Chip-to-chip data transmission efficiency improvement method and system for 3D packaging structure

Abstract

This invention provides a method and system for improving inter-chip data transmission efficiency in 3D packaging structures, relating to the field of 3D packaging technology. The method includes obtaining the spatial distribution matrix of a through-silicon via (TSV) array, calculating spatial distance and inter-layer hop count to cluster the TSVs. Data to be transmitted is fragmented according to the number of clusters and channel bandwidth, and assigned to corresponding clusters, establishing a bidirectional state mapping. Based on the mapping, the capacity saturation time of each cluster is predicted, and data fragments are migrated to unsaturated clusters before saturation. Feature parameters are generated by correlating the capacity saturation time with the spatial distance to adjust the granularity of subsequent data fragmentation, creating a time mismatch between the fragment size distribution and the saturation period of each cluster, thereby optimizing the transmission path load and improving overall data transmission efficiency.

Description

Technical Field

[0001] This invention relates to 3D packaging technology, and more particularly to a method and system for improving the efficiency of inter-chip data transmission in 3D packaging structures. Background Technology

[0002] In the field of 3D integrated circuit packaging, through-silicon vias (TSVs) are a key structure for realizing vertical interconnection and data transmission between chips. With the continuous increase in integration density and performance requirements, how to efficiently utilize TSV arrays for large-scale data communication has become a core technological challenge. In existing technologies, conventional approaches mainly rely on static resource allocation and fixed routing strategies for TSV arrays. Specifically, the system typically binds data streams statically to specific TSVs or channel sets for transmission based on the physical location of the TSVs or pre-defined logical groups. Data transmission scheduling is often based on simple polling or priority queue mechanisms, lacking dynamic perception and adaptation to the real-time status of the transmission channels and the 3D spatial topology.

[0003] These conventional practices have significant drawbacks. Due to static bonding and fixed routing, data transmission paths cannot be dynamically adjusted based on the real-time load of the through-silicon via (TSV) array. When a TSV or a group of TSVs reaches capacity saturation due to data congestion, subsequent data allocated to that channel will face queuing delays or even transmission failures, while other idle or low-load TSV resources remain underutilized, leading to decreased overall transmission efficiency and resource idleness. Furthermore, existing methods typically ignore the spatial distribution characteristics of TSVs in 3D packaging, such as the impact of interlayer hop count and physical distance differences on transmission delay and reliability. Fixed data partitioning granularity cannot adapt to the varying saturation periods of different TSV clusters due to spatial topology, easily causing multiple transmission clusters to reach capacity bottlenecks simultaneously, resulting in systemic transmission fluctuations and performance degradation. Therefore, a method capable of dynamically sensing the state of the 3D interconnect structure and intelligently scheduling data transmission is urgently needed to improve resource utilization and overall transmission efficiency. Summary of the Invention

[0004] This invention provides a method and system for improving the efficiency of inter-chip data transmission in 3D packaging structures, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a method for improving inter-chip data transmission efficiency in 3D packaging structures, comprising:

[0006] Obtain the spatial distribution matrix of the data to be transmitted and the through-silicon via array in the three-dimensional packaging structure. Calculate the spatial distance between through-silicon vias and the interlayer transmission hop count based on the spatial distribution matrix. Then, cluster the through-silicon vias according to the spatial distance and the interlayer transmission hop count to generate multiple through-silicon via transmission clusters.

[0007] The data to be transmitted is divided into fragments according to the number of the through-silicon via (TSV) transmission clusters and the channel bandwidth capacity of the TSVs in each cluster, generating a data fragment set corresponding to the TSV transmission clusters. Each data fragment is labeled with its corresponding TSV transmission cluster identifier. Each data fragment is assigned to the corresponding TSV transmission cluster according to the TSV transmission cluster identifier. During the transmission process, a bidirectional state mapping relationship between the data fragments and the TSV transmission clusters is established.

[0008] Based on the bidirectional state mapping relationship, the capacity saturation time of each through-silicon via (TSV) transmission cluster is predicted, and before the capacity saturation time is reached, the data to be transmitted is fragmented and migrated across clusters to unsaturated TSV transmission clusters.

[0009] The capacity saturation time of each through-silicon via (TSV) transmission cluster is correlated with the spatial distance to generate a capacity-topology characteristic parameter for the TSV transmission cluster. The capacity-topology characteristic parameter is used to adjust the granularity of subsequent data fragmentation to be transmitted, so that the size distribution of the data fragments is mismatched with the capacity saturation period of each TSV transmission cluster.

[0010] Based on the spatial distribution matrix, the spatial distance between through-silicon vias (TSVs) and the interlayer transport hop count are calculated. The TSVs are then clustered according to the spatial distance and the interlayer transport hop count to generate multiple TSV transport clusters, including:

[0011] The three-dimensional coordinates and interlayer interconnect topology of each through-silicon via (TSV) are extracted from the spatial distribution matrix. The spatial distance between TSVs is calculated based on the three-dimensional coordinates, and the interlayer transport hop count between TSVs is traced based on the interlayer interconnect topology.

[0012] Construct a through-silicon via (TSV) transmission cost function, and nonlinearly couple the spatial distance and the interlayer transmission hop count according to the transmission cost function, so that the influence weight of the interlayer transmission hop count on the transmission cost decreases nonlinearly as the spatial distance increases, and use the transmission cost function to calculate the transmission cost between any two TSVs.

[0013] Based on the transmission cost, a through-silicon via (TSV) correlation map is constructed. Connected components are identified in the TSV correlation map. TSVs with strong correlation edges are divided into the same connected component, and each connected component is used as an initial TSV transmission cluster. At the same time, the transmission cost dispersion of each initial TSV transmission cluster is calculated.

[0014] When the transmission cost dispersion of an initial through-silicon via (TSV) transport cluster exceeds the cohesion threshold, local peak points of the transmission cost value are identified within the initial TSV transport cluster. The initial TSV transport cluster is then re-clustered using the local peak points as the splitting cores to generate multiple TSV transport clusters with transmission cost cohesion.

[0015] Constructing a through-silicon via (TSV) transmission cost function, and nonlinearly coupling the spatial distance and the interlayer transmission hop count based on the transmission cost function, so that the weight of the interlayer transmission hop count on the transmission cost decreases nonlinearly with the increase of spatial distance, including:

[0016] Construct a through-silicon via (TSV) transport cost function, which includes a spatial distance term and an interlayer transport hops term;

[0017] An attenuation operator is constructed using spatial distance as the independent variable. The attenuation operator outputs an attenuation coefficient that decreases non-linearly as the spatial distance increases. The attenuation coefficient is then multiplied by the inter-layer transmission hop count to generate an attenuated inter-layer transmission hop count component.

[0018] The spatial distance and the interlayer transmission hop count components are weighted and summed to generate the transmission cost value between silicon via pairs. The transmission cost value between each silicon via pair is calculated according to the transmission cost function, thereby realizing the nonlinear coupling between the spatial distance and the interlayer transmission hop count, so that the influence weight of the interlayer transmission hop count on the transmission cost decreases nonlinearly as the spatial distance increases.

[0019] The data to be transmitted is divided into fragments according to the number of the through-silicon via (TSV) transmission clusters and the channel bandwidth capacity of the TSVs within each cluster, generating a data fragment set corresponding to each TSV transmission cluster. Each data fragment is labeled with its corresponding TSV transmission cluster identifier, including:

[0020] Obtain the number of through-silicon via (TSV) transmission clusters, traverse all TSVs within each TSV transmission cluster and read the channel bandwidth capacity parameters of each TSV, statistically summarize the channel bandwidth capacity of each TSV within the same TSV transmission cluster, and generate the cluster-level bandwidth capacity of each TSV transmission cluster.

[0021] Establish a mapping relationship between the cluster-level bandwidth capacity and the data allocation weight. The mapping relationship determines the data carrying weight of each silicon via transmission cluster based on the relative size of the cluster-level bandwidth capacity of each cluster-level bandwidth capacity in all cluster-level bandwidth capacities, and generates the bandwidth capacity ratio of each silicon via transmission cluster.

[0022] The total amount of data to be transmitted is obtained. The total amount of data is decomposed according to the bandwidth capacity ratio of each through-silicon via (TSV) transmission cluster. The total amount of data is then divided into data segments that match the bandwidth capacity ratio of each TSV transmission cluster, and the data segment size that each TSV transmission cluster should carry is generated.

[0023] The data to be transmitted is continuously segmented according to the data segment size. Starting from the beginning of the data to be transmitted, data segments are sequentially extracted according to the data segment size to generate data segments equal to the number of through-silicon via (TSV) transmission clusters. Each data segment is appended with its corresponding TSV transmission cluster identifier to generate a set of data segments with TSV transmission cluster identifiers.

[0024] Based on the through-silicon via (TSV) transport cluster identifier, each data fragment is assigned to the corresponding TSV transport cluster, and a bidirectional state mapping relationship is established between the data fragment and the TSV transport cluster during transmission, including:

[0025] Parse the through-silicon via (TSV) transport cluster identifier carried by each data fragment, and route each data fragment to its corresponding TSV transport cluster according to the TSV transport cluster identifier;

[0026] When a data shard enters a through-silicon via (TSV) transport cluster, a transport status awareness layer is established. The data shard is used as an observation object based on the transport status awareness layer, and the transport behavior characteristics of the data shard within the TSV transport cluster are captured. The transport behavior characteristics are then encoded as the transport behavior fingerprint of the data shard.

[0027] A cluster-level resource response layer is established synchronously. Based on the cluster-level resource response layer, the through-silicon via (TSV) transmission cluster is taken as the observation object, and the resource response characteristics of the TSV transmission cluster to the data fragmentation are monitored. The resource response characteristics are encoded as the resource response fingerprint of the TSV transmission cluster.

[0028] Cross-correlation analysis is performed on the transmission behavior fingerprint and the resource response fingerprint to establish a bidirectional state mapping relationship between data fragmentation and through-silicon via transmission clusters.

[0029] Based on the bidirectional state mapping relationship, the capacity saturation time of each through-silicon via (TSV) transmission cluster is predicted, and the data to be transmitted is fragmented and migrated across clusters to unsaturated TSV transmission clusters before the capacity saturation time is reached. This includes:

[0030] Extract the resource response fingerprint of each through-silicon via (TSV) transmission cluster and the transmission behavior fingerprint of each data segment from the bidirectional state mapping relationship.

[0031] Capacity evolution trajectory is constructed based on the load fluctuation pattern of each through-silicon via (TSV) transmission cluster. The capacity evolution trajectory depicts the nonlinear decay process of the available capacity of the TSV transmission cluster over time. The time node when the available capacity drops to the capacity threshold is marked on the capacity evolution trajectory, and the time node is taken as the capacity saturation moment.

[0032] Identify the through-silicon via transport clusters whose capacity saturation time falls within the warning time window, and select data fragments whose interlayer hopping frequency exceeds a preset hopping threshold from the through-silicon via transport clusters as migration candidate fragments;

[0033] Traverse the through-silicon via (TSV) transmission clusters whose capacity saturation time is outside the warning time window, and perform a compatibility evaluation based on the transmission behavior fingerprint of the migration candidate fragment and the resource response fingerprint of each TSV transmission cluster. Select the TSV transmission cluster with the best compatibility evaluation value as the receiving cluster.

[0034] Before the capacity saturation time is reached, the migration candidate fragment is migrated from the original through-silicon via (TSV) transmission cluster to the receiving cluster, and the association record between the migration candidate fragment and the receiving cluster is updated in the bidirectional state mapping relationship, thus completing the cross-cluster migration of the data fragment to be transmitted to the unsaturated TSV transmission cluster.

[0035] Adjusting the fragmentation granularity of subsequent data to be transmitted using the capacity-topology characteristic parameters, so that the size distribution of data fragments and the capacity saturation period of each through-silicon via (TSV) transmission cluster are mismatched in time, includes:

[0036] The capacity saturation period and topological connectivity of each through-silicon via transmission cluster are extracted from the capacity-topology characteristic parameters. The capacity saturation period is projected onto the cyclic time axis to construct a capacity saturation rhythm diagram. Peak superposition regions and trough isolation regions are identified in the capacity saturation rhythm diagram.

[0037] A piecewise granularity inverse control function is constructed. The piecewise granularity inverse control function uses the superposition intensity of the peak superposition region as input to generate a granularity suppression coefficient, uses the isolation spacing of the valley isolation region as input to generate a granularity release coefficient, and introduces the topological connectivity as a nonlinear modulation parameter into the calculation of the granularity suppression coefficient and the granularity release coefficient.

[0038] The fragmentation granularity inverse control function is used to perform variable granularity segmentation on the subsequent data to be transmitted, generating fine-grained data fragments at the silicon via transmission clusters corresponding to the peak superposition region and coarse-grained data fragments at the silicon via transmission clusters corresponding to the valley isolation region.

[0039] The fine-grained data fragments and the coarse-grained data fragments are allocated to each through-silicon via (TSV) transmission cluster, so that the size distribution of the data fragments and the capacity saturation period of each TSV transmission cluster are mismatched in time.

[0040] A second aspect of the present invention provides a system for improving the efficiency of inter-chip data transmission in 3D packaging structures, comprising:

[0041] The clustering calculation unit is used to obtain the spatial distribution matrix of the data to be transmitted and the through-silicon via array in the three-dimensional packaging structure, calculate the spatial distance between through-silicon vias and the interlayer transmission hops based on the spatial distribution matrix, and cluster the through-silicon vias according to the spatial distance and the interlayer transmission hops to generate multiple through-silicon via transmission clusters.

[0042] The data fragmentation unit is used to divide the data to be transmitted into fragments according to the number of the through-silicon via (TSV) transmission clusters and the channel bandwidth capacity of the TSVs in each cluster, generate a data fragment set corresponding to the TSV transmission clusters, and label each data fragment with its corresponding TSV transmission cluster identifier. Based on the TSV transmission cluster identifier, each data fragment is allocated to the corresponding TSV transmission cluster, and a bidirectional state mapping relationship between the data fragments and the TSV transmission clusters is established during the transmission process.

[0043] The cross-cluster migration unit is used to predict the capacity saturation time of each through-silicon via (TSV) transmission cluster based on the bidirectional state mapping relationship, and to migrate the data to be transmitted to the unsaturated TSV transmission clusters before the capacity saturation time is reached.

[0044] The feature adjustment unit is used to correlate the capacity saturation time of each through-silicon via (TSV) transmission cluster with the spatial distance, generate the capacity-topology feature parameters of the TSV transmission cluster, and use the capacity-topology feature parameters to adjust the fragmentation granularity of subsequent data to be transmitted, so that the size distribution of the data fragments forms a time mismatch with the capacity saturation period of each TSV transmission cluster.

[0045] A third aspect of the present invention provides an electronic device, comprising:

[0046] processor;

[0047] Memory used to store processor-executable instructions;

[0048] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0049] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0050] The beneficial effects of this application are as follows:

[0051] This method significantly improves the data transmission efficiency between chips within a 3D packaging structure. By obtaining the spatial distribution matrix of the through-silicon via (TSV) array and calculating the spatial distance and interlayer transmission hop count, comprehensive clustering of TSVs based on physical topology and logical paths is achieved. This clustering method aggregates spatially adjacent TSVs with low interlayer hop counts, effectively reducing signal transmission delay and crosstalk within the same transmission cluster, and establishing an optimized physical channel foundation for parallel data transmission.

[0052] Data fragmentation is performed based on the number of through-silicon via (TSV) transmission clusters and the channel bandwidth capacity of each cluster, ensuring that the data load matches the capacity of the transmission channels. Cluster identifiers are assigned to each data fragment, and a bidirectional state mapping relationship is established, enabling precise binding of data streams to physical transmission paths and real-time status monitoring. This refined mapping management allows the system to clearly understand the status of each data fragment within a specific transmission cluster, providing accurate information for dynamic scheduling.

[0053] Based on the bidirectional state mapping relationship, the capacity saturation time of each transmission cluster is predicted, and cross-cluster migration of data fragments is initiated before saturation, proactively avoiding channel congestion. This proactive scheduling mechanism diverts traffic from paths that are about to be overloaded to idle or less loaded paths in advance, maintaining the overall load balance of the transmission network and eliminating the problems of overall transmission delay and throughput reduction caused by local congestion. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating a method for improving inter-chip data transmission efficiency in 3D packaging structures according to an embodiment of the present invention.

[0055] Figure 2 This is a flowchart illustrating the construction process of a through-silicon via (TSV) transport cluster based on transmission cost, according to an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0058] Figure 1 This is a flowchart illustrating a method for improving inter-chip data transmission efficiency in 3D packaging structures according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0059] Obtain the spatial distribution matrix of the data to be transmitted and the through-silicon via array in the three-dimensional packaging structure. Calculate the spatial distance between through-silicon vias and the interlayer transmission hop count based on the spatial distribution matrix. Then, cluster the through-silicon vias according to the spatial distance and the interlayer transmission hop count to generate multiple through-silicon via transmission clusters.

[0060] The data to be transmitted is divided into fragments according to the number of the through-silicon via (TSV) transmission clusters and the channel bandwidth capacity of the TSVs in each cluster, generating a data fragment set corresponding to the TSV transmission clusters. Each data fragment is labeled with its corresponding TSV transmission cluster identifier. Each data fragment is assigned to the corresponding TSV transmission cluster according to the TSV transmission cluster identifier. During the transmission process, a bidirectional state mapping relationship between the data fragments and the TSV transmission clusters is established.

[0061] Based on the bidirectional state mapping relationship, the capacity saturation time of each through-silicon via (TSV) transmission cluster is predicted, and before the capacity saturation time is reached, the data to be transmitted is fragmented and migrated across clusters to unsaturated TSV transmission clusters.

[0062] The capacity saturation time of each through-silicon via (TSV) transmission cluster is correlated with the spatial distance to generate a capacity-topology characteristic parameter for the TSV transmission cluster. The capacity-topology characteristic parameter is used to adjust the granularity of subsequent data fragmentation to be transmitted, so that the size distribution of the data fragments is mismatched with the capacity saturation period of each TSV transmission cluster.

[0063] In one optional implementation, the spatial distance between vias and the interlayer transport hop count are calculated based on the spatial distribution matrix, and the vias are clustered according to the spatial distance and the interlayer transport hop count to generate multiple via transport clusters, including:

[0064] The three-dimensional coordinates and interlayer interconnect topology of each through-silicon via (TSV) are extracted from the spatial distribution matrix. The spatial distance between TSVs is calculated based on the three-dimensional coordinates, and the interlayer transport hop count between TSVs is traced based on the interlayer interconnect topology.

[0065] Construct a through-silicon via (TSV) transmission cost function, and nonlinearly couple the spatial distance and the interlayer transmission hop count according to the transmission cost function, so that the influence weight of the interlayer transmission hop count on the transmission cost decreases nonlinearly as the spatial distance increases, and use the transmission cost function to calculate the transmission cost between any two TSVs.

[0066] Based on the transmission cost, a through-silicon via (TSV) correlation map is constructed. Connected components are identified in the TSV correlation map. TSVs with strong correlation edges are divided into the same connected component, and each connected component is used as an initial TSV transmission cluster. At the same time, the transmission cost dispersion of each initial TSV transmission cluster is calculated.

[0067] When the transmission cost dispersion of an initial through-silicon via (TSV) transport cluster exceeds the cohesion threshold, local peak points of the transmission cost value are identified within the initial TSV transport cluster. The initial TSV transport cluster is then re-clustered using the local peak points as the splitting cores to generate multiple TSV transport clusters with transmission cost cohesion.

[0068] When extracting the 3D coordinates and interlayer interconnect topology of each through-silicon via (TSV) from the spatial distribution matrix, the spatial distribution matrix stores the TSV location information in a row-column format. Each row corresponds to one TSV, including its x, y, and z coordinates in 3D space, as well as the layer number of the chip layer to which it belongs. The 3D coordinates use a Cartesian coordinate system with the lower left corner of the chip packaging substrate as the origin, with the coordinate unit being micrometers and the coordinate values ​​having one decimal place precision. The interlayer interconnect topology is stored in the form of an adjacency list. Each TSV in the adjacency list corresponds to a linked list. The linked list nodes record the identifiers and connection types of other TSVs that have a direct electrical connection with that TSV. The connection types include intra-layer lateral connections and inter-layer vertical connections.

[0069] When calculating the spatial distance between through-silicon vias (TSVs) based on 3D coordinates, for any two TSVs i and j, their 3D coordinates are extracted as coordinates i and j, respectively. The differences between coordinates i and j in the x-axis, y-axis, and z-axis directions are calculated. The squares of the three differences are then summed, and the square root of the sum is taken to obtain the Euclidean spatial distance between TSVs i and j. When tracing the interlayer transmission hop count between through-silicon vias (TSVs) based on the interlayer interconnect topology, a breadth-first search is used to traverse the adjacency list. Starting from TSV i, a queue to be visited is initialized. TSV i is added to the queue and its hop count is marked as 0. The TSV at the head of the queue is taken as the current TSV. All adjacent TSVs of the current TSV in the adjacency list are traversed. If an adjacent TSV has not been visited, it is added to the queue and its hop count is marked as the hop count of the current TSV plus 1. This process is repeated until the queue is empty or TSV j is found. The hop count marked for TSV j is the interlayer transmission hop count between TSV i and TSV j.

[0070] When constructing the transmission cost function for through-silicon vias (TSVs), the function takes spatial distance and interlayer transmission hops as input parameters and outputs the transmission cost value as a comprehensive indicator of the data transmission cost between two TSVs. The transmission cost function non-linearly couples spatial distance and interlayer transmission hops. The influence of interlayer transmission hops on transmission cost decreases non-linearly with increasing spatial distance. Specifically, a spatial distance influence weighting coefficient is constructed, which is calculated using an exponential decay method. The coefficient is based on the natural constant, and the negative spatial distance is multiplied by a decay factor as the exponent. The default value of the decay factor is 0.01, expressed in micrometers to the power of -1.

[0071] The contribution of interlayer transport hops to transport cost is equal to the interlayer transport hops multiplied by the hop count weight parameter, and then multiplied by the spatial distance influence weight coefficient. The default value of the hop count weight parameter is 10 micrometers per hop. The transport cost is equal to the spatial distance plus the contribution of interlayer transport hops to transport cost. When calculating the transport cost between any two through-silicon vias (TSVs) using the transport cost function, all TSV pairs in the spatial distribution matrix are traversed. For each TSV pair, its spatial distance and interlayer transport hops are extracted, and substituted into the transport cost function to calculate the transport cost of that TSV pair. The transport cost of all TSV pairs is stored in the transport cost matrix, which is a symmetric matrix.

[0072] When constructing a through-silicon via (TSV) association graph based on transmission cost value, the TSV association graph is represented by a graph data structure. Each node in the graph corresponds to one TSV, and the edges between nodes represent the association relationships between TSVs. The weight of the edge is the transmission cost value of the corresponding TSV pair. When the transmission cost value between two TSVs is less than the transmission cost threshold, an edge is created between the nodes corresponding to these two TSVs. The default value of the transmission cost threshold is the 20th percentile of the transmission cost values ​​of all TSV pairs.

[0073] When identifying connected components in the through-silicon via (TSV) correlation graph, a depth-first search (DFS) is used to traverse all nodes in the graph. Starting from any unvisited node, the DFS is performed, marking the current node as visited and recursively visiting all its adjacent nodes. After the DFS is completed, all visited nodes constitute a connected component. Each connected component identified by the DFS is used as an initial TSV transmission cluster. The criterion for classifying TSVs with strong correlation edges into the same connected component is that there is an edge between two TSVs and the weight of the edge is less than the strong correlation edge weight threshold. The default value of the strong correlation edge weight threshold is 50% of the transmission cost threshold. When calculating the transmission cost dispersion of each initial TSV transmission cluster, the transmission cost value of all TSV pairs within each initial TSV transmission cluster is extracted, and the standard deviation of these transmission cost values ​​is calculated as the transmission cost dispersion.

[0074] When the transmission cost dispersion of an initial through-silicon via (TSV) transport cluster exceeds the cohesion threshold, it is determined that the transmission cost difference between TSVs within the initial TSV transport cluster is too large, and the cluster needs to be re-clustered. The default value of the cohesion threshold is the median of the transmission cost dispersion of all initial TSV transport clusters. When identifying local peak points of transmission cost values ​​within an initial TSV transport cluster, all TSVs within the cluster are traversed. For each TSV, the sum of its transmission cost value and that of other TSVs within the cluster is calculated as the cumulative transmission cost value of that TSV. TSVs whose cumulative transmission cost value forms a local maximum within their neighborhood are marked as local peak points.

[0075] The neighborhood range is defined as the set of all vias whose transmission cost is less than the neighborhood cost threshold. The default value of the neighborhood cost threshold is 30% of the transmission cost threshold. When re-clustering the initial via transmission cluster with local peak points as the splitting core, each local peak point serves as the center node of a new sub-cluster. The vias with all non-local peak points within the cluster are traversed, and the transmission cost of each via with each local peak point is calculated. The via is then assigned to the sub-cluster corresponding to the local peak point with the smallest transmission cost. After re-clustering, multiple sub-clusters are obtained, and each sub-cluster constitutes a via transmission cluster with transmission cost cohesion.

[0076] In one optional implementation, a through-silicon via (TSV) transmission cost function is constructed, and the spatial distance and the interlayer transmission hop count are nonlinearly coupled according to the transmission cost function, such that the weight of the interlayer transmission hop count on the transmission cost decreases nonlinearly as the spatial distance increases, including:

[0077] Construct a through-silicon via (TSV) transport cost function, which includes a spatial distance term and an interlayer transport hops term;

[0078] An attenuation operator is constructed using spatial distance as the independent variable. The attenuation operator outputs an attenuation coefficient that decreases non-linearly as the spatial distance increases. The attenuation coefficient is then multiplied by the inter-layer transmission hop count to generate an attenuated inter-layer transmission hop count component.

[0079] The spatial distance and the interlayer transmission hop count components are weighted and summed to generate the transmission cost value between silicon via pairs. The transmission cost value between each silicon via pair is calculated according to the transmission cost function, thereby realizing the nonlinear coupling between the spatial distance and the interlayer transmission hop count, so that the influence weight of the interlayer transmission hop count on the transmission cost decreases nonlinearly as the spatial distance increases.

[0080] Before clustering vias, it is necessary to accurately assess the transmission cost between any two vias. The transmission cost depends not only on the physical spatial distance but also on the number of interlayer hops. Spatial distance reflects the physical path length for signal propagation, while the number of interlayer hops reflects the number of vertical layers the data needs to traverse. When two vias are located on different chip layers, data needs to traverse vertically multiple times to reach its target location, each traversal introducing additional latency and power consumption.

[0081] The construction of the transport cost function begins with defining two core terms: the spatial distance term, which represents the Euclidean distance between vias, calculated based on the coordinates of the vias in three-dimensional space; and the interlayer transport hop count term, which records the difference in the number of layers the data needs to traverse from the source via to the target via. When the source via is located on the first chip layer and the target via is located on the third chip layer, the interlayer transport hop count is 2.

[0082] A key phenomenon exists in actual transmission: when the spatial distance is small, the number of hops between layers has a significant impact on the overall transmission cost; however, as the spatial distance increases to a certain extent, the cost of interlayer traversal decreases as a percentage of the cost of long-distance horizontal transmission. To accurately characterize this nonlinear relationship, an attenuation operator is introduced to modulate the number of hops between layers.

[0083] The attenuation operator adopts a negative exponential function form, with spatial distance as the independent variable. Specifically, a function of the form exp(-λd) is chosen as the attenuation operator, where d represents the spatial distance and λ is the attenuation rate parameter. When the spatial distance d is close to zero, the attenuation operator outputs a coefficient close to 1; as d increases, the coefficient output by the attenuation operator decreases rapidly according to an exponential law and tends towards zero. The value of the attenuation rate parameter λ is determined based on the specific characteristics of the 3D packaging process, with a typical range between 0.01 and 0.1.

[0084] Multiplying the attenuation coefficient output by the attenuation operator by the interlayer transmission hop count yields the attenuated interlayer transmission hop count component. This component retains interlayer crossing information, while its weight adaptively adjusts with spatial distance. When the spatial distance between two through-silicon vias is small, the interlayer transmission hop count component maintains a large weight; as the spatial distance increases, the weight of this component automatically decreases.

[0085] The transmission cost is generated through a weighted summation mechanism, which involves multiplying the spatial distance term and the attenuated inter-layer transmission hop count component by their respective weighting coefficients and then summing them. The weighting coefficients for the spatial distance term are typically set to 0.6 to 0.8, and the weighting coefficients for the inter-layer transmission hop count component are set to 0.2 to 0.4; the sum of the two is 1 to maintain normalization. For any two through-silicon vias (TSVs) in the array, their spatial coordinates and layer numbers are extracted, the spatial distance and inter-layer transmission hop count are calculated, and the corresponding transmission cost is obtained by substituting them into the transmission cost function.

[0086] The transmission cost matrix is ​​calculated by iterating through all pairs of vias in the array and calculating the transmission cost between each pair. This matrix serves as the input to the subsequent clustering algorithm, ensuring that the clustering results comprehensively consider the transmission characteristics of both physical distance and inter-layer hop count, thus achieving a reasonable allocation of via resources.

[0087] In one optional implementation, the data to be transmitted is fragmented according to the number of the through-silicon via (TSV) transmission clusters and the channel bandwidth capacity of the TSVs within each cluster, generating a data fragment set corresponding to the TSV transmission clusters, and each data fragment is labeled with its corresponding TSV transmission cluster identifier, including:

[0088] Obtain the number of through-silicon via (TSV) transmission clusters, traverse all TSVs within each TSV transmission cluster and read the channel bandwidth capacity parameters of each TSV, statistically summarize the channel bandwidth capacity of each TSV within the same TSV transmission cluster, and generate the cluster-level bandwidth capacity of each TSV transmission cluster.

[0089] Establish a mapping relationship between the cluster-level bandwidth capacity and the data allocation weight. The mapping relationship determines the data carrying weight of each silicon via transmission cluster based on the relative size of the cluster-level bandwidth capacity of each cluster-level bandwidth capacity in all cluster-level bandwidth capacities, and generates the bandwidth capacity ratio of each silicon via transmission cluster.

[0090] The total amount of data to be transmitted is obtained. The total amount of data is decomposed according to the bandwidth capacity ratio of each through-silicon via (TSV) transmission cluster. The total amount of data is then divided into data segments that match the bandwidth capacity ratio of each TSV transmission cluster, and the data segment size that each TSV transmission cluster should carry is generated.

[0091] The data to be transmitted is continuously segmented according to the data segment size. Starting from the beginning of the data to be transmitted, data segments are sequentially extracted according to the data segment size to generate data segments equal to the number of through-silicon via (TSV) transmission clusters. Each data segment is appended with its corresponding TSV transmission cluster identifier to generate a set of data segments with TSV transmission cluster identifiers.

[0092] When partitioning the data to be transmitted according to the number of through-silicon via (TSV) transmission clusters and the channel bandwidth capacity of the TSVs within each cluster, the total number of TSV transmission clusters is read from the management module of the 3D packaging structure. This total number is stored as an integer in the cluster count register, which has a bit width of 16 bits. When traversing all TSVs within each TSV transmission cluster, the TSV configuration table is accessed. This configuration table is organized using a hash table structure, where the key is a globally unique identifier for the TSV, and the value is a structure containing channel bandwidth capacity parameters. The structure fields include nominal bandwidth capacity, real-time available bandwidth capacity, bandwidth capacity unit identifier, and bandwidth capacity measurement timestamp. Nominal bandwidth capacity represents the theoretical maximum transmission rate under the TSV design specifications, in megabits per second (Mbps), ranging from 1 to 10000 Mbps. Real-time available bandwidth capacity represents the current actual available transmission rate after considering physical factors such as signal attenuation, crosstalk interference, and temperature drift. It is periodically measured and updated by the bandwidth detection module, with a default update period of 100 milliseconds.

[0093] When reading the channel bandwidth capacity parameters of each through-silicon via (TSV), the real-time available bandwidth capacity is used as the valid value first. If the real-time available bandwidth capacity field is empty or the timestamp exceeds the timeliness window, the nominal bandwidth capacity is used as a fallback and multiplied by a safety margin factor, which defaults to 0.8. When summarizing the channel bandwidth capacity of each TSV within the same TSV transmission cluster, the cluster-level bandwidth capacity accumulator is initialized to a floating-point type with an initial value of 0.0. All TSV identifiers contained in the TSV transmission cluster are traversed, and the valid bandwidth capacity value of each TSV is extracted from the TSV configuration table in turn. The extracted bandwidth capacity values ​​are converted to a unified unit and then accumulated into the cluster-level bandwidth capacity accumulator.

[0094] If, during the accumulation process, the effective bandwidth capacity of a certain through-silicon via (TSV) is found to be negative or exceeds the physical upper limit threshold, the TSV is marked as abnormal and excluded from the accumulation calculation. The physical upper limit threshold is set to 20,000 megabits per second. After completing the traversal of all TSVs, the final value of the cluster-level bandwidth capacity accumulator is the cluster-level bandwidth capacity of the TSV transmission cluster, in megabits per second, with precision retained to two decimal places.

[0095] When establishing the mapping relationship between cluster-level bandwidth capacity and data allocation weights, the total cluster-level bandwidth capacity of all through-silicon via (TSV) transmission clusters is calculated. The global bandwidth capacity accumulator is initialized to a floating-point value of 0.0. All elements in the cluster-level bandwidth capacity array are traversed, and the values ​​of each element are accumulated into the global bandwidth capacity accumulator. The total global cluster-level bandwidth capacity is used to subsequently calculate the bandwidth capacity ratio of each TSV transmission cluster. When the total global cluster-level bandwidth capacity is less than the minimum available bandwidth capacity threshold, it is determined that the overall available bandwidth of the TSV array in the 3D packaging structure is insufficient, triggering a transmission degradation strategy. The minimum available bandwidth capacity threshold is 100 megabits per second by default.

[0096] When determining the data carrying weight of each TSV transmission cluster based on its relative size among all cluster-level bandwidth capacities, the ratio of each TSV transmission cluster's cluster-level bandwidth capacity to the total global cluster-level bandwidth capacity is calculated. This ratio represents the bandwidth capacity percentage of that TSV transmission cluster. The bandwidth capacity percentage is a dimensionless floating-point number, ranging from 0 to 1, with a precision of four decimal places. Theoretically, the sum of all elements in the bandwidth capacity percentage array should equal 1.0. However, in actual calculations, slight deviations occur due to the loss of precision in floating-point operations. When the absolute value of the deviation exceeds the precision tolerance, percentage normalization is performed. The default precision tolerance is 0.0001.

[0097] When obtaining the total amount of data to be transmitted, the number of bytes of data to be transmitted is read from the data transmission task descriptor. The total amount of data is represented as a 64-bit unsigned integer. When decomposing the total amount of data according to the bandwidth capacity ratio of each through-silicon via (TSV) transmission cluster, the bandwidth capacity ratio array is traversed. For each TSV transmission cluster, the data fragment size that it should carry is calculated. The data fragment size is equal to the total amount of data multiplied by the bandwidth capacity ratio of that TSV transmission cluster. The product is a floating-point number and needs to be rounded down to an integer number of bytes. Rounding down results in the sum of the data fragment sizes of each TSV transmission cluster being less than the total amount of data, resulting in the remaining number of bytes.

[0098] The allocation of remaining bytes adopts a maximum balance priority strategy. The fractional part of each through-silicon via (TSV) transport cluster before rounding down is calculated as the balance value. The remaining bytes are allocated byte-by-byte to the TSV transport cluster with the largest balance value. After each byte is allocated, the data fragment size of that TSV transport cluster is updated and the balance values ​​are reordered. This process is repeated until all remaining bytes are allocated. The data fragment size to be carried by each TSV transport cluster, after allocation of remaining bytes, is stored in a data fragment size array. The array elements are 64-bit unsigned integers representing the number of bytes, and the sum of all elements in the data fragment size array is strictly equal to the total data volume.

[0099] When performing continuous segmentation of the data to be transmitted according to the data segment size, the data segmentation cursor is initialized to point to the address of the first byte of the data to be transmitted. The data to be transmitted is stored in the data buffer as a continuous byte stream. When extracting data segments sequentially according to the data segment size starting from the beginning of the data to be transmitted, the data segment size array is traversed, and the corresponding data segment size is read for each through-silicon via (TSV) transmission cluster. A specified number of bytes of data segment are extracted from the current position of the data segmentation cursor. The extraction operation is implemented through a memory copy instruction, with the source address being the current value of the data segmentation cursor, the destination address being the starting address of the newly allocated data segment buffer, and the copy length being the data segment size.

[0100] The data fragment buffer is dynamically allocated in heap memory, reserving additional space for a metadata header during allocation. This header includes fields such as the data fragment identifier, the through-silicon via (TSV) transport cluster identifier, the data fragment size, the data fragment checksum, and the data fragment creation timestamp. The data fragment identifier is generated using a globally incrementing sequence number, starting at 1 and incrementing by 1 after each data fragment is generated. The sequence number is represented as a 32-bit unsigned integer. The TSV transport cluster identifier is directly mapped from the TSV transport cluster sequence number, with a one-to-one correspondence between the sequence number and the identifier. The identifier is represented as a 16-bit unsigned integer. The data fragment checksum is calculated using a 32-bit cyclic redundancy check algorithm on the actual data content in the data fragment buffer. The checksum is represented as a 32-bit unsigned integer and is used for data integrity verification during subsequent transmission.

[0101] After a data segment is extracted, the data segmentation cursor is moved forward by the data segment size (bytes), and the updated cursor position points to the starting address of the next data segment to be extracted. When generating data segments equal to the number of through-silicon via (TSV) transport clusters, the total number of data segments equals the total number of TSV transport clusters, with each TSV transport cluster corresponding to one data segment. When attaching the corresponding TSV transport cluster identifier to each data segment, the TSV transport cluster identifier field in the data segment metadata header is filled with the TSV transport cluster sequence number corresponding to that data segment. This filling operation is completed after the data segment buffer is allocated and before the data content is copied. When generating a set of data segments with TSV transport cluster identifiers, the addresses of all generated data segment buffers are added to the data segment set container. The container is implemented using a dynamic array structure, where array elements are pointers to data segment buffers. The array length equals the total number of data segments, and the array index is consistent with the TSV transport cluster sequence number, supporting quick location of the corresponding data segment by TSV transport cluster sequence number.

[0102] In one optional implementation, allocating each data fragment to a corresponding through-silicon via (TSV) transport cluster based on the TSV transport cluster identifier, and establishing a bidirectional state mapping relationship between the data fragments and the TSV transport clusters during transmission includes:

[0103] Parse the through-silicon via (TSV) transport cluster identifier carried by each data fragment, and route each data fragment to its corresponding TSV transport cluster according to the TSV transport cluster identifier;

[0104] When a data shard enters a through-silicon via (TSV) transport cluster, a transport status awareness layer is established. The data shard is used as an observation object based on the transport status awareness layer, and the transport behavior characteristics of the data shard within the TSV transport cluster are captured. The transport behavior characteristics are then encoded as the transport behavior fingerprint of the data shard.

[0105] A cluster-level resource response layer is established synchronously. Based on the cluster-level resource response layer, the through-silicon via (TSV) transmission cluster is taken as the observation object, and the resource response characteristics of the TSV transmission cluster to the data fragmentation are monitored. The resource response characteristics are encoded as the resource response fingerprint of the TSV transmission cluster.

[0106] Cross-correlation analysis is performed on the transmission behavior fingerprint and the resource response fingerprint to establish a bidirectional state mapping relationship between data fragmentation and through-silicon via transmission clusters.

[0107] During the data fragment allocation and bidirectional state mapping process, the parsing mechanism first extracts the ToSAPV transport cluster identifier field carried in the data fragment header. This identifier field uses 16-bit binary encoding, where the high 8 bits represent the global index of the target ToSAPV transport cluster, and the low 8 bits record the priority parameter of the data fragment at the source end. The routing module queries a pre-built cluster identifier mapping table to match the identifier value with the physical address of the specific ToSAPV transport cluster, thus determining the routing path. When multiple data fragments point to the same ToSAPV transport cluster at the same time, the data fragments are sorted according to the priority parameter, and the data fragment with the higher priority value gets priority transmission.

[0108] The transmission status awareness layer is activated the instant a data fragment enters the through-silicon via (TSV) transmission cluster, achieved by deploying a status monitoring unit at the entry node of the TSV array. The monitoring unit uses a time window as the basic observation unit, with the length of the time window set to the single transmission cycle of the TSV. It captures behavioral characteristic parameters of the data fragment within this time window, such as transmission delay, queue dwell time, and retransmission count. Transmission delay is obtained by recording the difference between the timestamp of the data fragment entering the TSV transmission cluster and the timestamp of its completion. Queue dwell time reflects the time the data fragment spends in the waiting queue, and the retransmission count counts duplicate transmissions caused by transmission errors. These characteristic parameters are weighted and summed according to preset weights. The weighting coefficients are determined based on statistical analysis of historical transmission data: transmission delay has a weight of 0.5, queue dwell time has a weight of 0.3, and retransmission count has a weight of 0.2. The weighted result is mapped to a 64-bit transmission behavior fingerprint using a hash function.

[0109] The cluster-level resource response layer is established based on a global perspective of the through-silicon via (TSV) transmission cluster. Monitoring targets include resource response characteristics such as instantaneous bandwidth utilization, buffer remaining capacity, and power consumption levels of the TSV transmission cluster. Instantaneous bandwidth utilization is obtained by calculating the ratio of the current transmitted data volume to the theoretical maximum transmission capacity of the TSV transmission cluster. Buffer remaining capacity is obtained by real-time reading of the free space index from the buffer management unit. Power consumption levels are characterized by instantaneous power consumption data provided by the power management module. These resource response characteristics are also weighted to generate resource response fingerprints, with bandwidth utilization weighted at 0.4, buffer remaining capacity weighted at 0.35, and power consumption level weighted at 0.25. The combined result is encoded into a 64-bit resource response fingerprint using the same hash function as the transmission behavior fingerprint.

[0110] The bidirectional state mapping relationship is established through cross-correlation analysis, specifically using the sliding window correlation calculation method. The transmission behavior fingerprint sequence and the resource response fingerprint sequence are treated as two time-series signals, and their correlation coefficients are calculated at different time offsets. The peak position of the correlation coefficient reflects the delay in the impact of data fragmentation transmission behavior on the resource response of the through-silicon via (TSV) transmission cluster, and the peak amplitude represents the coupling strength between the two. When the correlation coefficient exceeds 0.7, a strong correlation is considered to exist between the data fragmentation and the TSV transmission cluster, and the established bidirectional state mapping relationship is marked as valid. This mapping relationship is stored in the form of an association matrix, where the row index corresponds to the data fragmentation identifier, the column index corresponds to the TSV transmission cluster identifier, and the matrix element values ​​are composite parameters of the correlation coefficient and the time offset, providing a quantitative basis for subsequent capacity saturation prediction and cross-cluster migration decisions.

[0111] In one optional implementation, predicting the capacity saturation time of each through-silicon via (TSV) transport cluster based on the bidirectional state mapping relationship, and migrating the data to be transmitted in fragments to unsaturated TSV transport clusters before the capacity saturation time arrives includes:

[0112] Extract the resource response fingerprint of each through-silicon via (TSV) transmission cluster and the transmission behavior fingerprint of each data segment from the bidirectional state mapping relationship.

[0113] Capacity evolution trajectory is constructed based on the load fluctuation pattern of each through-silicon via (TSV) transmission cluster. The capacity evolution trajectory depicts the nonlinear decay process of the available capacity of the TSV transmission cluster over time. The time node when the available capacity drops to the capacity threshold is marked on the capacity evolution trajectory, and the time node is taken as the capacity saturation moment.

[0114] Identify the through-silicon via transport clusters whose capacity saturation time falls within the warning time window, and select data fragments whose interlayer hopping frequency exceeds a preset hopping threshold from the through-silicon via transport clusters as migration candidate fragments;

[0115] Traverse the through-silicon via (TSV) transmission clusters whose capacity saturation time is outside the warning time window, and perform a compatibility evaluation based on the transmission behavior fingerprint of the migration candidate fragment and the resource response fingerprint of each TSV transmission cluster. Select the TSV transmission cluster with the best compatibility evaluation value as the receiving cluster.

[0116] Before the capacity saturation time is reached, the migration candidate fragment is migrated from the original through-silicon via (TSV) transmission cluster to the receiving cluster, and the association record between the migration candidate fragment and the receiving cluster is updated in the bidirectional state mapping relationship, thus completing the cross-cluster migration of the data fragment to be transmitted to the unsaturated TSV transmission cluster.

[0117] When predicting the capacity saturation time of each through-silicon via (TSV) transport cluster based on the bidirectional state mapping relationship and performing cross-cluster migration, the resource response fingerprint of each TSV transport cluster and the transmission behavior fingerprint of each data fragment are extracted from the bidirectional state mapping relationship. The bidirectional state mapping relationship is stored in a bidirectional hash table structure. The forward mapping table uses the TSV transport cluster identifier as the key and the value is the set of all data fragment identifiers currently carried by the TSV transport cluster. The reverse mapping table uses the data fragment identifier as the key and the value is the TSV transport cluster identifier to which the data fragment currently belongs.

[0118] Resource response fingerprints characterize the dynamic response properties of the TSV transmission cluster when carrying data transmission tasks, including multi-dimensional indicators such as instantaneous throughput, queue depth, latency jitter, and error rate. Each indicator is updated at a fixed sampling period, which defaults to 50 milliseconds. Instantaneous throughput represents the amount of effective data transmitted by the TSV transmission cluster within the current sampling period, measured in megabits per second (Mbps). Queue depth represents the number of data fragments currently backed up in the queue of data to be transmitted by the TSV transmission cluster; the rate of increase in queue depth reflects the load pressure on the TSV transmission cluster. Latency jitter represents the standard deviation of the transmission delay of the TSV transmission cluster, calculated by collecting the transmission delay samples of the most recent 100 data fragments, measured in milliseconds. Error rate represents the probability of transmission failure in the most recent 1000 transmission attempts by the TSV transmission cluster, ranging from 0 to 1.

[0119] Transmission behavior fingerprints characterize the behavioral patterns of data fragments during transmission, including metrics such as average transmission rate, transmission burst rate, inter-layer hop frequency, and retransmission count. Average transmission rate represents the cumulative average transmission rate of the data fragment from the start of transmission to the current moment, measured in megabits per second (Mbps), calculated by dividing the number of bytes transmitted by the elapsed transmission time. Transmission burst rate indicates the degree of temporal fluctuation in the data fragment transmission rate, obtained by calculating the coefficient of variation of the transmission rate time series. Inter-layer hop frequency represents the ratio of the number of cross-layer via jumps traversed by the data fragment in the transmission path to the total number of jumps, ranging from 0 to 1. A higher inter-layer hop frequency indicates a stronger dependence of the data fragment on cross-layer transmission capabilities. Retransmission count represents the cumulative number of retransmission operations triggered by transmission errors, recorded as an integer.

[0120] When constructing the capacity evolution trajectory based on the load fluctuation patterns of each through-silicon via (TSV) transmission cluster, historical load data sequences of the TSV transmission clusters are extracted. These historical load data sequences record the instantaneous throughput and queue depth of the TSV transmission cluster at each sampling moment within a 600-second time window. The capacity evolution trajectory depicts the nonlinear decay process of the available capacity of the TSV transmission cluster over time. Available capacity is defined as the cluster-level bandwidth capacity of the TSV transmission cluster minus the current instantaneous throughput, in megabits per second (Mbps). The nonlinear decay process is fitted using an exponentially weighted moving average model. Model parameters include a decay coefficient and a smoothing factor. The decay coefficient defaults to 0.95, and the smoothing factor defaults to 0.7. The predicted available capacity at the current moment is equal to the smoothing factor multiplied by the actual available capacity at the current moment plus 1 minus the smoothing factor, multiplied by the predicted available capacity at the previous moment. The recursive calculation starts from the beginning of the historical data sequence.

[0121] When the available capacity drops to the capacity threshold on the capacity evolution trajectory, the capacity threshold is set to 10% of the cluster-level bandwidth capacity, indicating that the available capacity of the through-silicon via (TSV) transmission cluster is nearing exhaustion. The available capacity values ​​at each moment on the capacity evolution trajectory are predicted, and the first moment when the predicted available capacity value is less than the capacity threshold is found. This moment is the capacity saturation moment, expressed as the time difference from the current moment in seconds. If the predicted available capacity values ​​at all moments within the predicted range of the capacity evolution trajectory for the next 300 seconds are not less than the capacity threshold, then it is determined that the TSV transmission cluster will not experience capacity saturation within the foreseeable timeframe.

[0122] When identifying via transport clusters whose capacity saturation time falls within the warning time window, the warning time window is defined as the time interval from the current moment to 60 seconds in the future, representing the warning period requiring early intervention. Via transport clusters with capacity saturation times less than 60 seconds are selected and added to the set of source clusters to be migrated. When selecting data fragments with interlayer hop frequencies exceeding a preset hop threshold from each via transport cluster in the set of source clusters to be migrated as migration candidate fragments, the preset hop threshold defaults to 0.6. The forward mapping table of the bidirectional state mapping relationship is accessed to extract all data fragment identifiers currently carried by each via transport cluster in the set of source clusters to be migrated. The interlayer hop frequencies of each data fragment are compared with the preset hop threshold, and data fragments with interlayer hop frequencies greater than 0.6 are selected to generate a set of migration candidate fragments.

[0123] When traversing through SWIFT transmission clusters whose capacity saturation time falls outside the warning time window, SWIFT transmission clusters with a capacity saturation time greater than or equal to 60 seconds are selected and added to the candidate receiver cluster set. When evaluating the compatibility between the transmission behavior fingerprint of the migration candidate fragments and the resource response fingerprint of each SWIFT transmission cluster, for each data fragment in the migration candidate fragment set, each SWIFT transmission cluster in the candidate receiver cluster set is traversed, and the compatibility evaluation value between the data fragment and the SWIFT transmission cluster is calculated. The compatibility evaluation value comprehensively considers the weighted scores of three dimensions: transmission rate matching degree, latency jitter compatibility, and error rate tolerance.

[0124] Transmission rate matching is equal to the smaller of the average transmission rate of data fragments and the instantaneous throughput of the through-silicon via (TSV) transmission cluster, divided by the larger value, ranging from 0 to 1. Delay jitter compatibility is equal to 1 minus the absolute value of the difference between the burst rate of data fragment transmission and the delay jitter of the TSV transmission cluster divided by 100 milliseconds. Error rate tolerance is equal to 1 minus the error rate of the TSV transmission cluster. The adaptability evaluation value is equal to transmission rate matching multiplied by 0.5, delay jitter compatibility multiplied by 0.3, and error rate tolerance multiplied by 0.2.

[0125] When selecting the through-silicon via (TSV) transmission cluster with the best adaptability evaluation value as the receiving cluster, the TSV transmission cluster with the largest adaptability evaluation value is selected. If multiple TSV transmission clusters have the same adaptability evaluation value and are all the maximum value, then the available capacity of these TSV transmission clusters is further compared, and the TSV transmission cluster with the largest available capacity is selected as the receiving cluster.

[0126] When migrating candidate fragments from the original through-silicon via (TSV) transmission cluster to the receiving cluster before the capacity saturation time is reached, the migration trigger time is calculated. The migration trigger time is equal to the capacity saturation time minus a 10-second migration preparation time. When the current time reaches or exceeds the migration trigger time, the cross-cluster migration process is initiated. Before migration, it is checked and verified that the current available capacity of the receiving cluster is greater than 1.2 times the average transmission rate of the data fragment to be migrated.

[0127] Data migration involves removing the data fragment to be migrated from the original through-silicon via (TSV) transmission cluster's transmission queue and adding it to the receiving cluster's transmission queue. State synchronization updates the associated records in the bidirectional state mapping relationship. In the forward mapping table, the identifier of the data fragment to be migrated is deleted from the data fragment set corresponding to the original TSV transmission cluster, and this identifier is added to the data fragment set corresponding to the receiving cluster. In the reverse mapping table, the TSV transmission cluster identifier corresponding to the data fragment to be migrated is updated to the receiving cluster identifier.

[0128] In one optional implementation, adjusting the fragmentation granularity of subsequent data to be transmitted using the capacity-topology characteristic parameters, so that the size distribution of the data fragments and the capacity saturation period of each through-silicon via (TSV) transmission cluster form a time mismatch, includes:

[0129] The capacity saturation period and topological connectivity of each through-silicon via transmission cluster are extracted from the capacity-topology characteristic parameters. The capacity saturation period is projected onto the cyclic time axis to construct a capacity saturation rhythm diagram. Peak superposition regions and trough isolation regions are identified in the capacity saturation rhythm diagram.

[0130] A piecewise granularity inverse control function is constructed. The piecewise granularity inverse control function uses the superposition intensity of the peak superposition region as input to generate a granularity suppression coefficient, uses the isolation spacing of the valley isolation region as input to generate a granularity release coefficient, and introduces the topological connectivity as a nonlinear modulation parameter into the calculation of the granularity suppression coefficient and the granularity release coefficient.

[0131] The fragmentation granularity inverse control function is used to perform variable granularity segmentation on the subsequent data to be transmitted, generating fine-grained data fragments at the silicon via transmission clusters corresponding to the peak superposition region and coarse-grained data fragments at the silicon via transmission clusters corresponding to the valley isolation region.

[0132] The fine-grained data fragments and the coarse-grained data fragments are allocated to each through-silicon via (TSV) transmission cluster, so that the size distribution of the data fragments and the capacity saturation period of each TSV transmission cluster are mismatched in time.

[0133] like Figure 2 As shown, the method includes:

[0134] When adjusting the fragmentation granularity of subsequent data to be transmitted using capacity-topology feature parameters, the capacity saturation period and topological connectivity of each through-silicon via (TSV) transport cluster are extracted from the capacity-topology feature parameters. The capacity saturation period represents the time span required for a TSV transport cluster to evolve from a state of sufficient available capacity to a state of capacity saturation and then back to a state of sufficient available capacity, measured in seconds. It is calculated by analyzing the time interval between two adjacent capacity saturation moments in the historical capacity evolution trajectory, and the average of the most recent 10 periods is taken, with a default value range of 60 to 600 seconds. Topological connectivity represents the density of spatial adjacency relationships between a TSV transport cluster and other TSV transport clusters in the three-dimensional packaging structure. It is calculated by counting the number of direct paths between the TSV transport cluster and adjacent TSV transport clusters and dividing by the theoretical maximum number of direct paths, with a value range of 0 to 1.

[0135] When constructing a capacity saturation rhythm diagram by projecting the capacity saturation period onto the cyclic time axis, the cyclic time axis uses the greatest common divisor of the capacity saturation periods of all through-silicon via (TSV) transport clusters as its period length, which is 60 seconds by default. The capacity saturation period of each TSV transport cluster is mapped onto the cyclic time axis. The capacity saturation time of the TSV transport cluster is marked at time 0 on the cyclic time axis, and subsequent capacity saturation times are marked at equal intervals according to the capacity saturation period. The capacity saturation rhythm diagram is presented as a two-dimensional planar graph, with the cyclic time axis on the horizontal axis and the TSV transport cluster identifiers on the vertical axis. When identifying peak overlap regions and trough isolation regions in the capacity saturation rhythm diagram, a sliding window scan is performed along the cyclic time axis with a time step of 1 second. The window length is 10 seconds by default. The number of marked points within each window position is counted as the overlap intensity. Window regions with an overlap intensity greater than 30% of the total number of TSV transport clusters are identified as peak overlap regions, and window regions with an overlap intensity less than 10% of the total number of TSV transport clusters are identified as trough isolation regions.

[0136] When constructing the granularity inverse control function, this function receives the superposition intensity of the crest superposition region, the isolation spacing of the trough isolation region, and the topological connectivity as input parameters, and outputs a granularity suppression coefficient and a granularity release coefficient. The superposition intensity represents the cumulative number of marker points within the crest superposition region, and the isolation spacing represents the duration of the trough isolation region on the cyclic time axis. The granularity suppression coefficient is used to reduce the data granularity size, with a value ranging from 0.1 to 1. The granularity release coefficient is used to increase the data granularity size, with a value ranging from 1 to 10. When using the superposition intensity of the crest superposition region as input to generate the granularity suppression coefficient, the granularity suppression coefficient is equal to 1 minus the ratio of the superposition intensity to the total number of through-silicon via (TSV) transmission clusters, multiplied by 0.8. If the result is less than 0.1, it is truncated to 0.1. When using the isolation spacing of the trough isolation region as input to generate the granularity release coefficient, the granularity release coefficient is equal to 1 plus the isolation spacing divided by the cyclic time axis period length, multiplied by 5. If the result is greater than 10, it is truncated to 10.

[0137] When topological connectivity is introduced as a nonlinear modulation parameter into the calculation of granularity suppression coefficient and granularity release coefficient, the modulated granularity suppression coefficient is equal to the original granularity suppression coefficient multiplied by the negative topological connectivity of the natural constant multiplied by the square root, and the modulated granularity release coefficient is equal to the original granularity release coefficient multiplied by the topological connectivity of the natural constant multiplied by the square root.

[0138] When performing variable-granularity segmentation of subsequent data to be transmitted using the fragmentation granularity inverse control function, the total data volume of the subsequent data to be transmitted is obtained. Each through-silicon via (TSV) transmission cluster is traversed, and the corresponding region type in the capacity saturation rhythm diagram is queried. Region types include peak superposition region, trough isolation region, and neutral region. For TSV transmission clusters corresponding to peak superposition regions, the fine-grained data fragment size is calculated using the modulated granularity suppression coefficient. The fine-grained data fragment size is equal to the baseline fragment size of the TSV transmission cluster multiplied by the granularity suppression coefficient. The baseline fragment size is calculated by multiplying the bandwidth capacity ratio of the TSV transmission cluster by the total data volume. For TSV transmission clusters corresponding to trough isolation regions, the coarse-grained data fragment size is calculated using the modulated granularity release coefficient. The coarse-grained data fragment size is equal to the baseline fragment size multiplied by the granularity release coefficient. For TSV transmission clusters corresponding to neutral regions, the baseline fragment size is directly used as the data fragment size.

[0139] When generating fine-grained data fragments at the via transmission clusters corresponding to the peak overlap region, data segments are extracted from subsequent data to be transmitted according to the fine-grained data fragment size. The data segmentation cursor starts from the beginning of the data to be transmitted and moves sequentially forward by the fine-grained data fragment size. The generated fine-grained data fragments are marked with a fine-grained identifier in the granularity type field of the metadata header. When generating coarse-grained data fragments at the via transmission clusters corresponding to the trough isolation region, data segments are extracted according to the coarse-grained data fragment size. The generated coarse-grained data fragments are marked with a coarse-grained identifier in the granularity type field of the metadata header.

[0140] When allocating fine-grained and coarse-grained data fragments to various through-silicon via (TSV) transport clusters, the allocation routing is performed according to the TSV transport cluster identifier field in the data fragment metadata header. Fine-grained data fragments are allocated to TSV transport clusters corresponding to the peak overlap region, while coarse-grained data fragments are allocated to TSV transport clusters corresponding to the trough isolation region. The size distribution of data fragments creates a time mismatch with the capacity saturation period of each TSV transport cluster. During peak overlap periods where capacity saturation events occur frequently, the data fragment size allocated to the corresponding TSV transport cluster for that period is reduced to fine-grained size, reducing the transmission time of a single data fragment and preventing multiple TSV transport clusters from simultaneously entering capacity saturation. During trough isolation periods where capacity saturation events occur sparsely, the data fragment size allocated to the corresponding TSV transport cluster for that period is enlarged to coarse-grained size, improving the transmission efficiency of a single data fragment and fully utilizing the surplus capacity of the TSV transport cluster during that period.

[0141] A second aspect of the present invention provides a system for improving the efficiency of inter-chip data transmission in 3D packaging structures, comprising:

[0142] The clustering calculation unit is used to obtain the spatial distribution matrix of the data to be transmitted and the through-silicon via array in the three-dimensional packaging structure, calculate the spatial distance between through-silicon vias and the interlayer transmission hops based on the spatial distribution matrix, and cluster the through-silicon vias according to the spatial distance and the interlayer transmission hops to generate multiple through-silicon via transmission clusters.

[0143] The data fragmentation unit is used to divide the data to be transmitted into fragments according to the number of the through-silicon via (TSV) transmission clusters and the channel bandwidth capacity of the TSVs in each cluster, generate a data fragment set corresponding to the TSV transmission clusters, and label each data fragment with its corresponding TSV transmission cluster identifier. Based on the TSV transmission cluster identifier, each data fragment is allocated to the corresponding TSV transmission cluster, and a bidirectional state mapping relationship between the data fragments and the TSV transmission clusters is established during the transmission process.

[0144] The cross-cluster migration unit is used to predict the capacity saturation time of each through-silicon via (TSV) transmission cluster based on the bidirectional state mapping relationship, and to migrate the data to be transmitted to the unsaturated TSV transmission clusters before the capacity saturation time is reached.

[0145] The feature adjustment unit is used to correlate the capacity saturation time of each through-silicon via (TSV) transmission cluster with the spatial distance, generate the capacity-topology feature parameters of the TSV transmission cluster, and use the capacity-topology feature parameters to adjust the fragmentation granularity of subsequent data to be transmitted, so that the size distribution of the data fragments forms a time mismatch with the capacity saturation period of each TSV transmission cluster.

[0146] A third aspect of the present invention provides an electronic device, comprising:

[0147] processor;

[0148] Memory used to store processor-executable instructions;

[0149] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0150] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0151] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for improving inter-chip data transmission efficiency in 3D packaging structures, characterized in that, include: Obtain the spatial distribution matrix of the data to be transmitted and the through-silicon via array in the three-dimensional packaging structure. Calculate the spatial distance between through-silicon vias and the interlayer transmission hop count based on the spatial distribution matrix. Then, cluster the through-silicon vias according to the spatial distance and the interlayer transmission hop count to generate multiple through-silicon via transmission clusters. The data to be transmitted is divided into fragments according to the number of the through-silicon via (TSV) transmission clusters and the channel bandwidth capacity of the TSVs in each cluster, generating a data fragment set corresponding to the TSV transmission clusters. Each data fragment is labeled with its corresponding TSV transmission cluster identifier. Each data fragment is assigned to the corresponding TSV transmission cluster according to the TSV transmission cluster identifier. During the transmission process, a bidirectional state mapping relationship between the data fragments and the TSV transmission clusters is established. Based on the bidirectional state mapping relationship, the capacity saturation time of each through-silicon via (TSV) transmission cluster is predicted, and before the capacity saturation time is reached, the data to be transmitted is fragmented and migrated across clusters to unsaturated TSV transmission clusters. The capacity saturation time of each through-silicon via (TSV) transmission cluster is correlated with the spatial distance to generate a capacity-topology characteristic parameter for the TSV transmission cluster. The capacity-topology characteristic parameter is used to adjust the granularity of subsequent data fragmentation to be transmitted, so that the size distribution of the data fragments is mismatched with the capacity saturation period of each TSV transmission cluster.

2. The method of claim 1, wherein, Based on the spatial distribution matrix, the spatial distance between through-silicon vias (TSVs) and the interlayer transport hop count are calculated. The TSVs are then clustered according to the spatial distance and the interlayer transport hop count to generate multiple TSV transport clusters, including: The three-dimensional coordinates and interlayer interconnect topology of each through-silicon via (TSV) are extracted from the spatial distribution matrix. The spatial distance between TSVs is calculated based on the three-dimensional coordinates, and the interlayer transport hop count between TSVs is traced based on the interlayer interconnect topology. Construct a through-silicon via (TSV) transmission cost function, and nonlinearly couple the spatial distance and the interlayer transmission hop count according to the transmission cost function, so that the influence weight of the interlayer transmission hop count on the transmission cost decreases nonlinearly as the spatial distance increases, and use the transmission cost function to calculate the transmission cost between any two TSVs. Based on the transmission cost, a through-silicon via (TSV) correlation map is constructed. Connected components are identified in the TSV correlation map. TSVs with strong correlation edges are divided into the same connected component, and each connected component is used as an initial TSV transmission cluster. At the same time, the transmission cost dispersion of each initial TSV transmission cluster is calculated. When the transmission cost dispersion of an initial through-silicon via (TSV) transport cluster exceeds the cohesion threshold, local peak points of the transmission cost value are identified within the initial TSV transport cluster. The initial TSV transport cluster is then re-clustered using the local peak points as the splitting cores to generate multiple TSV transport clusters with transmission cost cohesion.

3. The method of claim 2, wherein, Constructing a through-silicon via (TSV) transmission cost function, and nonlinearly coupling the spatial distance and the interlayer transmission hop count based on the transmission cost function, so that the weight of the interlayer transmission hop count on the transmission cost decreases nonlinearly with the increase of spatial distance, including: Construct a through-silicon via (TSV) transport cost function, which includes a spatial distance term and an interlayer transport hops term; An attenuation operator is constructed using spatial distance as the independent variable. The attenuation operator outputs an attenuation coefficient that decreases non-linearly as the spatial distance increases. The attenuation coefficient is then multiplied by the inter-layer transmission hop count to generate an attenuated inter-layer transmission hop count component. The spatial distance and the interlayer transmission hop count components are weighted and summed to generate the transmission cost value between silicon via pairs. The transmission cost value between each silicon via pair is calculated according to the transmission cost function, thereby realizing the nonlinear coupling between the spatial distance and the interlayer transmission hop count, so that the influence weight of the interlayer transmission hop count on the transmission cost decreases nonlinearly as the spatial distance increases.

4. The method of claim 1, wherein, The data to be transmitted is divided into fragments according to the number of the through-silicon via (TSV) transmission clusters and the channel bandwidth capacity of the TSVs within each cluster, generating a data fragment set corresponding to each TSV transmission cluster. Each data fragment is labeled with its corresponding TSV transmission cluster identifier, including: Obtain the number of through-silicon via (TSV) transmission clusters, traverse all TSVs within each TSV transmission cluster and read the channel bandwidth capacity parameters of each TSV, statistically summarize the channel bandwidth capacity of each TSV within the same TSV transmission cluster, and generate the cluster-level bandwidth capacity of each TSV transmission cluster. Establish a mapping relationship between the cluster-level bandwidth capacity and the data allocation weight. The mapping relationship determines the data carrying weight of each silicon via transmission cluster based on the relative size of the cluster-level bandwidth capacity of each cluster-level bandwidth capacity in all cluster-level bandwidth capacities, and generates the bandwidth capacity ratio of each silicon via transmission cluster. The total amount of data to be transmitted is obtained. The total amount of data is decomposed according to the bandwidth capacity ratio of each through-silicon via (TSV) transmission cluster. The total amount of data is then divided into data segments that match the bandwidth capacity ratio of each TSV transmission cluster, and the data segment size that each TSV transmission cluster should carry is generated. The data to be transmitted is continuously segmented according to the data segment size. Starting from the beginning of the data to be transmitted, data segments are sequentially extracted according to the data segment size to generate data segments equal to the number of through-silicon via (TSV) transmission clusters. Each data segment is appended with its corresponding TSV transmission cluster identifier to generate a set of data segments with TSV transmission cluster identifiers.

5. The method according to claim 1, characterized in that, Based on the through-silicon via (TSV) transport cluster identifier, each data fragment is assigned to the corresponding TSV transport cluster, and a bidirectional state mapping relationship is established between the data fragment and the TSV transport cluster during transmission, including: Parse the through-silicon via (TSV) transport cluster identifier carried by each data fragment, and route each data fragment to its corresponding TSV transport cluster according to the TSV transport cluster identifier; When a data shard enters a through-silicon via (TSV) transport cluster, a transport status awareness layer is established. The data shard is used as an observation object based on the transport status awareness layer, and the transport behavior characteristics of the data shard within the TSV transport cluster are captured. The transport behavior characteristics are then encoded as the transport behavior fingerprint of the data shard. A cluster-level resource response layer is established synchronously. Based on the cluster-level resource response layer, the through-silicon via (TSV) transmission cluster is taken as the observation object, and the resource response characteristics of the TSV transmission cluster to the data fragmentation are monitored. The resource response characteristics are encoded as the resource response fingerprint of the TSV transmission cluster. Cross-correlation analysis is performed on the transmission behavior fingerprint and the resource response fingerprint to establish a bidirectional state mapping relationship between data fragmentation and through-silicon via transmission clusters.

6. The method according to claim 1, characterized in that, Based on the bidirectional state mapping relationship, the capacity saturation time of each through-silicon via (TSV) transmission cluster is predicted, and the data to be transmitted is fragmented and migrated across clusters to unsaturated TSV transmission clusters before the capacity saturation time is reached. This includes: Extract the resource response fingerprint of each through-silicon via (TSV) transmission cluster and the transmission behavior fingerprint of each data segment from the bidirectional state mapping relationship. Capacity evolution trajectory is constructed based on the load fluctuation pattern of each through-silicon via (TSV) transmission cluster. The capacity evolution trajectory depicts the nonlinear decay process of the available capacity of the TSV transmission cluster over time. The time node when the available capacity drops to the capacity threshold is marked on the capacity evolution trajectory, and the time node is taken as the capacity saturation moment. Identify the through-silicon via transport clusters whose capacity saturation time falls within the warning time window, and select data fragments whose interlayer hopping frequency exceeds a preset hopping threshold from the through-silicon via transport clusters as migration candidate fragments; Traverse the through-silicon via (TSV) transmission clusters whose capacity saturation time is outside the warning time window, and perform a compatibility evaluation based on the transmission behavior fingerprint of the migration candidate fragment and the resource response fingerprint of each TSV transmission cluster. Select the TSV transmission cluster with the best compatibility evaluation value as the receiving cluster. Before the capacity saturation time is reached, the migration candidate fragment is migrated from the original through-silicon via (TSV) transmission cluster to the receiving cluster, and the association record between the migration candidate fragment and the receiving cluster is updated in the bidirectional state mapping relationship, thus completing the cross-cluster migration of the data fragment to be transmitted to the unsaturated TSV transmission cluster.

7. The method according to claim 1, characterized in that, Adjusting the fragmentation granularity of subsequent data to be transmitted using the capacity-topology characteristic parameters, so that the size distribution of data fragments and the capacity saturation period of each through-silicon via (TSV) transmission cluster are mismatched in time, includes: The capacity saturation period and topological connectivity of each through-silicon via transmission cluster are extracted from the capacity-topology characteristic parameters. The capacity saturation period is projected onto the cyclic time axis to construct a capacity saturation rhythm diagram. Peak superposition regions and trough isolation regions are identified in the capacity saturation rhythm diagram. A piecewise granularity inverse control function is constructed. The piecewise granularity inverse control function uses the superposition intensity of the peak superposition region as input to generate a granularity suppression coefficient, uses the isolation spacing of the valley isolation region as input to generate a granularity release coefficient, and introduces the topological connectivity as a nonlinear modulation parameter into the calculation of the granularity suppression coefficient and the granularity release coefficient. The fragmentation granularity inverse control function is used to perform variable granularity segmentation on the subsequent data to be transmitted, generating fine-grained data fragments at the silicon via transmission clusters corresponding to the peak superposition region and coarse-grained data fragments at the silicon via transmission clusters corresponding to the valley isolation region. The fine-grained data fragments and the coarse-grained data fragments are allocated to each through-silicon via (TSV) transmission cluster, so that the size distribution of the data fragments and the capacity saturation period of each TSV transmission cluster are mismatched in time.

8. A system for improving inter-chip data transmission efficiency in 3D packaging structures, used to implement the method of any one of claims 1-7, characterized in that, include: The clustering calculation unit is used to obtain the spatial distribution matrix of the data to be transmitted and the through-silicon via array in the three-dimensional packaging structure, calculate the spatial distance between through-silicon vias and the interlayer transmission hops based on the spatial distribution matrix, and cluster the through-silicon vias according to the spatial distance and the interlayer transmission hops to generate multiple through-silicon via transmission clusters. The data fragmentation unit is used to divide the data to be transmitted into fragments according to the number of the through-silicon via (TSV) transmission clusters and the channel bandwidth capacity of the TSVs in each cluster, generate a data fragment set corresponding to the TSV transmission clusters, and label each data fragment with its corresponding TSV transmission cluster identifier. Based on the TSV transmission cluster identifier, each data fragment is allocated to the corresponding TSV transmission cluster, and a bidirectional state mapping relationship between the data fragments and the TSV transmission clusters is established during the transmission process. The cross-cluster migration unit is used to predict the capacity saturation time of each through-silicon via (TSV) transmission cluster based on the bidirectional state mapping relationship, and to migrate the data to be transmitted to the unsaturated TSV transmission clusters before the capacity saturation time is reached. The feature adjustment unit is used to correlate the capacity saturation time of each through-silicon via (TSV) transmission cluster with the spatial distance, generate the capacity-topology feature parameters of the TSV transmission cluster, and use the capacity-topology feature parameters to adjust the fragmentation granularity of subsequent data to be transmitted, so that the size distribution of the data fragments forms a time mismatch with the capacity saturation period of each TSV transmission cluster.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images