Frequency domain-space domain fusion sensing method for rapid and high-precision thermal prediction of 2.5 D integrated circuit

Abstract

The invention relates to a frequency domain-space domain fusion sensing method for rapid high-precision thermal prediction of a 2.5 D integrated circuit. A network comprises a frequency-space thermal encoder module, a frequency domain cross-scale interaction module, a frequency domain-space domain mixed loss function (FSL) and a model input and preprocessing network (PPNet). The frequency air-heat encoder module is combined with adaptive multi-frequency embedding and a 3D convolutional neural network to extract heat dissipation gradient features and detailed spatial features from high frequency to low frequency; the frequency domain cross-scale interaction module realizes multi-scale feature interaction through a cross-attention mechanism and solves the semantic gap problem; and the frequency domain-space domain mixed loss function is combined with the space mean square error and the frequency domain loss to suppress the high-frequency thermal gradient noise. According to the method, frequency domain-space domain double-domain sensing thermal prediction is realized for the first time, the prediction precision and speed are remarkably improved, and the method has excellent generalization ability and can be widely applied to the field of thermal management of 2.5 D integrated circuits.

Description

Technical Field

[0001] The present invention relates to the technical field of integrated circuit thermal management, and in particular to a frequency domain-spatial domain fusion perception method (FSA-Heat) for fast and high-precision thermal prediction of 2.5D integrated circuits. Background Art

[0002] In the post-Moore era, although 2.5D / 3D chip technology has achieved higher density and lower power consumption designs, it has also brought about increased power density and hot spot problems, which seriously affect system reliability and life. Thermal analysis is essential for detecting hot spots on chips, providing thermally aware placement / layout solutions, and assisting rapid design turnaround. However, traditional thermal simulation methods such as the finite element method (FEM) and the finite difference method (FDM) are highly accurate but consume large computing resources, making it difficult to meet the needs of repeated thermal simulations during the design optimization of complex tasks.

[0003] In recent years, neural networks have been introduced to accelerate the calculation of discrete heat equations, but existing methods based on convolutional neural networks (CNN) and graph neural networks (GCN) are insufficient in capturing global thermal features, especially high-frequency components, which limits the improvement of prediction accuracy. For example, the CNN method is mainly used for 2D analysis. Due to its local perception characteristics and low-frequency deviation, it cannot effectively perceive long-distance relationships and changing heat dissipation gradients in each heat transfer layer, thus affecting the accuracy of 2.5D thermal prediction. In addition, methods based on graph neural networks and neural operators are also insufficient in global feature extraction, making it difficult to cope with more fine-grained and complex 2.5D chip designs.

[0004] In addition, although the existing global self-attention mechanism can perceive global information, it mainly focuses on low-frequency features and is still insufficient in extracting high-frequency global information, which is crucial for improving the prediction accuracy of hot spots and thermal dissipation gradients. Therefore, developing a thermal prediction model that can effectively capture multi-frequency global to local thermal dissipation features is of great significance for solving the thermal management problem of 2.5D integrated circuits. Summary of the invention

[0005] Purpose of the invention: The present invention aims to provide a frequency domain-spatial domain fusion perception method (FSA-Heat) for fast and high-precision thermal prediction of 2.5D integrated circuits, so as to solve the problem that the thermal prediction model in the prior art cannot effectively capture the global thermal characteristics and high-frequency components, and realize fast and high-precision thermal prediction.

[0006] In order to achieve the above object, the present invention provides a frequency domain-spatial domain fusion perception method (FSA-Heat) for fast and high-precision thermal prediction of 2.5D integrated circuits, comprising the following steps:

[0007] S1: The present invention first converts the geometric properties, heat exchange conditions and heat source parameters of the 2.5D integrated circuit into a picture form and performs normalization preprocessing. The geometric properties include material thickness t, length l, width w, and heat sink side length s; the heat exchange conditions include the convection resistance coefficient h r , thermal conductivity k; heat source parameters include layered heat source power Q s , total heat source power Q t .

[0008] S2: The parameters of step S1 are processed through the parameter preprocessing network (PPNet). PPNet uses the 3D stem layer to reduce the dimension of the input parameters, and then sends them to the K-layer convolution backbone to obtain the fusion map.

[0009] S3: After upsampling the fusion image, the preprocessed geometric and heat exchange parameter feature maps are obtained and concatenated with the original heat source parameters as the input of the frequency-spatial domain perception prediction network.

[0010] S4: The frequency-space hot encoder module uses a deep convolutional layer followed by two branches to process frequency domain and spatial domain features respectively. In the frequency domain branch, the three-dimensional discrete cosine transform (3D-DCT) is used to map the spatial features to the frequency domain to obtain the three-dimensional anisotropic frequency features, and combined with the learnable frequency weight W f The features of different frequency bands are adaptively adjusted, and the frequency features are converted back to the spatial domain through 3D inverse discrete cosine transform (3D-IDCT). The frequency branch uses the multiplication gated signal attention mechanism to dynamically adjust the feature importance. The learnable weight formula is:

[0011]

[0012] Due to the anisotropy of heat conduction, W f The three directions (u, w, p) in the Fourier domain and the high-frequency to low-frequency band (E) corresponding to the three heat conduction directions (x, y, z) in the spatial domain are u ,E w ,E p ) is embedded.

[0013] S5: In the spatial domain branch, a residual block consisting of two layers of 3D convolution, instance normalization, and SILU activation layers is used to extract local spatial features.

[0014] S6: Fuse the features of the frequency domain and spatial domain branches to form a comprehensive feature representation.

[0015] S7: The frequency domain cross-scale interaction module is used to fuse multi-scale global to local heat dissipation features to alleviate semantic gaps. First, the embedding vector output by the multi-scale frequency-space heat encoder module is added to the position embedding, and the query vector q is obtained through layer normalization (LN) and linear layer. 1 ,q 2 ,q 3 ,q 4 and fuse the embedded key k, value v.

[0016] S8: Convert the query vector and key to the frequency domain through 3D-DCT to obtain the frequency domain coordinate representation. Calculate the attention score S of vectors at different scales to represent the semantic similarity. The specific formula is:

[0017]

[0018] S represents the attention score, q i and k represent the multi-head query and key matrix respectively. k is the dimension of the value matrix k.

[0019] S9: Using 3D-IDCT and weighted summation to obtain feature f i , and then enrich the semantic content through the feedforward network (FFN) to enhance the feature expression ability. F Upsampling to align the decoder and ensure feature sizes match.

[0020] S10: The decoding block first performs trilinear interpolation upsampling on the feature map output by the encoding block, followed by 1×1×1 convolution, instance normalization, and GELU activation to obtain the upsampled feature map to be processed.

[0021] S11: The upsampled feature map to be processed is concatenated with the output of the frequency domain cross-scale interaction module and then input into the residual block of the next decoding layer. The residual block contains two layers of 3×3×3 convolution, instance normalization, and LeakyReLU activation function to enhance nonlinear learning capabilities. Finally, the hot prediction result is obtained through a 1×1×1 convolution prediction head.

[0022] Furthermore, the thermal prediction image obtained by S10 is converted to the frequency domain through two-dimensional discrete Fourier transform (DFT2D) to obtain the frequency domain function F(u,w). Thus, the frequency domain loss L is calculated. f , compare the predicted F(u,w) and the true value The amplitude and phase of are constrained by L1 norm. The specific formula is:

[0023]

[0024] β is a balancing factor hyperparameter, which is used to adjust the weights of amplitude and phase losses.

[0025] Furthermore, the mean square error (MSE) loss is used as the spatial domain loss. Weigh the predicted value T and the true value The difference in the spatial domain. The frequency domain loss L f and spatial domain loss L s Combined, the frequency-space mixing loss FSL is formed, and the formula is:

[0026] FSL=L s +αL f

[0027] α is a balancing hyperparameter used to balance the contribution of the two losses.

[0028] A frequency domain-spatial domain fusion perception system for fast and high-precision thermal prediction of 2.5D integrated circuits, the system comprising:

[0029] Parameter preprocessing module (PPNet), used for dimensionality reduction and feature fusion of input parameters;

[0030] Frequency-space hot encoding module, used to extract and fuse frequency domain and space domain heat diffusion features;

[0031] Frequency domain cross-scale interaction module for semantic alignment and fusion of multi-scale features;

[0032] Decoder module, used to generate multi-layer temperature field prediction results;

[0033] The loss calculation module is used to optimize the network parameters through the frequency-space hybrid loss function (FSL).

[0034] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the frequency domain-spatial domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits is implemented.

[0035] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the frequency domain-spatial domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits.

[0036] Beneficial effects: Aiming at the problem of fast and high-precision thermal prediction of 2.5D integrated circuits (ICs), the present invention proposes a frequency domain-spatial domain fusion perception method for the first time, which realizes:

[0037] 1. Efficient feature extraction: FSA-Heat innovatively integrates the high-to-low frequency and spatial domain encoder (FSTE) module with the frequency domain cross-scale interaction module (FCIFormer) to achieve heat dissipation feature extraction from high frequency to low frequency and from global to local. This multi-dimensional feature capture method can more comprehensively and meticulously depict the complex thermal distribution inside the 2.5D chip, especially for the more accurate prediction of high-frequency thermal gradients and hot spots.

[0038] 2. Targeted loss function design: The designed frequency-space mixed loss (FSL) function not only considers the mean square error in the spatial domain, but also supervises the amplitude and phase in the frequency domain. This dual supervision mechanism can effectively suppress high-frequency thermal gradient noise and reduce spatial structural misalignment, making the prediction results closer to the actual situation in both details and overall structure.

[0039] 3. Experimental verification of acceleration effect: Experimental results show that FSA-Heat has significant speed advantages over traditional commercial solvers based on the finite element method and the newly proposed graph network + PNA2.5D model. For example, compared with HotSpot, FSA-Heat's inference speed is 10.58 times faster and 4.23 times faster than GCN + PNA. Compared with the graph network + PNA method, MAE, RMSE, MAPE (reduced by 99%), PSNR (increased by 2.37 times). BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 : This is the architecture diagram of the frequency domain-spatial domain fusion perception method of the present invention, showing the connection relationship between the encoder (frequency-space hot encoder module), the decoder (residual convolution block) and the frequency domain cross-scale interaction module.

[0041] Figure 2 : This is an example of comparing the prediction results of the temperature field of multiple layers such as the heat source layer, the intermediate layer, and the heat dissipation layer of the present invention.

[0042] Figure 3 , 4 : It is a comparison diagram of the prediction errors of the present invention under the condition of no parameters (thermal conductivity, number of heat sources). DETAILED DESCRIPTION

[0043] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0044] Example: Figure 1 As shown, the overall process of the frequency domain-spatial domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits of the present invention is as follows Figure 1 The encoder part includes four encoder layers, each layer consists of L iIt consists of stacked frequency-space hot encoder blocks to extract high-to-low heat dissipation gradient features and spatial features. In addition, the traditional skip connection between the encoder and decoder parts is replaced by the proposed frequency-domain cross-scale interaction module, which implements the interaction of the input multi-scale encoder layers and the heat dissipation perception from global to local. The output of the frequency-domain cross-scale interaction module is connected with four decoder layers using residual convolution blocks. Finally, a scaled prediction head is used to generate 2.5D thermal predictions. In FSA-Heat training, it uses the designed frequency-space hybrid loss (FSL) for backpropagation. Figure 1 In the figure, D, H, W, and C represent the depth, height, width, and number of channels of the feature map, respectively. The present invention specifically includes the following steps:

[0045] S1: Use the HotSpot simulation tool to generate a training data set containing 6000 samples, covering the four layers of the 2.5D chip: heat source layer, thermal interface material layer, heat dissipation layer, and heat sink layer. The chip size of each layer is set to 18mm×18mm, the grid resolution is 64×64, the number of heat sources is randomly distributed between 4 and 35, and the power density follows a Gaussian distribution (mean 3W / mm 2 , variance 1.4 2 The input parameters include geometric parameters (material thickness, length, width, side length of heat sink), heat exchange parameters (thermal conductivity, convection coefficient) and heat source parameters (inter-layer power distribution and total power).

[0046] S2: The geometric parameters and heat exchange parameters are input into the 3D-Stem layer and reduced in dimension through a 3×3×1 convolution kernel (with a stride of (2×2×1) and a channel number of 32). Then, cross-parameter correlation features are extracted through two layers of 3D convolution blocks, each of which contains 3×3×3 convolution, instance normalization, and SILU activation functions, with the number of output channels being 16 and 32 respectively.

[0047] S3: Use trilinear interpolation to upsample the fused feature map to a resolution of 64×64×4×32, and concatenate it with the heat source parameters (64×64×4×2 grid power distribution) along the channel dimension to form the final input feature.

[0048] S4: The input features are encoded into 64×64×4×C dimensions through an embedding layer, and then the frequency domain feature dimensions are kept consistent with the input (64×64×4×C) through a 3D discrete cosine transform (3D-DCT). f Dynamically adjust the weight of each frequency band, where (E u ,E w ,E p) is initialized to 0.01 and updated with training. The weighted features are restored to the spatial domain using a 3D inverse discrete cosine transform (3D-IDCT) and the importance of the output frequency branch features is dynamically adjusted through the source input using a multiplication gated signal attention mechanism.

[0049] S5: Extract local spatial features using residual block structure: Use two layers of 3×3×3 convolution, instance normalization and SILU activation function, and the convolution layer uses reflection filling mode to maintain adiabatic boundaries. The frequency domain and spatial domain output features are concatenated by channel (total number of channels 2C), and compressed to C channels through 1×1×1 convolution (linear layer).

[0050] S6: Add learnable position encoding to the four-scale features output by the encoder (resolution 64×64×C, 32×32×C / / 2, 16×16×C / / 4, 8×8×C / / 8) and generate the query vector (q) through linear projection 1 ,q 2 ,q 3 ,q 4 ) and the key-value vector (k, v). Perform 3D-DCT transformation on the query vector and the key vector to the frequency domain respectively, calculate the cross-scale attention score, perform 3D-IDCT transformation on the attention score back to the spatial domain, and perform weighted summation with the value vector (v) to generate the fused feature. S7: Input the fused feature into the feedforward network (containing two fully connected layers, the hidden layer dimension is 4 times the input dimension, and the activation function is GELU), upsample to the corresponding scale of the decoder through trilinear interpolation, and concatenate with the output of the decoder residual block.

[0051] S8: The decoder contains 4 residual layers, each of which is upsampled to twice the resolution through trilinear interpolation, followed by 3×3×3 convolution (channel number 516→256), instance normalization, and GELU activation. The upsampled features are concatenated with the output of the frequency domain cross-scale interaction module and fused through 3×3×3 convolution. The final layer outputs the four-layer temperature field prediction results through 1×1×1 convolution (channel number 256→4) and scaling the output range.

[0052] In the specific implementation process, the loss function uses the frequency-space hybrid loss (FSL): the frequency domain loss calculates the Fourier transform amplitude and phase L1 norm ((β=1.0)) of the predicted and true values, and the spatial loss uses the mean square error (MSE). The total loss is (FSL=L s +0.5L f ). The Adam optimizer (initial learning rate 3e-4, weight decay 1e-5, batch size 8) was used to optimize the model parameters for 200 rounds of training by back-propagation algorithm, and the learning rate was reduced to 1e-4 in the 150th round. The model was trained on an Ubuntu server equipped with a GeForce RTX 3090 GPU to accelerate the computing process.

[0053] After the model training is completed, the new input parameters are input into the trained model to quickly generate thermal prediction results. The test data set is divided into three: one is a 1200 test set that is initially split from the training set, and the other two are 500 unseen test cases, which cover different thermal conductivity changes (±50%) and heat source number changes (maximum 2 times the number of heat sources in the training set) to verify the generalization ability of the model.

[0054] By comparing with existing methods (such as HotSpot and Graph Network + PNA), the performance of our model is evaluated in terms of root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), peak signal-to-noise ratio (PSNR), and inference speed. The experimental results show that our model significantly outperforms existing methods in all indicators, showing higher prediction accuracy and faster inference speed. Compared with the Graph Network + PNA method, MAE, RMSE, MAPE (reduced by 99%), and PSNR (increased by 2.37 times), which means that our method achieves higher prediction accuracy and better image details. In addition, our model achieves the fastest inference speed, which is 10.58 times faster than HotSpot and 4.23 times faster than Graph Network + PNA.

[0055] Figure 2 The comparison of the predicted results of the temperature and dissipation gradients of the heat source layer, TIM layer, heat diffusion layer, and heat sink layer is shown. The predicted images of FSA-Heat have higher structural fidelity in all four layers, and significantly smaller errors in multi-level (high and low temperature regions, especially in significant hot spots) and multi-scale (global and local) thermal regions. In addition, FSA-Heat also has higher prediction accuracy for the amplitude and direction of thermal gradients, especially in large amplitude regions. In contrast, GCN+PNA has higher prediction accuracy in the heat source layer and TIM layer, but the prediction accuracy in the heat diffusion layer and heat sink layer decreases rapidly, especially in the hot spot prediction.

[0056] Figure 3 and Figure 4 The comparison of thermal and gradient prediction between FSA-Heat and GCN+PNA under unseen conditions of thermal conductivity and number of heat sources is shown. The results show that FSA-Heat can achieve higher thermal and gradient prediction accuracy in all four layers, regardless of whether the thermal conductivity increases or decreases, and when the number of heat sources is 10, 60, and 80, which are unseen conditions. This shows that FSA-Heat has good adaptability and generalization performance when facing changes in thermal conductivity and the number of heat sources, and can accurately perform thermal predictions, providing a reliable tool for thermal analysis in the chip design stage.

[0057] It should be noted that the above embodiments are not intended to limit the protection scope of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions all fall within the protection scope of the claims of the present invention.

Claims

1. A frequency domain-spatial domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits, characterized in that: The following steps are involved: S1: Input geometric parameters, heat exchange parameters and heat source parameters, perform dimension reduction and fusion through parameter preprocessing network (PPNet) to generate joint input features; S2: The frequency-space heat encoder module performs dynamic weighting in the frequency domain and local feature extraction in the spatial domain on the joint input features to generate fused global-local heat diffusion features; S3: The frequency domain cross-scale interaction module is used to perform frequency domain cross-attention fusion on the multi-scale features output by the encoder to generate cross-scale semantic alignment features; S4: The decoder upsamples and residual maps the cross-scale semantic alignment features, combines the frequency-space mixed loss function (FSL), and outputs multi-layer temperature field and thermal gradient prediction results.

2. The frequency domain-spatial domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits according to claim 1, characterized in that: In step S2, the details are as follows: 2-1: Convert the input features to the frequency domain through 3D discrete cosine transform (3D-DCT) and use learnable weights Dynamically adjust the thermal gradient mode of each frequency band, where W f The three directions (u, w, p) in the Fourier domain and the high-frequency to low-frequency band (E) corresponding to the three heat conduction directions (x, y, z) in the spatial domain are u ,E w ,E p ) is embedded; 2-2: Perform 3D inverse discrete cosine transform (3D-IDCT) on the weighted frequency domain features to restore them to the spatial domain; 2-3: Use residual convolution blocks in parallel to extract local spatial features and maintain adiabatic boundary conditions through reflection filling; 2-4: After concatenating the frequency domain and spatial domain features, the feature weights are dynamically adjusted through the channel attention mechanism.

3. The frequency domain-spatial domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits according to claim 1, characterized in that: In step S3, the details are as follows: 3-1: Add learnable position encoding to the four-scale features output by the encoder to generate query vectors (q1, q2, q3, q4) and key-value vectors (k, v); 3-2: Convert the query and key vectors to the frequency domain through 3D-DCT and calculate the cross-scale attention score: S represents the attention score, q i and k represent multi-query and key matrices respectively, d k is the dimension of the value matrix k, 3-3: Perform 3D-IDCT on the attention score to transform it back to the spatial domain, and perform weighted summation with the value vector (v) to generate multi-scale fusion features; 3-4: The fused features are input into the feed-forward network (FFN) for nonlinear enhancement and concatenated with the decoder features through upsampling.

4. The frequency domain-spatial domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits according to claim 1, characterized in that: In step S4, the FSL loss function includes: Frequency domain loss f :Get the frequency domain function F(u,w) of the predicted transformation, and calculate the predicted F(u,w) and the true value in the Fourier domain The 2D Fourier transform amplitude and phase L1 norm of is: Space loss L s :Calculate the mean square error (MSE) between the prediction and the true value, and the total loss is FSL = L s +αL f , where the hyperparameters α=0.5, β=1.

0.

5. The frequency domain-spatial domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits according to claim 1, characterized in that: In step S1, the parameter preprocessing network (PPNet) specifically includes: 1-1: Convolutional dimension reduction of geometric parameters and heat exchange parameters is performed through the 3D-Stem layer; 1-2: Extract cross-parameter correlation features through 2 layers of 3D convolution blocks and activation functions; 1-3: Upsample the fused features to a resolution of 64×64×4 and concatenate them with the heat source parameters along the channel dimension as input features.

6. The frequency domain-spatial domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits according to claim 1, characterized in that: The decoder module contains 4 layers of residual blocks, each layer is upsampled by trilinear interpolation and concatenated with the corresponding scale features output by FCIFormer, and finally outputs four layers of temperature field prediction results through 1×1×1 convolution.

7. A frequency domain-spatial domain fusion perception system for fast and high-precision thermal prediction of 2.5D integrated circuits, characterized in that: A frequency domain-spatial domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits according to any one of claims 1 to 6, The system comprises: Parameter preprocessing module (PPNet), used for dimensionality reduction and feature fusion of input parameters; Frequency-space hot encoding module, used to extract and fuse frequency domain and space domain heat diffusion features; Frequency domain cross-scale interaction module for semantic alignment and fusion of multi-scale features; Decoder module, used to generate multi-layer temperature field prediction results; The loss calculation module is used to optimize the network parameters through the frequency-space hybrid loss function (FSL).

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the frequency domain-spatial domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits as described in any one of claims 1 to 6 above is implemented.

9. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by the processor, the frequency domain-spatial domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits as described in any one of claims 1 to 6 is implemented.

Images