contains 256 neurons with a CLM activation function; and the third hidden layer contains 128 neurons with a CLM activation function. The output layer has a number of neurons equal to the size of the action space A and outputs the Q-value for each possible action.

[0069] Training the Deep Q-Network: Historical airflow distribution ratio data and corresponding duct unit layout schemes under various operating conditions were collected as training samples. An experience replay mechanism was employed, storing the experience data generated by the agent's interaction with the environment in an experience pool. A small batch of data was randomly sampled from the experience pool each time for network parameter updates. The loss function used was the mean squared error function, calculated as: loss equals the expected value of the square of the current Q-value minus the target Q-value. The target Q-value was calculated as: the target Q-value equals the immediate reward R plus a discount factor γ multiplied by the maximum Q-value of the next state. The discount factor γ was set to 0.95. Gradient descent optimization was performed using the Adam optimizer with a learning rate of 0.001.

[0070] The current airflow distribution ratio R is input as a constraint into the trained deep Q-network. Using all spatial partitions within the server as nodes, the agent explores the environment and continuously adjusts the placement of the air duct unit partitions until it converges to the optimal layout. The output of the optimal layout is a boundary marker vector B of length n. Each element in the vector takes a value of 0 or 1. A value of 1 indicates that an air duct unit partition needs to be placed between this spatial partition and its adjacent partitions, while a value of 0 indicates that no partition is placed.

[0071] Based on the generated optimal spatial layout scheme, multiple independent airflow unit areas are actually divided within the physical space of the server. The specific implementation method of the division operation is as follows:

[0072] The output boundary identifier vector B is parsed to identify all spatial partition locations with a value of 1. These locations are the boundary locations where air duct unit partitions need to be installed.

[0073] Based on the boundary locations obtained from the analysis, removable air duct partitions are installed at the corresponding locations inside the server. These partitions divide the server's internal space into multiple physically isolated air duct unit areas. Each air duct unit area corresponds to a spatial partition cluster; that is, each heat load cluster is divided into an independent air duct unit. This ensures complete isolation between the air duct units corresponding to hot spots and those corresponding to cold spots, preventing mutual interference.

[0074] At the entrance of each air duct unit area, an independent airflow control valve is installed. This airflow control valve employs a stepper motor-driven louvered valve structure. The valve opening is controlled by the rotation angle of the stepper motor, thereby achieving independent airflow control at the entrance of that air duct unit. Each airflow control valve is equipped with an independent controller, which is connected to the server baseboard management controller via an internal bus, receiving airflow setpoint commands and providing feedback on the current valve opening status.

[0075] Each airflow ratio value rk1 in the calculated airflow distribution ratio R is assigned to the airflow control valve at the inlet of the corresponding duct unit area, serving as the initial airflow setpoint for that duct unit. The specific implementation method is as follows:

[0076] Based on the spatial partition cluster number corresponding to each duct unit region, a correspondence is established between each proportion value rk1 in the air volume allocation ratio R and the duct unit region. Specifically, the k1th air volume proportion value rk1 corresponds to the duct unit region formed by the k1th heat load cluster.

[0077] The total airflow Qtotal of the server's internal cooling fans is obtained. This total airflow is measured in real time by a fan speed sensor or obtained by looking up a table from the fan performance curve. For the k1th air duct unit area, the initial absolute airflow value qk1 is calculated as follows: qk1 equals the total airflow Qtotal multiplied by the airflow ratio rk1.

[0078] The calculated initial absolute airflow value qk1 is converted into a corresponding valve opening command. The conversion relationship is determined by a pre-calibrated valve opening-airflow characteristic curve. The valve opening command is sent to the controller of the airflow control valve at the inlet of the corresponding air duct unit area via the internal bus of the baseboard management controller. The controller drives the stepper motor to rotate to the specified opening position, completing the assignment of the initial airflow setpoint.

[0079] After the valve openings of all duct unit areas are set, the actual air volume is monitored in real time by the flow sensors deployed inside each duct unit and compared with the initial set value. If the deviation exceeds 5%, the valve opening is finely adjusted for compensation and correction to ensure that the actual air volume of each duct unit accurately matches the initial air volume set value.

[0080] The spatial location information of each duct unit area is summarized and organized with the initial airflow setting value of each duct unit area to form a duct unit layout dataset. The spatial location information includes the three-dimensional spatial coordinate range of each duct unit area, the position coordinates and geometric dimensions of the inlet section, the position coordinates and geometric dimensions of the outlet section, and the relative positional relationship between each duct unit.

[0081] Simulation analysis module: Performs fluid dynamics simulation on each air duct unit, obtains the airflow velocity field and temperature field distribution in each air duct unit, and determines whether the preset heat dissipation threshold is met.

[0082] Based on the spatial location information and initial airflow setting of each duct unit area, an initial duct model M0 containing multiple independent duct units is constructed in computer-aided design software, and the initial duct model M0 is then meshed. The specific implementation method is as follows:

[0083] The spatial location information of each air duct unit region is imported into the 3D modeling software. This spatial location information includes the starting and ending coordinates of each air duct unit region in the server's internal coordinate system, the cross-sectional shape and dimensions of the air duct unit, and the 3D model of the key heat-generating components deployed inside the air duct unit. All air duct unit regions are assembled according to their actual spatial locations to form a complete initial air duct model M0. The initial air duct model M0 contains multiple independent air duct units, each with an independent inlet cross-section, outlet cross-section, and internal flow space.

[0084] Based on the initial airflow setting value qk for each duct unit region, a velocity inlet boundary condition is assigned to the inlet section of each duct unit in the initial duct model M0. The inlet velocity value is calculated as follows: inlet velocity vk equals the initial airflow setting value qk divided by the area of ​​the inlet section of the duct unit. A pressure outlet boundary condition is assigned to the outlet section of each duct unit, with the outlet gauge pressure set to 0 Pascal. A no-slip boundary condition is assigned to the walls of all duct units, with the wall roughness set to 0.05 mm.

[0085] The completed initial air duct model M0 was imported into the meshing software, and the model was discretized using an unstructured tetrahedral mesh. Boundary layer meshes were set for the surfaces of key heating elements and the areas near the air duct walls to accurately capture velocity and temperature gradients. The boundary layer mesh parameters were: first layer height of 0.1 mm, growth factor of 1.2, and 5 boundary layers. The overall mesh size was controlled as follows: maximum mesh size not exceeding 2 mm, and minimum mesh size not less than 0.05 mm. After meshing, a mesh file was output, containing the coordinate information of all mesh nodes, the connection relationships of mesh cells, and the identification information of each boundary region.

[0086] The generated mesh file of the initial duct model M0 is input into a computational fluid dynamics solver for iterative solving. By solving the Navier-Stokes equations and the energy equations, the three-dimensional airflow velocity field distribution data and the three-dimensional temperature field distribution data inside each duct unit are obtained. The specific implementation method is as follows:

[0087] A pressure-based steady-state solver is selected, the standard k-ε model is used for turbulence, and the energy equation is enabled. Air is selected as the fluid medium, with air density calculated using an incompressible ideal gas model and air viscosity calculated using the Sutherland formula. The convergence criterion is set as follows: the continuity equation residual is less than 10. -4 The residual of the momentum equation is less than 10. -4 The residual of the energy equation is less than 10. -6 The residuals of the k-equation and the ε-equation are less than 10. -4 .

[0088] The inlet velocity boundary conditions, outlet pressure boundary conditions, and wall boundary conditions of each duct unit are applied to the corresponding boundary regions of the mesh file. Simultaneously, based on the heat dissipation data in the heat load distribution matrix Q, a volumetric heat source boundary condition is applied to each critical heating element. The heat source intensity is equal to the real-time heat dissipation value of that heating element divided by the volume of that element.

[0089] Initialize the flow and temperature fields, and set the relaxation factors: pressure relaxation factor is set to 0.3, momentum relaxation factor is set to 0.7, and energy relaxation factor is set to 1.0. Start iterative calculation, saving intermediate results every 100 iterations. Monitor the average temperature and average velocity values ​​at the outlet section of each duct unit. Stop iterative calculation when the monitored values ​​tend to stabilize and all residuals meet the convergence criteria.

[0090] After iterative convergence, the velocity vector values ​​and temperature scalar values ​​at all grid nodes within each airflow unit are exported from the solver, forming three-dimensional airflow velocity field distribution data and three-dimensional temperature field distribution data. The three-dimensional airflow velocity field distribution data includes the x-direction velocity component, y-direction velocity component, z-direction velocity component, and velocity amplitude of each grid node; the three-dimensional temperature field distribution data includes the temperature value of each grid node.

[0091] From the obtained three-dimensional temperature field distribution data, the temperature values ​​at the surface locations of all key heat-generating components within each air duct unit are extracted. These extracted temperature values ​​are then compared with a pre-set heat dissipation temperature threshold to determine if any heat-generating components exceed the threshold. The specific implementation method is as follows:

[0092] Based on the coordinates of the measuring points of the deployed thin-film thermocouple array, the grid nodes corresponding to the center and edge points of each key heating element's surface are located in the three-dimensional temperature field distribution data. The temperature values ​​at these grid nodes are read as the surface temperature measurements of that heating element. For each heating element, the maximum value of the temperature at all measuring points on its surface is taken as the representative temperature value Tj of that element, where j represents the j-th key heating element.

[0093] Based on the specifications of the key heat-generating components in the AI ​​computing server, the maximum allowable operating temperature for each type of heat-generating component is determined. The maximum allowable operating temperature refers to the upper limit of the temperature at which the component can operate stably for extended periods without performance degradation or damage. Based on this, and considering a certain safety margin, the heat dissipation temperature threshold Tthreshold is set to the maximum allowable operating temperature minus 5 degrees Celsius. Different heat-generating components can be assigned different heat dissipation temperature thresholds; for example, the heat dissipation temperature threshold for the graphics processor is set to 85 degrees Celsius, for the central processing unit (CPU) to 80 degrees Celsius, for the power supply module to 90 degrees Celsius, and for the memory module to 75 degrees Celsius.

[0094] The representative temperature value Tj of each key heat-generating element is compared with the heat dissipation temperature threshold Tthreshold corresponding to that element type. If Tj is less than or equal to Tthreshold, the heat-generating element is determined to meet the heat dissipation requirements; if Tj is greater than Tthreshold, the heat-generating element is determined to not meet the heat dissipation requirements, that is, there are heat-generating elements that exceed the heat dissipation temperature threshold.

[0095] Based on the judgment results, if there are heat-generating elements exceeding the heat dissipation temperature threshold, the air duct unit containing these heat-generating elements is marked as an air duct unit to be optimized, and the marking result is transmitted to the air duct unit parameter adjustment step. The specific implementation method is as follows:

[0096] The system iterates through the judgment result list. For each heat-generating element determined to exceed the heat dissipation temperature threshold, its corresponding air duct unit region is determined based on its spatial location information, and the identification number of that air duct unit region is recorded. After removing duplicates from all recorded identification numbers, a list of air duct unit identification numbers to be optimized is formed.

[0097] The list of duct unit identifiers to be optimized is associated with the detailed information of all duct unit areas to generate a dataset of duct unit labeling results. The dataset contains the identifier number, spatial location information, current initial airflow setting, list of internal heating elements exceeding the threshold, and current temperature value of each element for each duct unit to be optimized.

[0098] Adjust and update the module: If there are areas that do not meet the heat dissipation threshold, adjust the cross-sectional shape, inlet angle or fan speed of the corresponding air duct unit, update the air volume distribution ratio, and re-simulate until all areas meet the heat dissipation requirements to obtain the target asymmetric airflow duct model.

[0099] Based on the labeling results of the duct units to be optimized, a pre-constructed Bayesian optimization-based multi-parameter collaborative adjustment algorithm is used. Through iterative optimization calculations, the adjustment amounts for the cross-sectional shape parameter, inlet angle parameter, and corresponding fan speed parameter of each duct unit to be optimized are generated. The construction process and specific implementation of the Bayesian optimization-based multi-parameter collaborative adjustment algorithm are as follows:

[0100] For each air duct unit to be optimized, three adjustable parameter dimensions are determined: the first parameter is the cross-sectional shape parameter s1, which includes two sub-parameters: cross-sectional width and cross-sectional height, and its adjustment range is 0.8 to 1.2 times the original size; the second parameter is the inlet angle parameter θ, which refers to the angle between the airflow entering the air duct and the horizontal direction, and its adjustment range is -30 degrees to +30 degrees; the third parameter is the fan speed parameter ω, which refers to the rotational speed of the cooling fan supplying air to this air duct unit, and its adjustment range is 0.7 to 1.3 times the original speed. These three parameters together constitute a three-dimensional parameter space.

[0101] The highest temperature value Tmax of the key heat-generating element inside the duct unit to be optimized is used as the observed value of the optimization objective function f(x), where x represents a three-dimensional parameter vector composed of the cross-sectional shape parameter s, the inlet angle parameter θ, and the fan speed parameter ω. Gaussian process regression is used to model the objective function f(x), which is fully defined by the mean function m(x) and the covariance function k(x,x′). The mean function m(x) is set to a constant of 0, and the covariance function k(x,x′) adopts the squared exponential covariance function, the calculation formula of which is: ;in, The signal variance controls the overall variation of the function value. The length scale parameter of the d-th dimension is squared, which controls the smoothness of changes in that dimension; xd and x′d represent the values ​​of the parameter vectors x and x′ in the d-th dimension, respectively.

[0102] The desired improved acquisition function is used to guide the selection of the next sampling point. The formula for calculating the desired improved acquisition function is: EI(x) = E[max(f(x) - f(x)]]. + ),0)];where, f(x + The current best observed value is the minimum high temperature value among the sampled points; the expected improvement value EI(x) represents the expected performance improvement that can be obtained by sampling at parameter point x. The larger the sampling function value, the more potential that parameter point has to become a better solution.

[0103] Initialize the sampling point set. Randomly select 5 initial parameter combinations within the parameter space and perform simulation evaluations to obtain the corresponding maximum temperature value Tmax. Enter the Bayesian optimization main loop: First, update the Gaussian process surrogate model based on all current sampling point data; second, optimize the acquisition function EI(x) across the entire parameter space, finding the parameter point xnew that maximizes the acquisition function value; third, use xnew as a new parameter combination for simulation evaluation to obtain the corresponding maximum temperature value Tmax; fourth, add the new sampling point to the sampling point set. Repeat the above loop until the maximum number of iterations (20) is reached or the improvement in the maximum temperature value is less than 0.5 degrees Celsius after 5 consecutive iterations.

[0104] After Bayesian optimization converges, the optimal parameter combination that minimizes the maximum temperature value Tmax is selected from all sampling points. The current parameter value of the duct unit is subtracted from each parameter value in the optimal parameter combination to obtain the adjustment amounts Δs (cross-sectional shape parameter), Δθ (inlet angle parameter), and Δω (fan speed parameter). These parameter adjustments will be used in subsequent steps to update the geometric model of the duct unit and the fan operating parameters.

[0105] The generated parameter adjustments are applied to the corresponding duct unit to be optimized. Parameter updates are achieved by modifying the geometric model parameters and fan operating parameters of the duct unit, and the airflow distribution ratio RS at the inlet of each duct unit is recalculated. The specific implementation method is as follows:

[0106] For each duct unit to be optimized, its current cross-sectional width value Wcurrent and current cross-sectional height value Hcurrent are multiplied by an adjustment factor, which is determined by the cross-sectional shape parameter adjustment amount Δs. Specifically, the new cross-sectional width value... , New cross-sectional height value This allows for the adjustment of the cross-sectional shape proportions while maintaining a relatively constant cross-sectional area. In 3D modeling software, the geometric model of the duct unit is reconstructed based on the new width and height values.

[0107] The installation angle of the inlet guide plate of the air duct unit to be optimized is increased by Δθ degrees based on the original angle. The inlet guide plate is driven by a stepper motor. The baseboard management controller sends an angle adjustment command to the stepper motor, driving the guide plate to rotate to a new angular position, thereby changing the direction of airflow entering the air duct.

[0108] The current fan speed ωcurrent is multiplied by (1+Δω) to obtain the new fan speed ωnew. The baseboard management controller sends a speed adjustment command to the fan via a pulse width modulation signal to adjust the fan speed to ωnew.

[0109] Obtain the adjusted fan speed values ​​for all duct units, and look up the airflow of each fan at the current speed using the fan performance curve table. The new total air volume is obtained by summing the air supply volumes of all duct units. Then calculate the fresh air volume allocation ratio for each duct unit. The calculation formula is: The calculated air volume distribution ratios of each duct unit are combined into an updated air volume distribution ratio vector RS.

[0110] The updated duct element geometric model and airflow distribution ratio RS are re-input into the computational fluid dynamics solver for iterative solving. By resolving the Navier-Stokes equations and energy equations, new airflow velocity field distribution data and temperature field distribution data are obtained. The specific implementation method is as follows:

[0111] The reconstructed air duct unit geometric model is reassembled to form an updated complete air duct model. The updated air duct model is then re-meshed to generate a new mesh file.

[0112] The calculated updated airflow distribution ratio vector RS is converted into velocity boundary conditions at the inlet of each duct unit. For the k-th duct unit, the new inlet velocity vk is calculated as follows: vk equals the updated airflow qk divided by the updated inlet cross-sectional area of ​​the duct unit. The calculated inlet velocity value is then applied to the inlet cross-section of the corresponding duct unit.

[0113] The updated model is iteratively solved according to the solver parameters, including the turbulence model, convergence criteria, and relaxation factor. The average temperature and average velocity values ​​at the outlet section of each duct unit are monitored. The iterative calculation stops when all residuals meet the convergence criteria and the monitored values ​​are stable.

[0114] The updated 3D airflow velocity field distribution data and 3D temperature field distribution data are exported from the solver and used as the evaluation basis for this round of optimization iteration. The process continues until the surface temperature values ​​of the key heat-generating components inside all duct units are less than or equal to the corresponding heat dissipation temperature threshold. At this point, the combination of the duct unit geometric model and fan operating parameters is determined as the final target asymmetric airflow duct model. The specific implementation method is as follows:

[0115] After each optimization iteration, the temperature values ​​at the surface locations of key heat-generating components inside all air duct units are extracted and compared with the corresponding heat dissipation temperature thresholds. If the representative temperature value Tj of all heat-generating components satisfies Tj≤threshold, the optimization iteration process is terminated; if the temperature value of any heat-generating component still exceeds the threshold, the next optimization iteration is executed.

[0116] During each iteration, the current iteration number, the parameter adjustments for each duct unit to be optimized, the updated airflow distribution ratio, and the highest temperature values ​​of all heating elements are recorded. This recorded data is used to monitor the convergence of the optimization process and can serve as a basis for subsequent analysis.

[0117] When the termination condition is met, the geometric model parameters of the air duct unit corresponding to the current iteration (including the cross-sectional width, cross-sectional height, and inlet angle of each air duct unit) and the fan operating parameters (the rotational speed of each fan) are fixed and saved. These parameters are combined to define the final target asymmetric airflow duct model.

[0118] All geometric and operational parameters corresponding to the target asymmetric airflow duct model are output to the server control system via the data transmission interface of the baseboard management controller, serving as the baseline configuration for duct unit layout and fan control during actual operation. Simultaneously, the model parameters are stored in non-volatile memory for quick retrieval under subsequent identical load conditions.

[0119] Execution module: Controls the internal airflow of the server to perform heat dissipation operations according to the target asymmetric airflow model.

[0120] Based on the determined target asymmetric airflow model, control commands are sent to each actuator through the server baseboard management controller to control the internal airflow of the server to perform heat dissipation operations, achieving precise matching between the asymmetric airflow layout and actual operation. The specific implementation method is as follows:

[0121] Read all parameter data corresponding to the target asymmetric airflow duct model from the server's non-volatile memory. The parameter data includes the cross-sectional width value, cross-sectional height value, inlet angle value, and corresponding fan speed value of each airflow duct unit.

[0122] For servers equipped with adjustable duct partitions, the baseboard management controller sends control commands to the partition position adjustment mechanism, driving the partitions to move to the target position so that the cross-sectional width and height of each duct unit reach the values ​​specified in the target asymmetric airflow model. For servers using fixed partitions, this step is skipped.

[0123] The baseboard management controller sends angle control commands to the stepper motor at the inlet of each air duct unit, driving the inlet guide plate to rotate to the inlet angle value specified in the target asymmetric airflow duct model, ensuring that the airflow enters each air duct unit in the optimal direction.

[0124] The baseboard management controller sends speed control commands to each cooling fan via pulse width modulation signals, adjusting the fan speed to the speed value specified in the target asymmetric airflow model. Simultaneously, closed-loop fan speed control is activated, using the fan's built-in speed sensor to monitor the actual speed in real time. Automatic correction is performed when the actual speed deviates from the target speed by more than 5%.

[0125] During server operation, temperature data is continuously collected in real time by thin-film thermocouple arrays deployed on the surfaces of key heat-generating components and compared with heat dissipation temperature thresholds. If the temperature of all heat-generating components is within the threshold range, the current airflow configuration and fan parameters remain unchanged; if an abnormal temperature increase occurs, the optimization process is retried to dynamically adjust the airflow parameters.

[0126] Through the above steps, the target asymmetric airflow model is accurately deployed in the actual server, so that the airflow layout is precisely matched with the non-uniform heat load distribution inside the server, thereby improving heat dissipation efficiency, reducing fan energy consumption, and ensuring the stable operation of the AI ​​computing server under high load conditions.

[0127] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes 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.

Claims

1. An asymmetric airflow cooling system for AI computing servers, characterized in that: include: Thermal power consumption data acquisition module: Acquires thermal power consumption data of key heat-generating components of AI computing server under typical load conditions, and constructs thermal load distribution matrix inside the server; Layout strategy generation module: Based on the heat load distribution matrix, identify hot and cold areas inside the server and generate an asymmetric airflow layout strategy, including the air volume distribution ratio of each airflow inlet and the airflow direction. Air duct unit construction module: Based on the air volume distribution ratio, multiple independently controllable air duct units are constructed, each corresponding to a different heat generation area; Simulation analysis module: Performs fluid dynamics simulation on each air duct unit, obtains the airflow velocity field and temperature field distribution in each air duct unit, and determines whether the preset heat dissipation threshold is met; Adjust and update the module: If there are areas that do not meet the heat dissipation threshold, adjust the cross-sectional shape, inlet angle or fan speed of the corresponding air duct unit, update the air volume distribution ratio, and re-simulate until all areas meet the heat dissipation requirements to obtain the target asymmetric airflow duct model. Execution module: Controls the internal airflow of the server to perform heat dissipation operations according to the target asymmetric airflow model; Identifying hot and cold areas within a server based on the heat load distribution matrix includes the following steps: The heat load distribution matrix is ​​input into a spatial heat distribution feature extraction model based on graph convolutional neural network to extract the heat load coupling features and spatial correlation features between each spatial partition; the extracted feature vectors are input into the density peak clustering algorithm for cluster analysis to divide adjacent spatial partitions with similar heat load features into the same heat load cluster. Calculate the average heat flux density value of each heat load cluster, and compare the average heat flux density value with a preset heat flux density threshold. If the average heat flux density value is greater than or equal to the preset heat flux density threshold, the cluster is marked as a hot spot area. If the average heat flux density value is less than the preset heat flux density threshold, the cluster is marked as a cold spot area. The graph convolutional layer uses the Chebyshev polynomial approximation method to achieve spectral graph convolution, and its core calculation formula is as follows: ； Among them, H (1) H represents the feature matrix of the l-th layer node. (1+1) W represents the feature matrix of the (l+1)th layer node. (1) Let be the trainable weight matrix of the l-th layer, and σ represent the linear rectified activation function; The normalized Laplace matrix is ​​calculated using the following formula: ； Where A is the adjacency matrix of the graph, A{ij}=1 indicates that there is an edge connection between node i and node j, otherwise A{ij}=0, D is the degree matrix, which is a diagonal matrix, and its diagonal element D{ii} is equal to the sum of all elements in the i-th row of the adjacency matrix A, that is, the number of edges connected to node i.

2. The asymmetric airflow cooling system for AI computing power servers according to claim 1, characterized in that: The thermal power consumption data acquisition module specifically includes the following steps: Under typical load conditions of AI computing servers, temperature data is collected in real time by thin-film thermocouple arrays deployed on the surface of each key heat-generating element; The collected temperature data is input into the pre-built thermal resistance network model. By combining the thermal resistance parameters of each heating element with the inlet temperature of the cooling medium, the real-time heat power consumption of each heating element is derived in reverse. The real-time thermal power consumption value is normalized according to the component type and spatial location to generate an m-row n-column heat load distribution matrix, where the rows of the matrix correspond to different types of heat-generating components and the columns of the matrix correspond to the spatial partitions inside the server.

3. The asymmetric airflow cooling system for AI computing power servers according to claim 1, characterized in that: Constructing multiple independently controllable duct units based on airflow distribution ratios specifically includes the following steps: The air volume distribution ratio is input into the air duct unit layout optimization model based on deep Q network. Taking the internal space partition of the server as nodes, and taking the maximization of heat dissipation efficiency and the minimization of flow resistance as dual objectives, the optimal spatial layout scheme of the air duct unit is generated. According to the optimal spatial layout scheme, multiple independent air duct unit areas are divided inside the server. Each air duct unit area corresponds to a spatial partition cluster, and an independent air volume control valve is set at the entrance of each air duct unit area.

4. The asymmetric airflow cooling system for AI computing power servers according to claim 3, characterized in that: Each air volume ratio value in the air volume distribution ratio is assigned to the air volume control valve at the entrance of the corresponding air duct unit area as the initial air volume setting value of the air duct unit.

5. The asymmetric airflow cooling system for AI computing power servers according to claim 1, characterized in that: Performing fluid dynamics simulations on each air duct unit and determining whether the preset heat dissipation threshold is met includes the following steps: Based on the spatial location information and initial air volume setting of each air duct unit area, an initial air duct model M0 containing multiple independent air duct units is constructed, and the initial air duct model M0 is subjected to mesh division processing. The mesh file of the initial duct model M0 is input into the computational fluid dynamics solver for iterative solution to obtain the three-dimensional airflow velocity field distribution data and the three-dimensional temperature field distribution data inside each duct unit; Extract the temperature values ​​at the surface locations of all key heat-generating elements within each air duct unit, compare the extracted temperature values ​​with a pre-set heat dissipation temperature threshold, and determine whether there are any heat-generating elements that exceed the heat dissipation temperature threshold. If there are heat-generating elements that exceed the heat dissipation temperature threshold, the air duct unit containing the heat-generating element will be marked as an air duct unit to be optimized.

6. The asymmetric airflow cooling system for AI computing power servers according to claim 1, characterized in that: The adjustment of the cross-sectional shape, inlet angle, or fan speed of the corresponding duct unit to update the air volume distribution ratio specifically includes the following steps: Based on the labeling results of the duct unit to be optimized, a multi-parameter collaborative adjustment algorithm based on Bayesian optimization is used to generate the adjustment amount of the cross-sectional shape parameter, the adjustment amount of the inlet angle parameter, and the adjustment amount of the corresponding fan speed parameter of the duct unit to be optimized. The generated parameter adjustments are applied to the corresponding duct unit to be optimized, the geometric model and fan operating parameters of the duct unit are updated, and the air volume distribution ratio RS at the inlet of each duct unit is recalculated.

7. The asymmetric airflow cooling system for AI computing power servers according to claim 6, characterized in that: The updated duct unit geometric model and airflow distribution ratio RS are re-input into the computational fluid dynamics solver for iterative solution to obtain new airflow velocity field distribution data and temperature field distribution data. Until the surface temperature values ​​of the key heat-generating elements inside all duct units are less than or equal to the corresponding heat dissipation temperature threshold, the duct unit geometric model and fan operating parameters at this time are determined as the final target asymmetric airflow duct model.

8. The asymmetric airflow cooling system for AI computing power servers according to claim 2, characterized in that: The key heat-generating components include a graphics processor, a central processing unit, a power supply module, and a memory module.

9. The asymmetric airflow cooling system for AI computing power servers according to claim 2, characterized in that: The calculation formula for the reverse thermal simulation algorithm is: heat dissipation equals surface temperature minus the inlet temperature of the cooling medium, and then divided by the total thermal resistance.

Images