A network planning method for multi-grain integrated systems

By employing a two-layer structure of the original graph and the solution graph, along with a shrinking clustering pruning strategy, the limitations of modeling capabilities and the difficulty of solving problems in multi-die integrated systems are addressed. This enables efficient multi-layer network path modeling and convex block allocation, improving computational efficiency and parallel scalability.

CN123116993APending Publication Date: 2026-11-17FUZHOU LIXIN TECH CO LTD +1
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202611318028.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively handle global connectivity coupling in multi-die integrated systems, especially in scenarios involving vertical stacking of three or more dies and 2.5D interposer layers. Existing methods fail to address complex cross-layer connectivity issues simultaneously within a unified optimization model, resulting in low computational efficiency and limited modeling capabilities.

Method used

It adopts a two-layer structure of original graph-solution graph, and through edge shrinking and clustering pruning strategies, the size of the solution graph is compressed to 1.5%~2% of the original size. It has good parallel scalability, and the independent solution graph mechanism supports multi-core CPU, GPU and distributed parallel computing, realizing flexible modeling and accurate solution across multi-layer network paths.

Benefits of technology

It significantly improves computational efficiency, reducing solution time from hours to minutes, while ensuring solution quality. It also adapts to the modeling capabilities of complex multi-layer network paths and supports efficient convex block allocation for multi-die integrated systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116993A_ABST
    Figure CN123116993A_ABST
Patent Text Reader

Abstract

This invention proposes a network planning method for multi-die integrated systems, belonging to the field of flip-chip and 3D integrated packaging technology. The steps are as follows: input the physical layout information, bump positions, and network connection requirements between each die; construct the original graph; obtain a candidate path set; construct a solution graph for each candidate path; execute edge reduction and pruning strategies; execute clustering and coarsening pruning strategies; solve the simplified solution graph using a graph theory solver; output the network bump allocation results that satisfy each die. This invention provides a network planning solution for multi-die integrated systems that balances modeling flexibility and solution efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of flip chip and three-dimensional integrated packaging technology, specifically relating to a network planning method for multi-die integrated systems. Background Technology

[0002] In flip-chip and 3D integrated packaging technologies, bump assignment is a core step connecting chip I / O ports to the package substrate or interposer, and its result directly affects the routing capability and signal quality of subsequent wiring. Existing technologies can be divided into three stages according to their evolution path: early two-die bump assignment, extended methods in 2.5D scenarios, and research on 3D ICs and multi-die systems for multi-layer stacking.

[0003] I. Early Die-to-Die Bump Assignment

[0004] Early research focused on micro-bump mapping and single-layer RDL routing between two adjacent dies, resulting in relatively mature solutions. The paper "Efficient Micro-Bump Assignment for RDL Routing in 3D ICs" employs a two-stage heuristic algorithm combining initial matching with teardown and rerouting, while the paper "ILP-based Inter-Die Routing for 3D ICs" constructs an integer linear programming model to solve for optimal allocation. Subsequent work has jointly optimized bump assignment with I / O planning, such as multi-commodity flow network models (see "Flip-Chip Routing With I / O Planning Considering PracticalPad Assignment Constraints") and methods to guarantee routingability through area assignment (see "Area-I / ORDL routing for chip-package co-design considering regional assignment"). These methods all assume a fixed signal topology of "A-plane to B-plane" and only two physical layers of RDL, making direct extension to interconnection scenarios with three or more dies impossible.

[0005] II. 2.5D Scenes

[0006] With the widespread adoption of 2.5D interposers and chiplet architectures, multiple dies are placed side-by-side on the interposer, leading to a surge in signal networks and complex topologies. Early researchers modeled this problem as bipartite graph matching, transforming signal connections into optimal matching of weighted bipartite graphs. US patent "Two-chip co-design and co-optimization in three-dimensional integrated circuit net assignment (US8392870)" applies an extended Munkres algorithm to allocate interconnect paths between two chips; the paper "Multiple chip planning for chip-interposer codesign" extends the problem to multiple chips, using a hierarchical B*-tree for placement and then assigning microbumps using bipartite graph matching; US patent "Method of redistribution layer routing for 2.5-dimensional integrated circuit packages (US10558779)" uses minimum-weight maximum-cardinality bipartite graph matching for RDL routing; and the patent "Method for co-designing flip-chip and interposer" independently assigns power bumps to each chip before synthesizing and routing the interposer, employing a divide-and-conquer strategy.

[0007] In recent years, researchers have begun to examine I / O planning problems from a more macro-system perspective. Rossi et al., in their 2019 paper "I / O Planning and Cross-Hierarchical Optimization for Advanced Electronic 'Systems 4.0'", pointed out that in the increasingly complex and highly heterogeneous electronic systems required by Industry 4.0, I / O power consumption accounts for 50% of the total system power consumption, making I / O planning a key factor in system feasibility. This work proposed a cross-hierarchical I / O optimization algorithm framework for 2.5D and 3D-IC designs, aiming to address the design challenges of complex heterogeneous integrated systems through system-level I / O planning and cross-hierarchical optimization. The 2022 work published in ASP-DAC, "Signal-Integrity-Aware Interposer Bus Routing in 2.5D Heterogeneous Integration", extends the problem to signal integrity-aware interposer bus routing in 2.5D heterogeneous integration. By escaping the chip physical layer and sharing the bus topology, it achieves a compact interconnection of thousands of bump I / Os in a short time.

[0008] Although the above methods have expanded in terms of problem scale and optimization dimensions, they still essentially decompose the multi-die problem into several relatively independent sub-problems for solving, or perform post-processing of convex block allocation after the layout is completed. They fail to handle the global connectivity coupling of three or more dies simultaneously in a unified optimization model.

[0009] III. 3DIC Integrated System

[0010] In vertically stacked 3D ICs, the complexity of inter-layer signal allocation increases dramatically when the number of stacked layers exceeds two. The paper "Lagrangian Relaxation Based Inter-Layer Signal Via Assignment for 3-D ICs" models the inter-layer signal allocation of three or more dies as an integral minimum-cost multi-commodity flow problem, and uses the Lagrangian relaxation method to solve it. By decomposing it into a shortest path sub-problem, it gradually approaches the optimal solution and can handle signal allocation in multiple vertical directions simultaneously.

[0011] In recent years, with the penetration of machine learning technology into the EDA field, researchers have begun to explore data-driven bump assignment methods. Pin-Opt, in "Graph Representation Learning for Large-Scale Pin Assignment Optimization of Microbumps Considering Signal and Power Integrity," proposed a large-scale microbump pin assignment optimization framework based on graph representation learning. This framework addresses the increasing I / O requirements in 2.5D / 3D ICs by comprehensively considering both signal integrity and power integrity for optimization.

[0012] US Patent “US20190043808A1 - Interposer for an integrated system and corresponding design method” proposes an interposer for integrated systems and a corresponding design method, which achieves electrical interconnection of multi-chip systems by optimizing the arrangement of first and second connectors on the interposer. This patent considers the bump arrangement and interconnection optimization of multiple chips at the interposer level, and both it and this invention address the connection planning problem in multi-die integrated systems. However, this patent mainly focuses on the physical arrangement optimization of connectors on the interposer and does not involve the flexible graphing strategy and systematic graph size reduction method for cross-layer network paths described in this invention.

[0013] In terms of commercial EDA tools, the Cadence Integrity 3D-IC Platform provides the industry's first comprehensive 3D-IC design planning, implementation, and analysis platform (refer to "Cadence Integrity 3D-IC Platform Datasheet. Cadence Design Systems"). Built on the Innovus Implementation System, this platform supports the planning, implementation, and analysis of any type of stacked die system (2.5D or 3D). Regarding bump and TSV planning, the platform provides complete stack management, chip-to-package signal mapping, and advanced bump and TSV planning capabilities. However, the platform's core capabilities focus on system-level PPA (performance, power, area) optimization, thermal analysis, power analysis, inter-die static timing analysis, and physical verification. Its bump allocation function is embedded within the overall physical design flow, primarily targeting engineering implementations in adjacent die scenarios. It lacks publicly available algorithm models for simultaneously handling global connectivity coupling of three or more dies, and it does not address the flexible graphing strategies and targeted graph scaling methods described in this invention for adapting to complex multi-layer network paths. Summary of the Invention

[0014] The purpose of this invention is to propose a network planning method for multi-grain integrated systems. This method utilizes a two-layer structure of original graph and solution graph to achieve flexible specification and accurate modeling of any multi-layer network path, exhibiting good parallel scalability. Through edge reduction and clustering pruning strategies, the solution graph size is compressed to 1.5%~2% of the original size. In an instance containing 104 bumps, the computation time is reduced from several hours to minutes while maintaining solution quality. This provides a network planning solution for multi-grain integrated systems that balances modeling flexibility and solution efficiency.

[0015] To achieve the above objectives, the technical solution of the present invention is as follows:

[0016] A network planning method for multi-grain integrated systems, characterized by comprising the following steps:

[0017] Step S1: Obtain the physical layout information, bump location information, and network connection requirements between each die in the multi-die integrated system; wherein, each die includes several link regions, and each link region includes several bumps.

[0018] Step S2: Treat each linked region as a node in the original graph, and establish edges between the corresponding nodes according to the physical connectivity between the linked regions to obtain the original graph used to describe the global topology of the multi-grain integrated system.

[0019] Step S3: Perform path search on the original graph according to network connection requirements to generate several candidate paths. Each candidate path is recorded as a sequence of nodes arranged in order of the link regions passing through each grain in the direction of signal transmission.

[0020] Step S4: For each candidate path, construct an independent solution graph. The solution graph is constructed as follows: take all available bumps on all linked regions traversed by the candidate path as nodes in the solution graph, and establish directed edges between each available bump in the previous linked region and each available bump in the next linked region according to the order of adjacent linked regions in the node sequence, forming a full connection of available bumps between adjacent linked regions.

[0021] Step S5: Perform graph size reduction processing on each solution graph. The graph size reduction processing includes sequentially executed edge shrinking strategy and spatial clustering coarsening strategy:

[0022] Edge shrinking strategy: Count the number of available bump pairs that are physically aligned between adjacent linked regions on the candidate path. If the number of available bump pairs is greater than or equal to the number of networks that the current candidate path needs to support, then the adjacent linked regions are determined to be strongly correlated. The bump nodes corresponding to the strongly correlated adjacent linked regions in the solution graph are merged into a single node, and the fully connected edges between the merged adjacent linked regions are omitted.

[0023] Spatial clustering coarsening strategy: Based on the distribution density of available convexities in the linked regions, the available convexities in each linked region are adaptively divided into several blocks according to their spatial location. A representative node is generated for each block. A coarsened solution graph is constructed based on all representative nodes and an initial solution is performed. Based on the initial solution results, fully connected edges are generated only for convexity nodes between block pairs connected by the same network, and the remaining block pairs are pruned.

[0024] Step S6: Perform convex allocation on the solution graph after graph size reduction processing, and output the complete convex node sequence allocation result of each network from the source link region to the target link region.

[0025] Preferably, in the solution graph construction method of step S4:

[0026] The weight of each directed edge is determined based on the spatial distance between the two corresponding available bumps, and the capacity of each directed edge is set to 1.

[0027] Each node records the following attributes: its grain layer number, its link region number, available bump physical coordinates, and bump type.

[0028] Preferably, the complete physical alignment means:

[0029] In the scenario of vertically stacked 3D integrated circuits, the projection coordinates of the available bumps on the previous link area and the available bumps on the next link area completely coincide in the vertical direction.

[0030] Or, in a 2.5D intermediary layer scenario, the available bumps on the previous link region and the available bumps on the next link region have completely coincident coordinates on the intermediary layer plane.

[0031] Preferably, in the spatial clustering coarsening strategy, the number of blocks divided into each link region is positively correlated with the available bump distribution density of the link region.

[0032] Preferably, in the spatial clustering coarsening strategy, the number of blocks into which each link region is divided is calculated as follows:

[0033]

[0034] in, This represents the number of blocks the linked region is divided into during the cluster coarsening stage. For linking areas The number of available bumps on the surface For the area of ​​the linked region, An adjustable empirical coefficient. These are the preset lower and upper threshold values ​​for the number of cluster blocks, respectively.

[0035] Preferably, in the construction of the solution graph, each block is represented by the geometric centroid of all available convex blocks within the block.

[0036] Preferably, in the construction of the coarsened solution graph, directed edges are established between the representative nodes corresponding to adjacent linked regions according to the path direction, and weights and capacities are set to form a full connection between the representative nodes of adjacent linked regions; the edge weight is the spatial distance between the centroids of the blocks corresponding to the two representative nodes, and the edge capacity is set to the number of available convex blocks in the block.

[0037] Preferably, based on the initial solution results, fully connected edges are generated only between convex node pairs of blocks connected by the same network, as follows:

[0038] The activated block pairs are determined based on the initial solution results, and the convex node corresponding to the activated block pairs is inserted into the coarsened solution graph.

[0039] For an activated block pair, zero-weight edges are established between the representative node of the first block and each available convex node in the first block, zero-weight edges are established between each available convex node in the second block and the representative node of the second block, and fully connected edges are established between each available convex node in the first block and each available convex node in the second block. The weight of the fully connected edges is the spatial distance between the corresponding available convex nodes.

[0040] Preferably, the convex allocation solution process in step S6 is as follows: with the goal of minimizing the total path cost, and based on the network connection requirements, under the premise of satisfying the node capacity and edge capacity constraints, the global optimal mapping from the convex area of ​​the path's starting point region to the convex area of ​​the ending point region is completed.

[0041] Preferably, it also includes parallel computing steps:

[0042] When the sets of linked regions traversed by different candidate paths do not overlap at all, the solution graph construction, graph size reduction processing, and convex block allocation solution of the corresponding candidate paths are assigned to different computing threads or computing nodes for parallel execution.

[0043] When different candidate paths share the same link region, node capacity constraints are maintained independently in the solution graph of the corresponding candidate paths, and atomic locks are set to avoid resource contention.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] The network planning method for multi-die integrated systems provided by this invention demonstrates significant technological advancements in terms of computational efficiency, system architecture, and optimization performance.

[0046] (1) It realizes flexible specification and accurate modeling of complex network paths across multiple layers: The present invention uses a two-layer graph construction structure of original graph-solution graph. In the original graph, each die is divided into a link region and used as a node. In the solution graph, the candidate path is explicitly recorded as a node sequence and the corresponding convex fully connected edge is constructed, so that any complex network path across multiple layers can be accurately modeled, which greatly expands the applicability of the convex allocation method.

[0047] (2) Multi-level pruning strategy for solution graph scale

[0048] This invention proposes three progressive pruning methods based on graph construction. First, addressing the common design feature of perfectly aligned bumps between adjacent dies in 2.5D intermediary layers and vertically stacked 3D ICs, the number of available bump pairs with perfect alignment between adjacent linked regions is counted. Under strong correlation conditions, corresponding bump nodes are merged into a single supernode, reducing the number of edges from O(|B1|×|B2|) to O(1), achieving a scale reduction with zero precision loss. Second, addressing the problem of quadratic growth in the number of edges in high-density bump scenarios, the region is dynamically divided into 8 to 64 blocks using an adaptive formula based on the linked region size and bump density. After solving the initial solution on the coarsened graph, only the involved block pairs are activated for refined edge construction, compressing the number of edges from O(|B1|×|B2|) to O(∑|B_active|). 2This allows subsequent graph theory solutions to significantly reduce the search space while preserving the possibility of finding the optimal solution, enabling the present invention to compress the solution time from several hours to minutes while ensuring the quality of the solution.

[0049] (3) Independent solution graph mechanism supports parallel acceleration

[0050] This invention constructs a solution graph independently for each candidate path. When different paths do not share the same link region, there are no data dependencies or read / write conflicts between the solution graphs, making it naturally compatible with multi-core CPUs, GPUs, and distributed parallel computing. For cases where shared link regions exist, resource contention can be avoided and parallel acceleration capabilities can be preserved by independently maintaining node capacity constraints and setting atomic locks in each solution graph.

[0051] In summary, this invention outperforms existing technologies in terms of modeling flexibility, solution efficiency, adaptability, and parallel scalability, effectively solving the two core technical problems of "limited modeling capabilities" and "difficulty in solving large-scale instances" in convex block allocation of multi-die integrated systems. Attached Figure Description

[0052] Figure 1 This is a flowchart of the network planning method of the present invention;

[0053] Figure 2 This is a front view of an example of a multi-die integrated system according to the present invention.

[0054] Figure 3 The original diagram constructed for this invention;

[0055] Figure 4 The solution diagram constructed for this invention;

[0056] Figure 5 This is the solution diagram generated by the present invention based on the coarsening solution results. Detailed Implementation

[0057] The following is in conjunction with the appendix Figure 1-5 The technical solution of the present invention will be described in detail below.

[0058] This invention proposes a network planning method for multi-die integrated systems. It achieves rapid solution for convex block allocation within the internal network of a multi-die integrated system through the organic combination of two key strategies: first, a graph construction method adapted to advanced processes; and second, a graph pruning strategy tailored to practical design scenarios. The specific technical solution is as follows:

[0059] 1. Mapping methods for multi-die integrated systems and their internal networks

[0060] The graph construction method proposed in this invention uses a directed graph as the core data structure, which includes two levels: the Original Graph and the Solution Graph. These are used to describe the global topology of the system and the allocatable resources under a specific path, respectively.

[0061] 1.1 Original Image

[0062] The original graph is used to globally model the link regions and interconnections of each die in a multi-die integrated system. Unlike traditional methods that treat each die as a multi-layered structure of identical size with completely overlapping projections, this invention, to adapt to more flexible stacking and linking methods under advanced processes, divides each die into several link regions (IO Regions) based on its actual linking state. Each link region serves as a node in the original graph. For a system containing n link regions, the node sequence of the original graph is denoted as [R0, R1, …, R…]. n-1 In the original graph, the edges between nodes represent the possibility of physical connectivity between two linked regions (such as vertical vias in adjacent dies or horizontal routing resources on intermediary layers). However, the original graph does not participate in the specific convex allocation solution; its role is to provide topological constraints and reachability judgment criteria for the generation of subsequent candidate paths.

[0063] 1.2 Solution Diagram

[0064] The solution graph is a computational structure constructed separately for each path, oriented towards convex allocation. The solution process begins with the network connectivity requirements obtained by the user through the overall design of the multi-die integrated system. These requirements specify a sequence of intermediate dies that signals need to traverse from the source die to the target die. This invention transforms these requirements into a path set [P0, P1, …, P…]. m-1 ], each path P i Mark a link region on each die it traverses, and follow the network-driven direction with a node sequence [R]. i1 , R i2 , R i3 , …, R ik Recorded in the form of ].

[0065] For each candidate path P i This invention constructs an independent solution graph for it. In the solution graph, the node is P. i All available bumps on all traversed link regions are recorded. Each node needs to record its Die layer number, physical coordinates, bump type (signal / power / ground), and whether it is currently occupied. The edges of the solution graph are constructed sequentially according to the path node sequence: for any two adjacent link regions (R...ij , R ij+1 In this invention, a directed edge is established between each available convex block in the preceding region and each available convex block in the following region in the solution graph, forming a full connection of convex blocks between the two layers of regions. The weight of each edge is defined as the spatial distance between two convex blocks, and the capacity of each edge is set to 1, indicating that the connection can be assigned to at most one network. Starting from any convex block in the path starting region in the solution graph, the network proceeds layer by layer along the path direction, passing through the convex blocks in the intermediate link regions, and finally reaching any convex block in the ending region, thus forming a complete convex network planning scheme that perfectly corresponds to the candidate path. Its total cost is the sum of the weights of all edges on the path.

[0066] 2. Graph pruning strategies for actual design scenarios

[0067] Based on the construction of the original graph and the solution graph, this invention further proposes two graph pruning strategies for practical design scenarios, aiming to effectively reduce the graph size, reduce the computational complexity of subsequent graph theory solutions, and keep the solution accuracy unaffected.

[0068] 2.1 Edge contraction strategy based on highly correlated adjacent dies

[0069] In 2.5D interposer architectures and vertically stacked 3D integrated circuits, designers often place perfectly aligned bumps (i.e., the projected coordinates of the bumps on the interposer or in the vertical direction completely coincide) at corresponding positions on adjacent dies to reduce inter-die communication latency and simplify routing complexity. This invention fully utilizes this design feature and proposes an edge-shrinking strategy based on the correlation of adjacent link regions. Specifically, for a candidate path P... i This invention statistically analyzes the adjacent link regions R on this path. ij With R ij+1 All available bump pairs are identified, and the number of physically aligned bump pairs is determined. Simultaneously, the number of networks that the current path needs to support is calculated. If the number of physically aligned bump pairs is greater than or equal to the number of networks in the current path, then these two linked regions are considered to be on path P. i The two linked regions in the graph have a strong correlation, so the corresponding convex nodes of these two linked regions are merged into a single node. This edge reduction operation can reduce the number of edges between adjacent layers from O(|B1|×|B2|) to O(1) without losing the accuracy of the solution, which effectively reduces the size of the graph.

[0070] 2.2 Spatial Clustering-Based Solution Graph Coarsening Strategy

[0071] To address the performance challenges posed by high-density, large-scale bumps in advanced processes—where the number of bumps in a single linked region can reach thousands or even tens of thousands, and the number of edges increases quadratically if adjacent regions' bumps are fully connected—this invention proposes a parameter-adaptive spatial clustering pruning strategy. The specific process is as follows: First, based on the geometric dimensions and bump distribution density of the linked region, the system adaptively divides each linked region into 8 to 64 bins of similar size. Second, in the coarsened solution graph, the centroids of the geometric center points of all available bumps within a bin are used as representative nodes for initial solution calculation. Finally, based on the coarsening solution's indications, fully connected edges are generated only between bump nodes in pairs of blocks connected by the same network; pairs of blocks not involved in the coarsening solution are directly pruned. This strategy achieves an adaptive balance between solution accuracy and computational efficiency, effectively improving the practicality and scalability of this invention in large-scale multi-die systems.

[0072] The overall flow of the multi-die integrated system network planning method of the present invention is as follows: Figure 1 As shown, the main steps include: Step S101, input each die The steps are as follows: Step S102: Construct the original graph; Step S103: Determine whether to trigger the pruning strategy based on the BUMP size; if the BUMP size is less than the preset empirical value, then execute steps S104 and S108 in sequence; otherwise, execute steps S105-S108 in sequence; Step S104: Construct a solution graph for each candidate path; Step S105: Construct a solution graph for each candidate path (the screenshot construction method is the same as in step S104), and execute the edge shrinking pruning strategy; Step S106: Execute the clustering coarsening pruning strategy; Step S107: Construct a simplified solution graph based on the coarsening solution graph result; Step S108: Perform bump allocation on the solution graph obtained in step S104 or step S107 in the graph theory solver; Step S109: Output the network bump allocation result that satisfies each grain. The following is combined with... Figures 2 to 5 The specific examples shown illustrate each step in detail.

[0073] like Figure 2 As shown, this embodiment uses a multi-die integrated system containing three dies: Chip1, Chip2, and Chip3 as an example for illustration. Chip1 contains two link regions R0 and R5, Chip2 contains two link regions R1 and R2, and Chip3 contains two link regions R3 and R4. The arrows in the figure indicate the signal transmission direction of interest in this embodiment: starting from R0, passing through R1, R2, R3, and R4, and finally reaching R5, forming the path P: R0→R1→R2→R3→R4→R5. In this example, the number of available bumps in each link region is set to |B(R i)| = 4000, and the number of networks carried by path P is N_net = 1500.

[0074] For ease of subsequent description, the following symbols are defined:

[0075] B(R i ): Linked region R i The complete set of available bumps on

[0076] |B(R i )|:Linked region R i Number of available bumps on

[0077] N_net: The number of networks that the current path needs to support.

[0078] K: The number of blocks each linked region is divided into during the cluster coarsening stage.

[0079] w(u, v): Spatial distance weight between convex blocks u and v

[0080] (1) Construction of the original graph (S101~S102)

[0081] like Figure 3 As shown, according to the graph construction method of the present invention, the original graph G_original = (V_original, E_original) is first constructed. The six linked regions R0 to R5 in the system are each designated as a node in the original graph, and the node sequence is as follows:

[0082]

[0083] In the original graph, edge E_original represents the possibility of physical connectivity between adjacent linked regions. In this embodiment, based on the chip physical layout, there are physical connectivity conditions between R0 and R1, R1 and R2, R2 and R3, R3 and R4, and R4 and R5. Therefore, directed edges are established between these node pairs in graph G_original, as follows: Figure 3 The black directed edge is shown in the image:

[0084]

[0085] The original graph does not directly participate in the convex block assignment solution. Its role is to provide physical topological constraints and connectivity judgment basis for subsequent candidate path generation and pruning strategies.

[0086] (2) Solving graph generation (S104)

[0087] Based on the system's network connectivity requirements, the designers specified that the signal originates from region R0 of Chip1, passes through regions R1 and R2 of Chip2, regions R3 and R4 of Chip3, and finally returns to region R5 of Chip1. According to the original diagram, this path is confirmed to be valid. Figure 3 The directed edge in red is shown in the middle. In this embodiment, the candidate path P0 corresponding to this requirement is recorded as a node sequence:

[0088]

[0089] For candidate path P0, this invention constructs a dedicated solution graph G_solution = (V_solution, E_solution). The nodes of the solution graph are all available bumps on each link region traversed by path P0, i.e.:

[0090]

[0091] Let |B(R) i If | = 4000, then the total number of nodes in the graph is N.

[0092]

[0093] Each node records the following attributes: its Die layer number, its linked region number, the physical coordinates of the bump, and the bump type.

[0094] The edges of the solution graph are constructed in the order of the path node sequence. For any two adjacent linked regions (R) in the sequence... i , R i+1 In the solution diagram, B(R) i Each bump in ) and B(R) i+1 Directed edges are established between each convexity in the graph, forming a fully connected bipartite graph, such as... Figure 4 As shown. Taking B(R0) and B(R1) as examples, the number of edges is |E(R0, R1)|:

[0095]

[0096] For 5 pairs of adjacent regions in the path, the total number of edges in the solution graph is |E_solution|:

[0097]

[0098] Each directed edge e is represented as:

[0099]

[0100] The weight function is defined as the spatial distance between two convex blocks.

[0101]

[0102] The capacity of each edge is set to 1, meaning that the edge can carry the signal transmission of at most one network.

[0103] After the graph is constructed, the convex allocation problem can be described as finding the N_net paths with the minimum total weight on the graph.

[0104] (3) Edge shrinking strategy based on highly correlated adjacent dies (S105)

[0105] In this embodiment, the designer arranged perfectly aligned bump pairs at corresponding positions in the R1 and R2 regions of Chip2, that is, there exist bumps u∈B(R1) and v∈B(R2), such that the two are aligned in the vertical direction (vertical stacking scene) or completely overlap in the intermediate layer plane (2.5D scene).

[0106] In the edge-shrinking strategy, this invention systematically counts the number of available bump pairs whose physical positions completely overlap between B(R1) and B(R2). Specifically, for each bump u in B(R1), all bumps v in B(R2) are traversed. If the center positions of u and v are exactly the same, then the bump pair is determined to be completely overlapping. In this embodiment, the bumps of B(R1) and B(R2) are arranged in a completely regular grid alignment, so the number of completely aligned bump pairs is counted as 4000 pairs (i.e., min(|B(R1)|, |B(R2)|) = 4000).

[0107] The current candidate path P0 needs to support N_net = 1500 networks. Since the number of aligned convex block pairs (4000 ≥ N_net = 1500) is greater than or equal to 4000, R1 and R2 are determined to be strongly correlated on path P0. Based on this, the corresponding convex block node pairs that originally belonged to R1 and R2 in the solution graph can be merged into a single supernode R. 12 After merging, the fully connected edges that originally needed to be established between R1 and R2 are completely omitted, and the path P0 becomes:

[0108]

[0109] The number of edges is |E(R1, R2)| = |B(R1)| × |B(R2)| = 1.6 × 10 7 Reduced to 0 (no need to build edges, R1 and R2 in the original graph have become the same node). This operation results in zero loss of accuracy in subsequent solutions because when the number of aligned bumps is greater than or equal to the number of networks, the optimal allocation of R1→R2 must be from a bump on R1 to the corresponding bump on R2 that completely overlaps with it.

[0110] (4) Spatial clustering-based graph coarsening strategy (S106~S107)

[0111] In this embodiment, each link region contains 4000 available bumps, which is a high-density scenario. The system calculates the number of cluster blocks K according to an adaptive formula. The design goal of the number of cluster blocks is to compress the size of the solution graph to the greatest extent while maintaining sufficient spatial resolution to preserve solution accuracy. The formula for calculating the number of cluster blocks used in this embodiment is:

[0112]

[0113] Where, the number of bumps = |B(R i The area of ​​the linked region is Area (unit: mm). 2 α is an adjustable empirical coefficient (default 1.0). The `clamp` function limits the calculation results to the range [8, 64] to ensure the clustering granularity is within a reasonable range. The formula is designed based on the following: the number of cluster blocks is positively correlated with the number of bumps—more bumps require more detailed spatial partitioning to retain sufficient resolution; it is negatively correlated with the area of ​​linked regions and has a weak correlation—the larger the area, the sparser the bump distribution, allowing for a reduction in the number of blocks; the ratio of the number of bumps to the area essentially reflects the bump density. When the bump density is high, the formula automatically generates a larger K value (tending towards 64) to retain sufficient spatial resolution; when the density is low, the K value tends towards 8 to avoid the additional computational overhead caused by overly fine partitioning. The coefficient α allows designers to fine-tune the clustering granularity according to specific process scenarios.

[0114] In this embodiment, Area = 20mm 2 ,|B(R i Given || = 4000, α = 1.0, we can calculate:

[0115]

[0116] For ease of explanation, this embodiment uses K = 20 (equivalent to α ≈ 1.41), meaning each link region is divided into 20 blocks.

[0117] Based on the partitioning results, a clustering algorithm is used to merge the available bumps on each linked region according to their spatial location, forming K clusters C. ij (j = 1, 2, …, K), each cluster contains approximately |B(R) i K = 200 convex blocks. The representative node of each cluster is the centroid of the coordinates of all convex blocks within the cluster.

[0118]

[0119] The construction process of the coarsened solution graph G_coarse is as follows: The representative node of each cluster is used as a node in the coarsened graph, resulting in K = 20 nodes per layer. After edge shrinking and merging, R1 and R2 are merged into R... 12 There are 5 floors in total (R0, R...) 12 R0, R3, R4, R5), totaling 100 representative nodes. On the coarsened graph, according to the path P0' = [R0, R...]... 12 The order is [R3, R4, R5], and edges are established between adjacent layer representative nodes using a fully connected method. Representative node C ij With C i+1,1 The edge weight is the spatial distance between the two cluster centroids, and the edge capacity is set to the number of convexities within the cluster (200).

[0120] The system runs a graph theory algorithm on the coarsened graph to obtain an initial solution, which indicates the block number sequence traversed by the N_net = 1500 network at each layer. Since each path activates a block pair between adjacent layers, a total of 5 layers generate 4 pairs of adjacent layers (R0→R...). 12 R 12 →R3, R3→R4, R4→R5), the total number of block pairs activated in the initial solution is:

[0121]

[0122] Subsequently, based on the block pair activation information indicated by the coarse resolution, the system performs fine-grained edge construction only on the convex nodes between block pairs actually involved in the network. For the activated representative node pairs (C... ij C i+1,1 The system inserts the corresponding set of fine nodes B(C) into the solution graph. ij ) and B(C i+1,1 Each node contains approximately 200 convex nodes. The refined edge construction process is as follows: Representing node C... ij With B(C) ij Each node in B(C) is connected to an edge with a weight of 0; i+1,1 Each node in ) is associated with a representative node C. i+1,1 Connect B(C) with an edge of weight 0. ij ) and B(C i+1,1 Fully connected edges are established between nodes in the convex region, and the weight of each edge is the spatial distance w(u, v) between the corresponding convex regions, such as... Figure 5 As shown.

[0123] In this embodiment, the total number of fine edges that need to be established between the 32 active block pairs is:

[0124]

[0125] Compare the original solution plot (8.0 × 10) 7 After pruning, the number of edges is reduced to 1.5% of the original size, resulting in a significant reduction in the size of the solution graph.

[0126] (5) Graph theory solution and result output (S108)

[0127] On the solution graph with completed pruning optimization output in step S107 or the screenshot obtained in step S104, this invention uses mainstream graph theory algorithms to solve for convex allocation in the simplified graph. This solution process aims to minimize the total path cost, and under the premise of satisfying node capacity and edge capacity constraints, completes the globally optimal mapping from the convex region of the path's starting point region to the convex region of the ending point region. With the solution graph size significantly reduced after pruning, the graph theory solver can complete the calculation in a relatively short time: taking this embodiment as an example, the original solution graph contains approximately 24,000 nodes and 8.0 × 10⁻⁶ nodes. 7 The estimated time for directly solving this problem is several hours; after two-stage pruning involving edge reduction and clustering, the number of nodes in the solved graph remains unchanged (approximately 24,000 nodes), while the number of edges is reduced to approximately 1.2 × 10⁻⁶. 6 The actual solution time is approximately 1 minute.

[0128] After the solution is completed, the system generates a complete node sequence for each network on path P0, starting from a convex node in region R0, passing through each layer of convex nodes, and finally reaching a convex node in region R5. The start and end points of this node sequence are recorded as the source and target convex node allocation results for the network, respectively, while the intermediate convex nodes are recorded as the relay convex node allocation results on the corresponding dies. The system organizes these allocation results according to the network identifier corresponding to path P0 to form a complete convex allocation scheme for the path, and uses this as the final output.

[0129] (6) Parallel computing scalability

[0130] The graph construction and pruning method proposed in this invention naturally possesses excellent multi-threaded parallel computing scalability. In the original graph construction stage, the system decomposes the overall network planning problem into several candidate paths P0, P1, …, P based on the physical layout and signal connection requirements of different dies. n-1 For any two candidate paths P i and P jIf the sets of linked regions they traverse are completely non-overlapping (i.e., there is no shared linked region R), then the sets of nodes and edges involved in the construction and pruning of the solution graph are independent of each other, with no data dependencies or read / write conflicts, and can be allocated to different computing threads or computing nodes for fully parallel execution. Even when different paths share the same linked region, this invention can still maintain a high parallel speedup ratio by independently maintaining node capacity constraints in the solution graph (i.e., each bump can only be occupied by one network) and setting atomic locks to avoid resource contention. The fundamental source of this parallel scalability lies in the solution architecture of this invention, which decomposes the global network planning problem into several path-level independent sub-problems. Unlike the traditional ILP method, which incorporates all networks and bumps into a single mathematical model for unified solution and is difficult to parallelize, the independent construction mechanism of the solution graph in this invention enables the overall running time to be approximately reduced to the solution time of a single path multiplied by the total number of paths and then divided by the number of parallel computing units in a large-scale multi-die system when the number of candidate paths is large and distributed in different die regions. The speedup effect becomes more and more significant as the system scale increases.

[0131] In summary, the convex allocation method for multi-die integrated systems proposed in this invention significantly improves solution efficiency while ensuring solution quality through the organic combination of flexible graph construction, multi-level pruning, and graph theory solving. It can effectively support network planning and convex allocation for complex heterogeneous integrated systems containing three or more dies. Those skilled in the art can make adaptive adjustments based on the above specific embodiments and the actual needs of different application scenarios; these adjustments do not depart from the protection scope of this invention.

[0132] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A network planning method for a multi-chip integrated system, characterized by, Includes the following steps: Step S1: Obtain the physical layout information, bump location information, and network connection requirements between each die in the multi-die integrated system; wherein, each die includes several link regions, and each link region includes several bumps. Step S2: Treat each linked region as a node in the original graph, and establish edges between the corresponding nodes according to the physical connectivity between the linked regions to obtain the original graph used to describe the global topology of the multi-grain integrated system. Step S3: Perform path search on the original graph according to network connection requirements to generate several candidate paths. Each candidate path is recorded as a sequence of nodes arranged in order of the link regions passing through each grain in the direction of signal transmission. Step S4: For each candidate path, construct an independent solution graph. The solution graph is constructed as follows: take all available bumps on all linked regions traversed by the candidate path as nodes in the solution graph, and establish directed edges between each available bump in the previous linked region and each available bump in the next linked region according to the order of adjacent linked regions in the node sequence, forming a full connection of available bumps between adjacent linked regions. Step S5: Perform graph size reduction processing on each solution graph. The graph size reduction processing includes sequentially executed edge shrinking strategy and spatial clustering coarsening strategy: Edge shrinking strategy: Count the number of available bump pairs that are physically aligned between adjacent linked regions on the candidate path. If the number of available bump pairs is greater than or equal to the number of networks that the current candidate path needs to support, then the adjacent linked regions are determined to be strongly correlated. The bump nodes corresponding to the strongly correlated adjacent linked regions in the solution graph are merged into a single node, and the fully connected edges between the merged adjacent linked regions are omitted. Spatial clustering coarsening strategy: Based on the distribution density of available convexities in the linked regions, the available convexities in each linked region are adaptively divided into several blocks according to their spatial location. A representative node is generated for each block. A coarsened solution graph is constructed based on all representative nodes and an initial solution is performed. Based on the initial solution results, fully connected edges are generated only for convexity nodes between block pairs connected by the same network, and the remaining block pairs are pruned. Step S6: Perform convex allocation on the solution graph after graph size reduction processing, and output the complete convex node sequence allocation result of each network from the source link region to the target link region.

2. The network planning method for a multi-chip integrated system according to claim 1, wherein, In the solution graph construction method of step S4: The weight of each directed edge is determined based on the spatial distance between the two corresponding available bumps, and the capacity of each directed edge is set to 1. Each node records the following attributes: its grain layer number, its link region number, available bump physical coordinates, and bump type.

3. The network planning method for a multi-chip integrated system according to claim 1, wherein, The complete physical alignment means: In the scenario of vertically stacked 3D integrated circuits, the projection coordinates of the available bumps on the previous link area and the available bumps on the next link area completely coincide in the vertical direction. Or, in a 2.5D intermediary layer scenario, the available bumps on the previous link region and the available bumps on the next link region have completely coincident coordinates on the intermediary layer plane.

4. The network planning method for a multi-chip integrated system according to claim 1, wherein, In the spatial clustering coarsening strategy, the number of blocks divided into each link region is positively correlated with the available bump distribution density of the link region.

5. The network planning method for a multi-chip integrated system according to claim 4, wherein, In the spatial clustering coarsening strategy, the number of blocks into which each linked region is divided is calculated as follows: wherein, is the number of blocks in which the linking area is divided in the clustering coarsening phase, is the number of available bumps on the linking area is the area of the linking area, is the area of the linking area, is an adjustable empirical coefficient, are respectively a preset lower threshold value and an upper threshold value of the number of clustering blocks.

6. The network planning method for a multi-chip integrated system according to claim 1, wherein, In the construction of the solution graph, each block is represented by the geometric centroid of all available convex blocks within the block.

7. The network planning method for a multi-chip integrated system according to claim 1, wherein In the construction of the coarsened solution graph, directed edges are established between the representative nodes corresponding to adjacent linked regions according to the path direction, and weights and capacities are set to form a full connection between the representative nodes of adjacent linked regions. The edge weight is the spatial distance between the centroids of the blocks corresponding to the two representative nodes, and the edge capacity is set to the number of available bumps within the block.

8. The network planning method for a multi-chip integrated system according to claim 7, wherein, Based on the initial solution results, fully connected edges are generated only between convex node pairs of blocks connected by the same network, as follows: The activated block pairs are determined based on the initial solution results, and the convex node corresponding to the activated block pairs is inserted into the coarsened solution graph. For an activated block pair, zero-weight edges are established between the representative node of the first block and each available convex node in the first block, zero-weight edges are established between each available convex node in the second block and the representative node of the second block, and fully connected edges are established between each available convex node in the first block and each available convex node in the second block. The weight of the fully connected edges is the spatial distance between the corresponding available convex nodes.

9. The network planning method for a multi-chip integrated system according to claim 1, wherein, The process of solving the convex allocation in step S6 is as follows: with the goal of minimizing the total path cost, and based on the network connection requirements, under the premise of satisfying the node capacity and edge capacity constraints, the global optimal mapping from the convex area of ​​the path's starting point region to the convex area of ​​the ending point region is completed.

10. The network planning method for a multi-chip integrated system according to claim 1, wherein, It also includes parallel computing steps: When the sets of linked regions traversed by different candidate paths do not overlap at all, the solution graph construction, graph size reduction processing, and convex block allocation solution of the corresponding candidate paths are assigned to different computing threads or computing nodes for parallel execution. When different candidate paths share the same link region, node capacity constraints are maintained independently in the solution graph of the corresponding candidate paths, and atomic locks are set to avoid resource contention.

Citation Information

Patent Citations

  • Method of redistribution layer routing for 2.5-dimensional integrated circuit packages

    US10558779B2

  • Interposer for an integrated system and corresponding design method

    US20190043808A1

  • Two-chip co-design and co-optimization in three-dimensional integrated circuit net assignment

    US8392870B2