GPU adaptive liquid cooling control method based on deep learning

Abstract

The invention discloses a GPU (Graphics Processing Unit) adaptive liquid cooling control method based on deep learning, relates to the technical field of GPU adaptive liquid cooling control, and overcomes the core defect of the existing GPU liquid cooling control in a sparse load scene through deep fusion of a deep learning method and a physical law. The model structure analyzer dynamically associates the neural network calculation characteristics with the physical layout of the chip to realize the pre-judgment type tracking of the heat source movement track, so that the cooling blind area is reduced from the source; a space-time attention mechanism is combined with a microdiscretization technology, so that cooling liquid distribution can respond to a millisecond-level heat wave peak, and the physical realizability of flow control is kept; and a resistance residual module and thermodynamic hard constraint are introduced, so that the robustness of the system under extreme working conditions such as pipeline abnormity is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of GPU adaptive liquid cooling control, and in particular to a GPU adaptive liquid cooling control method based on deep learning. Background Art

[0002] As large language models evolve towards a sparse architecture, the application of technologies such as hybrid expert MoE has enabled GPU computing units to exhibit dynamic activation characteristics. When performing tasks such as text generation and logical reasoning, expert modules at different levels of the model will be randomly activated at millisecond intervals, resulting in highly localized instantaneous heat loads on the chip surface; the cooling unit needs to quickly respond to randomly appearing heat islands on a microscale while avoiding energy loss caused by excessive cooling of inactive areas.

[0003] The current mainstream adaptive liquid cooling solutions mainly adopt a dynamic partitioning control strategy, which realizes regional regulation of cold plate flow through a temperature sensor array. Existing liquid-cooled GPUs have introduced reinforcement learning controllers to adjust the pump speed based on historical load predictions. Such solutions perform well under continuous uniform load scenarios, but have significant defects when facing sudden heat wave spikes in the MoE model: sensor feedback delay causes the response action to lag behind the heat source migration speed, resulting in passive catch-up cooling; in order to ensure the overall uniform temperature of the chip, the global optimization algorithm often imposes unnecessary cooling intensity on non-core areas, resulting in wasted pump power.

[0004] To address the above bottlenecks, some solutions attempt to establish a mapping relationship between the neural network layer and the physical chip coordinates, predict the movement trajectory of the heat source by pre-loading the model architecture information; use generative AI to design a microfluidic structure with fractal characteristics, so that the coolant can autonomously gather in the high thermal conductivity area; although such solutions improve response agility, when the sparse activation pattern of the model exceeds the distribution of training data, the system may still fall into the local optimization trap. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides a deep learning-based GPU adaptive liquid cooling control method to solve the problem of local instantaneous high heat of the GPU caused by sparse large models, and the slow response and low energy efficiency of existing liquid cooling solutions.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] An embodiment of the present invention provides a GPU adaptive liquid cooling control method based on deep learning, which includes:

[0009] Step S1: construct a model structure parser to parse the connection relationship between the neural network layers running on the GPU and generate a mapping relationship table between the computing units and the physical coordinates of the chip;

[0010] Step S2: Deploy a dual-channel time series prediction network. The first channel receives the sparsity feature vector of the model activation pattern, and the second channel receives the chip temperature distribution matrix within a preset time window, and outputs the probability distribution of heat load in each area in the future time slice.

[0011] Step S3, generating a coolant dynamic allocation strategy based on the heat load probability distribution, and mapping the continuous probability to the flow control instructions of each branch of the cold plate through a differentiable discretization algorithm;

[0012] In step S4, an adversarial residual module is embedded in the controller, which receives a preset fault mode feature vector and generates a flow compensation signal to offset the cooling efficiency attenuation caused by pipeline abnormality.

[0013] As a preferred solution of the GPU adaptive liquid cooling control method based on deep learning described in the present invention, the mapping relationship table in step S1 is constructed in the following manner:

[0014] Extracting operator dependencies of neural network computation graphs;

[0015] Map dense matrix multiplication operators to the dense floating-point operation unit area of ​​the GPU chip;

[0016] Mapping sparse attention operators to the chip’s cached neighborhood;

[0017] The dense matrix multiplication operators are classified according to the operators whose MAC / cycle values ​​in the computation graph feature vector are greater than a threshold, and the floating-point operation unit dense area is an area on the chip that contains more than 80% of the ALU logic units.

[0018] As a preferred solution of the GPU adaptive liquid cooling control method based on deep learning described in the present invention, the attention mechanism adopted by the dual-channel timing prediction network in step S2 includes a trainable position encoding module, which embeds the chip physical coordinate information into the feature vector output by the first channel, and establishes the correlation between the thermal load probability distribution and the spatial position by calculating the attention score matrix. Finally, the fused feature tensor is filtered by the gated linear unit and input into the fully connected layer.

[0019] As a preferred solution of the GPU adaptive liquid cooling control method based on deep learning described in the present invention, wherein: in step S2, the process of establishing the correlation between the heat load probability distribution and the spatial position first constructs a query key pair in the attention space, and then outputs the probability matrix through the gating unit, including:

[0020] Define the query vector as:

[0021] q=f t W q ,

[0022] Where q represents the query vector, f t represents the sparse activation feature vector at time t, t represents the time index, W q is a trainable linear mapping matrix;

[0023] Define the key vector as:

[0024]

[0025] Among them, k ij is the key vector of unit (i, j), i and j represent the hot grid row and column index respectively, E ij is the physical position vector, W k is the linear mapping matrix, Indicates the average window temperature, w t is the trainable weight vector;

[0026] The attention score calculation formula is:

[0027]

[0028] Among them, α ij is the normalized score, d q is the query vector dimension, m, n are the sum indexes, M, N are the number of hot grid rows and columns respectively;

[0029] The heat load probability calculation formula is:

[0030] p ij =σ(α ij g t +b),

[0031] Among them, p ij is the heat load probability, σ is the Sigmoid function, g t is the gate scale coefficient, and b is the bias constant.

[0032] As an optimal solution of the GPU adaptive liquid cooling control method based on deep learning described in the present invention, the differentiable discretization algorithm adopts a combination of relaxed probability distribution and hard threshold truncation when performing flow distribution. The forward propagation stage generates the flow ratio of each branch through the relaxed probability matrix, and the backward propagation stage constructs the proxy gradient to retain the discrete decision-making characteristics, wherein the discretization degree of the flow control instruction is dynamically adjusted by the temperature-sensitive parameter.

[0033] As a preferred solution of the GPU adaptive liquid cooling control method based on deep learning described in the present invention, the differentiable discretization algorithm described in step S3 adopts the Gumbel-Softmax relaxation technology, and controls the discretization degree of flow distribution through the temperature parameter, wherein the temperature value decreases exponentially with the increase of the average chip temperature.

[0034] As a preferred solution of the GPU adaptive liquid cooling control method based on deep learning described in the present invention, in step S3, the step of performing flow distribution includes:

[0035] The branch load aggregation is defined as,

[0036]

[0037] Among them, y b is the load aggregation of branch b, S b is the set of grids covered by branch b, ε is a very small positive number to avoid logarithmic singularity;

[0038] The relaxation weight is defined as:

[0039]

[0040] in, is the relaxation weight of branch b, g b is Gumbel noise, τ is the temperature parameter, c is the branch index, g c is the Gumbel(0,1) random perturbation of branch c, and g b Same distribution, B is the total number of branches;

[0041] The flow instructions are defined as:

[0042]

[0043] Among them, F b is the branch flow instruction, ρ is the discrete weight coefficient, h b is a hard weight, when The maximum value is 1, and the rest are 0. tot is the total pump flow;

[0044] Discrete weights

[0045] Where κ is the sensitivity coefficient, is the current average temperature of the chip, and T0 is the temperature threshold.

[0046] As a preferred solution of the GPU adaptive liquid cooling control method based on deep learning described in the present invention, the adversarial residual module in step S4 includes a generator and a discriminator. The generator receives the current pipeline pressure characteristics and historical fault codes and outputs a flow compensation increment. The discriminator determines whether the compensation amount exceeds the physically feasible region based on the law of conservation of thermodynamics.

[0047] Hard-coded thermodynamic constraints in the controller include:

[0048] The heat dissipation power of a single area shall not exceed the theoretical power consumption of the corresponding computing unit;

[0049] The coolant temperature gradient in adjacent areas does not exceed the material thermal expansion safety threshold;

[0050] The ratio of the total pump power consumption to the chip's real-time power consumption is maintained within a preset range.

[0051] As a preferred solution of the deep learning-based GPU adaptive liquid cooling control method described in the present invention, it also includes: an offline strategy optimization stage, generating a thermal interference spectrum diagram under a multi-task load scenario through numerical simulation, and using the spectrum energy distribution to constrain the output layer activation function of the control strategy network;

[0052] The numerical simulation uses the finite element method to solve the unsteady Navier-Stokes equations to generate a coupled data set including velocity field, pressure field and temperature field. The spectrum diagram extracts energy distribution characteristics in the 0.1-10 Hz frequency band through short-time Fourier transform.

[0053] As a preferred solution of the deep learning-based GPU adaptive liquid cooling control method described in the present invention, in the offline strategy optimization stage, a velocity, pressure, and temperature three-field coupled data set is established for a multi-task load scenario. The cold plate chip flow channel domain Ω is first divided into a tetrahedral grid, where Ω represents the cold plate-chip flow channel calculation domain. Then, a weighted residual form is constructed for the unsteady-state equation system, including:

[0054] Construct the momentum-pressure weak form:

[0055]

[0056] Where a represents the fluid density, u represents the velocity vector, t represents the time, v is the momentum test function, c1 represents the dynamic viscosity, is the gradient bilinear inner product, p represents the pressure field;

[0057] Construct the weak form of mass conservation:

[0058]

[0059] in, is the pressure test function;

[0060] Construct the energy weak form:

[0061]

[0062] Where c2 represents specific heat capacity, T represents temperature field, s represents energy test function, d1 represents thermal conductivity, Φ t Indicates instantaneous body heat source;

[0063] Using the implicit Euler format, denoting Δt as the time step, the semi-discrete equation of momentum in the n+1th step is written as:

[0064] A u u n+1 +C u (u n )u n+1 +K u u n+1 -Gp n+1 =A u u n ,

[0065] Among them, A u is the mass matrix, C u (u n ) is the convection matrix, K u is the viscosity matrix, G is the gradient matrix, u n ,u n+1 Respectively represent the discrete velocity vector at time n, n+1;

[0066] The energy equation is written at the same step size as:

[0067] A T T n+1 +C T (u n+1 )T n+1 +K T T n+1 =A T T n +F n+1 ,

[0068] Among them, A T is the thermal mass matrix, C T is the convection thermal coupling matrix, K T is the thermal conductivity matrix, F n+1 is the heat source vector;

[0069] The numerical solution process includes:

[0070] Prediction correction: solve the velocity and pressure block first, then update the temperature.

[0071] Adaptive time step: Decrement Δt according to the maximum Courant number constraint,

[0072] Multi-task boundary: Load the power distribution boundary conditions for each operator and solve them cyclically to obtain T max Indicates the total offline simulation time, which is used to truncate the transient solution;

[0073] In each scene, u, p, T are sampled at fixed intervals and stored as tensors in, Indicates the Scene No. Frame coupling sample tensor, l is the scene index, k is the time frame index, forming a coupled sample set containing velocity field, pressure field, and temperature field, which is used for subsequent spectrum energy constraint and strategy network training.

[0074] The beneficial effects of the present invention are as follows: the present invention overcomes the core defects of existing GPU liquid cooling control in sparse load scenarios through the deep integration of deep learning methods and physical laws; the model structure analyzer dynamically associates the computing characteristics of the neural network with the physical layout of the chip to realize the predictive tracking of the movement trajectory of the heat source, thereby reducing the cooling blind spot from the root; the combination of spatiotemporal attention mechanism and differentiable discretization technology enables the coolant distribution to respond to millisecond-level heat wave spikes while maintaining the physical feasibility of flow control; the introduction of adversarial residual modules and thermodynamic hard constraints significantly improves the robustness of the system under extreme working conditions such as pipeline abnormalities; compared with traditional temperature averaging strategies, this solution effectively reduces ineffective cooling energy consumption while ensuring chip safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0076] Figure 1 This is a flow chart of the deep learning-based GPU adaptive liquid cooling control method in Example 1. DETAILED DESCRIPTION

[0077] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0078] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0079] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0080] Example 1, with reference to Figure 1 , this embodiment provides a GPU adaptive liquid cooling control method based on deep learning, comprising the following steps:

[0081] Step S1: construct a model structure parser to parse the connection relationship between the neural network layers running on the GPU and generate a mapping relationship table between the computing units and the physical coordinates of the chip;

[0082] The mapping relationship table in step S1 is constructed in the following way:

[0083] Extracting operator dependencies of neural network computation graphs;

[0084] Map dense matrix multiplication operators to the dense floating-point operation unit area of ​​the GPU chip;

[0085] Mapping sparse attention operators to the chip’s cached neighborhood;

[0086] Dense matrix multiplication operators are classified according to the operators with MAC / cycle values ​​greater than the threshold in the computation graph feature vector. The floating-point unit dense area is the area on the chip that contains more than 80% of the ALU logic units;

[0087] Step S2: Deploy a dual-channel time series prediction network. The first channel receives the sparsity feature vector of the model activation pattern, and the second channel receives the chip temperature distribution matrix within a preset time window, and outputs the probability distribution of heat load in each area in the future time slice.

[0088] The attention mechanism used by the dual-channel timing prediction network in step S2 includes a trainable position encoding module, which embeds the chip physical coordinate information into the feature vector output by the first channel. The correlation between the thermal load probability distribution and the spatial position is established by calculating the attention score matrix. The fused feature tensor is finally filtered by the gated linear unit and input into the fully connected layer.

[0089] In step S2, in the process of establishing the correlation between the heat load probability distribution and the spatial position, the query key pair is first constructed in the attention space, and then the probability matrix is ​​output through the gating unit, including:

[0090] Define the query vector as:

[0091] q=f t W q ,

[0092] Where q represents the query vector, f t represents the sparse activation feature vector at time t, t represents the time index, W q is a trainable linear mapping matrix;

[0093] Define the key vector as:

[0094]

[0095] Among them, k ij is the key vector of unit (i, j), i and j represent the hot grid row and column index respectively, E ij is the physical position vector, W k is the linear mapping matrix, Indicates the average window temperature, w t is the trainable weight vector;

[0096] The attention score calculation formula is:

[0097]

[0098] Among them, α ij is the normalized score, d q is the query vector dimension, m, n are the sum indexes, M, N are the number of hot grid rows and columns respectively;

[0099] The heat load probability calculation formula is:

[0100] p ij =σ(α ij g t +b),

[0101] Among them, p ij is the heat load probability, σ is the Sigmoid function, g t is the gate scale coefficient, b is the bias constant;

[0102] Specifically, the coupling process first compresses sparse activation information with linear mapping, then injects physical coordinates and temperature history into the key vector to achieve dual-channel fusion of spatial functions. Dot product attention normalizes all units simultaneously, automatically balancing hot zone competition. Gating scale enables the model to dynamically adjust probability gradients at different load stages. Sigmoid restricts the output range to maintain probabilistic characteristics. The entire mechanism can obtain a probability matrix with an M×N resolution in a single forward propagation, providing coherent, differentiable, and physically consistent input for subsequent traffic distribution.

[0103] Step S3, generating a coolant dynamic allocation strategy based on the heat load probability distribution, and mapping the continuous probability to the flow control instructions of each branch of the cold plate through a differentiable discretization algorithm;

[0104] The differentiable discretization algorithm combines a relaxed probability distribution with a hard threshold cutoff when performing flow allocation. In the forward propagation phase, the relaxed probability matrix is ​​used to generate the flow ratio of each branch. In the backward propagation phase, a proxy gradient is constructed to preserve the discrete decision-making characteristics. The degree of discretization of the flow control instructions is dynamically adjusted by a temperature-sensitive parameter.

[0105] In step S3, the differentiable discretization algorithm uses the Gumbel-Softmax relaxation technique to control the discretization degree of the flow distribution through the temperature parameter, where the temperature value decreases exponentially as the average chip temperature increases;

[0106] In step S3, the steps of performing traffic distribution include:

[0107] The branch load aggregation is defined as,

[0108]

[0109] Among them, y b is the load aggregation of branch b, S b is the set of grids covered by branch b, ε is a very small positive number to avoid logarithmic singularity;

[0110] The relaxation weight is defined as:

[0111]

[0112] in, is the relaxation weight of branch b, g b is Gumbel noise, τ is the temperature parameter, c is the branch index, g c is the Gumbel(0,1) random perturbation of branch c, and g b Same distribution, B is the total number of branches;

[0113] The flow instructions are defined as:

[0114]

[0115] Among them, F b is the branch flow instruction, ρ is the discrete weight coefficient, h b is a hard weight, when The maximum value is 1, and the rest are 0. tot is the total pump flow;

[0116] Discrete weights

[0117] Where κ is the sensitivity coefficient, is the current average temperature of the chip, T0 is the temperature threshold;

[0118] Specifically, using the probability cumulant y b Measure the thermal intensity of each branch, map the logarithmic load to a derivable weight through Gumbel Soft max, and the noise g b To ensure that soft decisions are exploratory during training, the hardening coefficient h b During the inference phase, a single on-off valve control is provided to avoid frequent valve fine-tuning. The discrete weight ρ decays exponentially with the average temperature, quickly locking the hottest branch during high-temperature phases and smoothly distributing flow during low-temperature phases to suppress pump power consumption. This allocation strategy combines gradient propagation during training with physical feasibility during execution, achieving dynamic optimal control of multi-branch cold plates.

[0119] Step S4: embedding an adversarial residual module in the controller, receiving a preset fault mode feature vector, and generating a flow compensation signal to offset the cooling efficiency attenuation caused by pipeline abnormality;

[0120] In step S4, the adversarial residual module includes a generator and a discriminator. The generator receives the current pipeline pressure characteristics and historical fault codes and outputs the flow compensation increment; the discriminator determines whether the compensation amount exceeds the physical feasible region based on the law of conservation of thermodynamics.

[0121] Hard-coded thermodynamic constraints in the controller include:

[0122] The heat dissipation power of a single area shall not exceed the theoretical power consumption of the corresponding computing unit;

[0123] The coolant temperature gradient in adjacent areas does not exceed the material thermal expansion safety threshold;

[0124] The ratio of total pump power consumption to chip real-time power consumption is maintained within a preset range;

[0125] The GPU adaptive liquid cooling control method also includes an offline strategy optimization phase:

[0126] Through numerical simulation, we generate thermal interference spectrum under multi-task load scenario, and use spectrum energy distribution to constrain the output layer activation function of the control strategy network.

[0127] The numerical simulation uses the finite element method to solve the unsteady Navier-Stokes equations to generate a coupled data set containing velocity field, pressure field, and temperature field. The spectrum is then analyzed using short-time Fourier transform to extract the energy distribution characteristics in the 0.1-10 Hz frequency band.

[0128] During the offline strategy optimization phase, a coupled data set of velocity, pressure, and temperature fields is established for the multi-task load scenario. The cold plate-chip flow channel domain Ω is first divided into a tetrahedral mesh, where Ω represents the cold plate-chip flow channel computational domain. A weighted residual form is then constructed for the unsteady equations, including:

[0129] Construct the momentum-pressure weak form:

[0130]

[0131] Where a represents the fluid density, u represents the velocity vector, t represents the time, v is the momentum test function, c1 represents the dynamic viscosity, is the gradient bilinear inner product, p represents the pressure field;

[0132] Construct the weak form of mass conservation:

[0133]

[0134] in, is the pressure test function;

[0135] Construct the energy weak form:

[0136]

[0137] Where c2 represents specific heat capacity, T represents temperature field, s represents energy test function, d1 represents thermal conductivity, Φ t Indicates instantaneous body heat source;

[0138] Using the implicit Euler format, denoting Δt as the time step, the semi-discrete equation of momentum in the n+1th step is written as:

[0139] A u u n+1 +C u (u n )u n+1 +K u u n+1 -Gp n+1 =A u u n ,

[0140] Among them, A u is the mass matrix, C u (u n ) is the convection matrix, K u is the viscosity matrix, G is the gradient matrix, un ,u n+1 Respectively represent the discrete velocity vector at time n, n+1;

[0141] The energy equation is written at the same step size as:

[0142] A T T n+1 +C T (u n+1 )T n+1 +K T T n+1 =A T T n +F n+1 ,

[0143] Among them, A T is the thermal mass matrix, C T is the convection thermal coupling matrix, K T is the thermal conductivity matrix, F n+1 is the heat source vector;

[0144] The numerical solution process includes:

[0145] Prediction correction: solve the velocity and pressure block first, then update the temperature.

[0146] Adaptive time step: Decrement Δt according to the maximum Courant number constraint,

[0147] Multi-task boundary: Load the power distribution boundary conditions for each operator and solve them cyclically to obtain T max Indicates the total offline simulation time, which is used to truncate the transient solution;

[0148] In each scene, u, p, T are sampled at fixed intervals and stored as tensors in, Indicates the Scene No. Frame coupling sample tensor, where l is the scene index and k is the time frame index, forms a coupled sample set containing velocity field, pressure field, and temperature field, which is used for subsequent spectrum energy constraint and strategy network training;

[0149] Specifically, the weak finite element form projects the unsteady control equations into the experimental space, automatically satisfies the continuity of boundary conditions, implicit time discretization improves large step stability, the coupling of mass and convection matrix keeps numerical dissipation to a minimum, adaptive step size suppresses numerical oscillations in high Courant regions, and ensures that multi-task loads still converge under transient high thermal shocks. The velocity, pressure, and temperature sequential splitting reduces the solution dimension of the coupled system and improves the simulation throughput. The final sample covers the complete velocity vortex, pressure gradient, and temperature laminar disturbance transition information, providing high-fidelity supervision signals for the control strategy network, and realizing the joint learning of the three physical quantities of convection, heat, and structure.

[0150] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention 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 invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A GPU adaptive liquid cooling control method based on deep learning, characterized in that: include, Step S1: construct a model structure parser to parse the connection relationship between the neural network layers running on the GPU and generate a mapping relationship table between the computing units and the physical coordinates of the chip; Step S2: Deploy a dual-channel time series prediction network. The first channel receives the sparsity feature vector of the model activation pattern, and the second channel receives the chip temperature distribution matrix within a preset time window, and outputs the probability distribution of heat load in each area in the future time slice. Step S3, generating a coolant dynamic allocation strategy based on the heat load probability distribution, and mapping the continuous probability to the flow control instructions of each branch of the cold plate through a differentiable discretization algorithm; In step S4, an adversarial residual module is embedded in the controller, which receives a preset fault mode feature vector and generates a flow compensation signal to offset the cooling efficiency attenuation caused by pipeline abnormality.

2. The GPU adaptive liquid cooling control method based on deep learning according to claim 1, characterized in that: The mapping relationship table in step S1 is constructed in the following manner: Extracting operator dependencies of neural network computation graphs; Map dense matrix multiplication operators to the dense floating-point operation unit area of ​​the GPU chip; Mapping sparse attention operators to the chip’s cached neighborhood; The dense matrix multiplication operators are classified according to the operators whose MAC / cycle values ​​in the computation graph feature vector are greater than a threshold, and the floating-point operation unit dense area is an area on the chip that contains more than 80% of the ALU logic units.

3. The GPU adaptive liquid cooling control method based on deep learning according to claim 1, characterized in that: The attention mechanism adopted by the dual-channel timing prediction network in step S2 includes a trainable position encoding module, which embeds the chip physical coordinate information into the feature vector output by the first channel, establishes the correlation between the thermal load probability distribution and the spatial position by calculating the attention score matrix, and finally the fused feature tensor is filtered by the gated linear unit and input into the fully connected layer.

4. The GPU adaptive liquid cooling control method based on deep learning according to claim 3, characterized in that: In step S2, in the process of establishing the correlation between the heat load probability distribution and the spatial position, a query key pair is first constructed in the attention space, and then a probability matrix is ​​outputted through the gating unit, including: Define the query vector as: q=f t W q , Where q represents the query vector, f t represents the sparse activation feature vector at time t, t represents the time index, W q is a trainable linear mapping matrix; Define the key vector as: Among them, k ij is the key vector of unit (i, j), i and j represent the hot grid row and column index respectively, E ij is the physical position vector, W k is the linear mapping matrix, Indicates the average window temperature, w t is the trainable weight vector; The attention score calculation formula is: Among them, α ij is the normalized score, d q is the query vector dimension, m, n are the sum indexes, M, N are the number of hot grid rows and columns respectively; The heat load probability calculation formula is: p ij =σ(α ij g t +b), Among them, p ij is the heat load probability, σ is the Sigmoid function, g t is the gate scale coefficient, and b is the bias constant.

5. The GPU adaptive liquid cooling control method based on deep learning according to claim 1, characterized in that: The proposed differentiable discretization algorithm combines a relaxed probability distribution with a hard threshold cutoff when performing flow allocation. In the forward propagation phase, the relaxed probability matrix is ​​used to generate the flow ratio of each branch. In the backward propagation phase, a proxy gradient is constructed to retain the discrete decision-making characteristics. The discretization degree of the flow control instructions is dynamically adjusted by a temperature-sensitive parameter.

6. The GPU adaptive liquid cooling control method based on deep learning according to claim 5, characterized in that: The differentiable discretization algorithm in step S3 adopts the Gumbel-Softmax relaxation technique, and controls the discretization degree of the flow distribution through the temperature parameter, wherein the temperature value decreases exponentially as the average chip temperature increases.

7. The GPU adaptive liquid cooling control method based on deep learning according to claim 6, characterized in that: In step S3, the steps of performing traffic distribution include: The branch load aggregation is defined as, Among them, y b is the load aggregation of branch b, S b is the set of grids covered by branch b, ε is a very small positive number to avoid logarithmic singularity; The relaxation weight is defined as: in, is the relaxation weight of branch b, g b is Gumbel noise, τ is the temperature parameter, c is the branch index, g c is the Gumbel(0,1) random perturbation of branch c, and g b Same distribution, B is the total number of branches; The flow instructions are defined as: Among them, F b is the branch flow instruction, ρ is the discrete weight coefficient, h b is a hard weight, when The maximum value is 1, and the rest are 0. tot is the total pump flow; Discrete weights Where κ is the sensitivity coefficient, is the current average temperature of the chip, and T0 is the temperature threshold.

8. The GPU adaptive liquid cooling control method based on deep learning according to claim 1, characterized in that: The adversarial residual module in step S4 includes a generator and a discriminator. The generator receives the current pipeline pressure characteristics and historical fault codes and outputs the flow compensation increment; the discriminator determines whether the compensation amount exceeds the physical feasible region based on the law of conservation of thermodynamics. Hard-coded thermodynamic constraints in the controller include: The heat dissipation power of a single area shall not exceed the theoretical power consumption of the corresponding computing unit; The coolant temperature gradient in adjacent areas does not exceed the material thermal expansion safety threshold; The ratio of the total pump power consumption to the chip's real-time power consumption is maintained within a preset range.

9. The GPU adaptive liquid cooling control method based on deep learning according to claim 8, characterized in that: It also includes an offline strategy optimization phase, which generates a thermal interference spectrum under a multi-task load scenario through numerical simulation, and uses the spectrum energy distribution to constrain the output layer activation function of the control strategy network; The numerical simulation uses the finite element method to solve the unsteady Navier-Stokes equations to generate a coupled data set including velocity field, pressure field and temperature field. The spectrum diagram extracts energy distribution characteristics in the 0.1-10 Hz frequency band through short-time Fourier transform.

10. The GPU adaptive liquid cooling control method based on deep learning according to claim 9, characterized in that: In the offline strategy optimization phase, a coupled data set of velocity, pressure, and temperature fields is established for the multi-task load scenario. The cold plate chip flow channel domain Ω is first divided into a tetrahedral grid, where Ω represents the cold plate-chip flow channel computational domain. A weighted residual form is then constructed for the unsteady state equations, including: Construct the momentum-pressure weak form: Where a represents the fluid density, u represents the velocity vector, t represents the time, v is the momentum test function, c1 represents the dynamic viscosity, is the gradient bilinear inner product, p represents the pressure field; Construct the weak form of mass conservation: in, is the pressure test function; Construct the energy weak form: Where c2 represents specific heat capacity, T represents temperature field, s represents energy test function, d1 represents thermal conductivity, Φ t Indicates instantaneous body heat source; Using the implicit Euler format, denoting Δt as the time step, the semi-discrete equation of momentum in the n+1th step is written as: AND u in n+1 +C u (in n )in n+1 +K u in n+1 -Gp n+1 =A u in n , Among them, A u is the mass matrix, C u (u n ) is the convection matrix, K u is the viscosity matrix, G is the gradient matrix, u n ,u n+1 Respectively represent the discrete velocity vector at time n, n+1; The energy equation is written at the same step size as: A T T n+1 +C T (u n+1 )T n+1 +K T T n+1 =A T T n +F n+1 , Among them, A T is the thermal mass matrix, C T is the convection thermal coupling matrix, K T is the thermal conductivity matrix, F n+1 is the heat source vector; The numerical solution process includes: Prediction correction: solve the velocity and pressure block first, then update the temperature. Adaptive time step: Decrement Δt according to the maximum Courant number constraint, Multi-task boundary: Load the power distribution boundary conditions for each operator and solve them cyclically to obtain T max Indicates the total offline simulation time, which is used to truncate the transient solution; In each scene, u, p, T are sampled at fixed intervals and stored as tensors in, Indicates scene l Frame coupling sample tensor, l is the scene index, k is the time frame index, forming a coupled sample set containing velocity field, pressure field, and temperature field, which is used for subsequent spectrum energy constraint and strategy network training.

Images