A frequency-space domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits

Abstract

The present application relates to a kind of frequency domain-space domain fusion perception methods for 2.5D integrated circuit fast high-precision thermal prediction, the network includes frequency space thermal encoder module, frequency domain cross-scale interaction module, frequency domain-space domain hybrid loss function (FSL) and model input and pre-processing network (PPNet).Frequency space thermal encoder module combines adaptive multi-frequency embedding and 3D convolutional neural network, extracts high-to-low frequency thermal dissipation gradient feature and detailed spatial feature;Frequency domain cross-scale interaction module realizes multi-scale feature interaction through cross attention mechanism, solves semantic gap problem;Frequency domain-space domain hybrid loss function loss function combines spatial mean square error and frequency domain loss, suppresses high-frequency thermal gradient noise.The present application first realizes the thermal prediction of frequency domain-space domain dual-domain perception, significantly improves prediction accuracy and speed, and has excellent generalization ability, and can be widely applied to the thermal management field of 2.5D integrated circuit.

Description

TECHNICAL FIELD

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

[0002] In the post-Moore era, although 2.5D / 3D chip technology realizes higher density and lower power consumption design, it also brings the problem of increasing power density and hot spots, which seriously affects the system reliability and life. Thermal analysis is crucial for detecting hot spot areas on the chip, providing thermal-aware placement / layout solutions, and assisting in fast design turnaround. However, traditional thermal simulation methods such as finite element method (FEM) and finite difference method (FDM) have high accuracy but consume a lot of computing resources, making it difficult to meet the needs of repeated thermal simulation in complex task design optimization process.

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

[0004] In addition, although existing global self-attention mechanisms can perceive global information, they mainly focus on low-frequency features and still lack the extraction of high-frequency global information, which is crucial for improving the prediction accuracy of hot spots and heat dissipation gradients. Therefore, developing a thermal prediction model that can effectively capture multi-frequency global to local heat dissipation features is of great significance for solving the thermal management problem of 2.5D integrated circuits. SUMMARY

[0005] The present application aims to provide a frequency-space domain fusion perception method (FSA-Heat) for fast and high-precision thermal prediction of 2.5D integrated circuits to solve the problem that existing thermal prediction models cannot effectively capture global thermal features and high-frequency components, and to realize fast and high-precision thermal prediction.

[0006] To achieve the above-mentioned purpose, the present application provides a frequency-space 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 application first converts the geometric properties, heat exchange conditions and heat source parameters of the 2.5D integrated circuit into 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 convective resistance coefficient h r , thermal conductivity k; the heat source parameters include layered heat source power Q s , and total heat source power Qt.

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

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

[0010] S4: The frequency-space thermal encoder module adopts a deep convolution layer followed by two branches, which process the frequency domain and spatial domain features respectively. In the frequency domain branch, the spatial features are mapped to the frequency domain using three-dimensional discrete cosine transform (3D-DCT) to obtain three-dimensional anisotropic frequency features, and the learnable frequency weight is used to adaptively adjust the features of different frequency bands. The frequency features are converted back to the spatial domain through 3D inverse discrete cosine transform (3D-IDCT), and the frequency branch uses a multiplication gate signal attention mechanism to dynamically adjust the importance of the features. The formula of the learnable weight is:

[0011]

[0012] Due to the anisotropy of heat conduction, three directions in the Fourier domain , and high to low frequency bands corresponding to the three heat conduction directions (x, y, z) in the spatial domain are embedded.

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

[0014] S6: The features of the frequency domain and spatial domain branches are fused 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 and alleviate the semantic gap. First, the embedding vectors output by the multi-scale frequency-space thermal encoder module are added to the position embedding to obtain query vectors and key ,value .

[0016] S8: Transform the query vector and key to the frequency domain using 3D-DCT to obtain their frequency domain coordinate representation. Calculate the attention score S between vectors at different scales, representing semantic similarity. The specific formula is as follows:

[0017]

[0018] S represents the attention score. and These represent multi-head queries and key matrices, respectively. Value matrix Dimensions.

[0019] S9: Feature acquisition using 3D-IDCT and weighted summation Then, a feedforward network (FFN) is used to enrich the semantic content and enhance the feature representation capability. The processed features... Upsampling is performed to align the decoder and ensure feature size matching.

[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 input into the residual block of the next decoding layer. The residual block contains two layers of 3×3×3 convolutions, instance normalization, and LeakyReLU activation function to enhance nonlinear learning ability. Finally, the hot prediction result is obtained through a 1×1×1 convolution prediction head.

[0022] Furthermore, the thermal prediction image obtained from S10 is transformed into the frequency domain using a two-dimensional discrete Fourier transform (DFT2D) to obtain the frequency domain function. Therefore, the frequency domain loss is calculated. Comparison and prediction and the true value The amplitude and phase are constrained using the L1 norm, and the specific formula is as follows:

[0023]

[0024] This is a balancing factor hyperparameter used to adjust the weights of amplitude and phase loss.

[0025] Furthermore, mean squared error (MSE) loss is used as the spatial domain loss. Measure the predicted value Compared with the true value Difference in spatial domain. The frequency domain loss and the spatial domain loss are combined to form a frequency-spatial hybrid loss FSL, whose formula is:

[0026]

[0027] is used to balance the contributions of the two losses.

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

[0029] a parameter preprocessing module (PPNet) for dimensionality reduction and feature fusion of input parameters;

[0030] a frequency-space thermal coding module for extracting and fusing frequency domain and spatial domain thermal diffusion features;

[0031] a frequency domain cross-scale interaction module for multi-scale feature semantic alignment and fusion;

[0032] a decoder module for generating multi-layer temperature field prediction results;

[0033] a loss calculation module for optimizing network parameters through a frequency-spatial hybrid loss function (FSL).

[0034] An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the frequency-spatial fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits when executing the program.

[0035] A computer-readable storage medium having computer instructions stored thereon, the computer instructions being executed by a processor to implement the frequency-spatial fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits.

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

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

[0038] 2. Targeted Loss Function Design: The designed Frequency-Spatial Hybrid 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 effectively suppresses high-frequency thermal gradient noise and reduces spatial structural misalignment, thus making the prediction results closer to the real situation in both detail and overall structure.

[0039] 3. Experimental Verification of Acceleration Effect: Experimental results show that FSA-Heat exhibits significant speed advantages compared to traditional commercial solvers based on the finite element method and the newly proposed graph network + PNA 2.5D model. For example, compared to HotSpot, FSA-Heat's inference speed is 10.58 times faster, and compared to GCN + PNA, it is 4.23 times faster. Compared to the graph network + PNA method, MAE, RMSE, and MAPE are reduced by 99%, while PSNR is improved by 2.37 times. Attached Figure Description

[0040] Figure 1 The diagram below shows the architecture of the frequency-spatial domain fusion sensing method of the present invention, illustrating the connection relationships between the encoder (frequency-spatial thermal encoder module), the decoder (residual convolutional block), and the frequency domain cross-scale interaction module.

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

[0042] Figure 3 Figure 1 shows a comparison of the predicted errors of thermal conductivity under the condition that the parameters are not available.

[0043] Figure 4 The image shows a comparison of the prediction errors of the number of heat sources under the condition that the parameters are not known. Detailed Implementation

[0044] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0045] Example: Figure 1 As shown in Figure 1, the overall flow of the frequency-spatial domain fusion sensing method for fast and high-precision thermal prediction of 2.5D integrated circuits according to the present invention is illustrated. Its encoder section includes four encoder layers, each consisting of… The stacked spatial-thermal encoder block is composed to extract the thermal dissipation gradient features and spatial features from high to low. In addition, the traditional skip connection between the encoder and decoder part is replaced by the proposed frequency domain cross-scale interaction module, which realizes the interaction of input multi-scale encoder layers and the perception of thermal dissipation 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, the 2.5D thermal prediction is generated using the scaled prediction head. In the FSA-Heat training, it uses the designed frequency-space hybrid loss (FSL) for back propagation. In Fig. 1, 、 、 、 respectively represent the depth, height, width, and channel number of the feature map. The present application specifically includes the following steps:

[0046] S1: Use the HotSpot simulation tool to generate a training data set containing 6000 samples, covering the four-layer structure 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 18 mm x 18 mm, and the grid resolution is 64 x 64. The number of heat sources is randomly distributed between 4 and 35, and the power density follows a Gaussian distribution (mean 3 W / mm², variance 1.4²). The input parameters include geometric parameters (material thickness, length, width, heat sink side length), heat exchange parameters (thermal conductivity, convection coefficient), and heat source parameters (interlayer power distribution and total power).

[0047] S2: Input the geometric parameters and heat exchange parameters into the 3D-Stem layer, and reduce the dimension through a 3x3x1 convolution kernel (stride (2x2x1), channel number increased to 32). Then extract the cross-parameter correlation features through 2 layers of 3D convolution blocks, each containing a 3x3x3 convolution, instance normalization, and a SILU activation function, with the output channel number being 16, 32 in turn.

[0048] S3: Use tri-linear interpolation to upsample the fused feature map to 64x64x4x32 resolution, and concatenate it with the heat source parameters (64x64x4x2 grid power distribution) along the channel dimension to form the final input feature.

[0049] S4: Perform embedding layer encoding on the input feature to 64x64x4x dimension, and then pass it through a 3D discrete cosine transform (3D-DCT), and the frequency domain feature dimension remains consistent with the input (64x64x4x ). The weight is dynamically adjusted by the learnable weight initialized to 0.01 and updated with training. Perform 3D inverse discrete cosine transform (3D-IDCT) on the weighted features, restore to spatial domain and utilize the multiplication gating signal attention mechanism to dynamically adjust the output frequency branch feature importance through the source input.

[0050] S5: Extract local spatial features using residual block structure: use two layers of 3x3x3 convolution, instance normalization and SILU activation function, and the convolution layer uses reflective padding mode to maintain adiabatic boundary. The frequency domain and spatial domain output features are spliced by channel (total channel number ), compressed to channels through 1x1x1 convolution (linear layer).

[0051] S6: Add learnable position encoding to the four-scale features (resolution 64x64xC, 32x32xC / / 2, 16x16xC / / 4, 8x8xC / / 8) output by the encoder, generate query vector and key-value vector through linear projection. Perform 3D-DCT conversion to the frequency domain on the query vector and the key vector respectively, calculate the cross-scale attention score, perform 3D-IDCT conversion back to the spatial domain on the attention score, and weighted sum with the value vector to generate the fused feature. S7: Input the fused feature into the feedforward network (contains 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 splice with the decoder residual block output.

[0052] S8: The decoder contains 4 residual layers, each layer is upsampled by trilinear interpolation to double the resolution, followed by 3x3x3 convolution (channel number 516->256), instance normalization and GELU activation. The upsampled feature is spliced with the output of the frequency domain cross-scale interaction module, and fused through 3x3x3 convolution. The final layer is output by 1x1x1 convolution (channel number 256->4) and scaled by the output range, and the four-layer temperature field prediction results are output.

[0053] In the specific implementation process, the loss function uses frequency-space hybrid loss (FSL): the frequency domain loss calculates the Fourier transform amplitude and phase L1 norm between the prediction and the true value, and the spatial loss uses mean square error (MSE), and the total loss is . Use Adam optimizer (initial learning rate 3e-4, weight decay 1e-5, batch size 8) to optimize model parameters through back propagation algorithm for 200 rounds, and the learning rate is reduced to 1e-4 at the 150th round. The model is trained on an Ubuntu server equipped with a GeForce RTX 3090 GPU to speed up the calculation process.

[0054] After the model training is completed, 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 the initial 1200 test set divided with the training set, and the other two are 500 unseen test cases, which cover different thermal conductivity changes (± 50%) and heat source quantity changes (maximum 2 times the number of heat sources in the training set) respectively, to verify the generalization ability of the model.

[0055] By comparing with existing methods (such as HotSpot and Graph Network + PNA), the performance of our model 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 is evaluated. Experimental results show that our model is significantly better than 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%), 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, 10.58 times faster than HotSpot and 4.23 times faster than Graph Network + PNA.

[0056] Figure 2 The temperature and its dissipation gradient prediction results of the heat source layer, TIM layer, heat diffusion layer and heat sink layer are compared. FSA-Heat has higher structural fidelity in all four layers of predicted images, with 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 magnitude and direction of thermal gradient, especially in large amplitude areas. In contrast, GCN+PNA has higher prediction accuracy in the heat source layer and TIM layer, but its prediction accuracy decreases rapidly in the heat diffusion layer and heat sink layer, especially in hot spot prediction.

[0057] Figure 3 and Figure 4 FSA-Heat and GCN+PNA thermal and gradient prediction comparisons are shown in the case of unseen thermal conductivity and heat source quantity. The results show that FSA-Heat can achieve higher thermal and gradient prediction accuracy in all four layers, regardless of whether the thermal conductivity is increased or decreased, and the number of heat sources is 10, 60 and 80, etc. unseen. This indicates that FSA-Heat has good adaptability and generalization performance when facing changes in thermal conductivity and heat source quantity, and can accurately perform thermal prediction, providing a reliable tool for thermal analysis in the chip design stage.

[0058] It should be noted that the above examples are not intended to limit the scope of the present application, and equivalent transformations or substitutions made on the basis of the above technical solutions all fall within the scope of the claims of the present application.

Claims

1. A frequency-space domain fusion perception method for 2.5D integrated circuit fast high-precision thermal prediction, characterized in that, The method comprises the following steps: S1: input geometric parameters, heat exchange parameters and heat source parameters, and perform dimension reduction and fusion through a parameter preprocessing network PPNet to generate joint input features; S2: perform frequency domain dynamic weighting and spatial domain local feature extraction on the joint input features through a frequency-space thermal encoder module to generate fused global-local thermal diffusion features; S3: perform frequency domain cross-attention fusion on the multi-scale features output by the encoder through a frequency domain cross-scale interaction module to generate cross-scale semantic alignment features; S4: the decoder performs upsampling and residual mapping on the cross-scale semantic alignment features, combines a frequency-space hybrid loss function FSL, and outputs multi-layer temperature field and thermal gradient prediction results; In step S2, the following steps are performed, 2-1: convert the input features to the frequency domain by 3D-Discrete Cosine Transform 3D-DCT, and pass through learnable weights dynamically adjust the thermal gradient pattern for each frequency band, wherein is embedded by three directions under the Fourier domain ( ), and the high-to-low frequency bands corresponding to the three heat conduction directions (x, y, z) in the spatial domain ​ 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: Extract local spatial features in parallel using residual convolution blocks, and maintain the adiabatic boundary condition through reflection padding; 2-4: After concatenating the frequency domain and spatial domain features, dynamically adjust the feature weights through a channel attention mechanism; In step S3, the following steps are performed, 3-1: Add learnable positional encoding to the four-scale features output by the encoder to generate query vectors and key value vectors ; 3-2: Convert the query and key vectors to the frequency domain through 3D-DCT, and calculate the cross-scale attention score: S denotes an attention score, and denote multi-headed query and key matrices, respectively, is a value matrix of dimensionality 3-3: Perform 3D-IDCT transform back to spatial domain on the attention score, with value vector weighted sum, to generate multi-scale fusion features; 3-4: Input the fused features into a feedforward network FFN for nonlinear enhancement, and perform upsampling and decoder feature concatenation.

2. The frequency-domain and 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 comprises: Frequency domain loss : obtaining a predicted transformed frequency domain function , computing the 2D Fourier transform magnitude and phase L1 norm of the predicted and true values , formula: spatial loss : compute the mean squared error (MSE) of the prediction vs. the true value, total loss is where the hyperparameters , .

3. The frequency-space 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 comprises: 1-1: Perform convolution dimension reduction on the geometric parameters and heat exchange parameters through a 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 64x64x4, and concatenate them with the heat source parameters along the channel dimension as input features.

4. The frequency-domain and 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 includes 4 residual blocks, each of which is upsampled through trilinear interpolation, and the corresponding scale features output by the FCIFormer are concatenated, and finally a 1x1x1 convolution is performed to output four-layer temperature field prediction results.

5. A frequency-space domain fusion perception system for 2.5D integrated circuit fast high-precision thermal prediction, characterized in that, The frequency-space domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits, The system comprises: a parameter preprocessing module PPNet for dimension reduction and feature fusion of input parameters; a frequency-space thermal coding module for extracting and fusing frequency domain and spatial domain thermal diffusion features; a frequency domain cross-scale interaction module for multi-scale feature semantic alignment and fusion; a decoder module for generating multi-layer temperature field prediction results; a loss calculation module for optimizing network parameters through a frequency-space hybrid loss function FSL.

6. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that: The processor executes the program to implement the frequency-space domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits as claimed in any one of claims 1 to 4.

7. A computer readable storage medium having stored thereon computer instructions, characterized in that: The computer instructions are executed by the processor to implement the frequency-space domain fusion perception method for fast and high-precision thermal prediction of 2.5D integrated circuits as claimed in any one of claims 1 to 4.

Images