Generating integrated circuit placements using multimodal neural networks

The multimodal neural network addresses the inefficiencies of conventional chip design by generating suggested regions and using in-context reinforcement learning to optimize chip placement, resulting in high-quality designs with reduced power consumption and improved performance metrics.

WO2026241753A1PCT designated stage Publication Date: 2026-11-19GDM HOLDING LLC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/028660
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-16
Filing Date
2026-05-18
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Conventional floorplanning methods for integrated circuit chip design require significant human intervention and consume excessive computational resources, often failing to generate high-quality floorplans due to the vast space of potential node placement combinations, and struggle with limited training data and poor generalization to unseen data.

Method used

A multimodal neural network is used to provide spatial and visual reasoning capabilities, generating suggested regions for chip placement and constraining the action space of a policy model, enabling high-quality floorplans without extensive training or optimization algorithms, and incorporating in-context reinforcement learning for iterative improvement.

Benefits of technology

The system generates integrated circuit chips with reduced power consumption, increased computing power, and improved performance metrics such as wire length and congestion, while being flexible across different grid sizes and requiring minimal computational resources.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026028660_19112026_PF_FP_ABST
    Figure US2026028660_19112026_PF_FP_ABST
Patent Text Reader

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating an integrated circuit chip placement. One of the methods includes obtaining netlist data for an integrated circuit chip; and generating a set of one or more integrated circuit chip placements, comprising: generating a respective set of one or more initial integrated circuit chip placements at each of a sequence of iterations, comprising, at each of one or more particular iterations in the sequence: determining, using a multimodal neural network, one or more respective sets of suggested regions on the surface of the integrated circuit chip; generating, for each set of suggested regions and using a policy model, one or more respective initial integrated circuit chip placements; and determining, for each respective initial integrated circuit chip placement, one or more performance metrics that each measure a quality of the initial integrated circuit chip placement.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Attorney Docket No.: 45288-0612WO1

[0002] GENERATING INTEGRATED CIRCUIT PLACEMENTS USING MULTIMODAL NEURAL NETWORKS

[0003] CROSS REFERENCE TO RELATED APPLICATION

[0004] This application claims priority' to U.S. Provisional Application No. 63 / 807,466, filed on May 16. 2025. The disclosure of the prior application is considered part of and is incorporated by reference in the disclosure of this application.

[0005] BACKGROUND

[0006] This specification relates to using neural networks for electronic design automation and, more specifically, for generating a computer chip placement.

[0007] Computer chip placements are schematic representations of the placement of some or all of the circuits of a computer chip on the surface, i.e., the chip area, of the computer chip.

[0008] Neural networks are machine learning models that employ one or more layers of nonlinear units to predict an output for a received input. Some neural networks include one or more hidden layers in addition to an output layer. The output of each hidden layer is used as input to the next layer in the network, i.e., the next hidden layer or the output layer. Each layer of the network generates an output from received inputs in accordance with current values of a respective set of parameters.

[0009] SUMMARY

[0010] This specification describes a system implemented as computer programs on one or more computers in one or more locations that generates a chip placement for an integrated circuit. The integrated circuit for which the chip placement is being generated will be referred to in this specification as a ‘’computer chip’’ but should generally be understood to mean any collection of electronic circuits that are fabricated on one piece of semiconductor material. References herein to “a chip” refer to an integrated circuit chip. The chip placement can be generated for use in manufacturing an integrated circuit chip.

[0011] Particular embodiments of the subject matter described in this specification can be implemented so as to realize one or more of the following advantages.

[0012] Floorplanning, which involves placing the components of a chip on the surface of the chip, is a crucial step in the chip design process. The placement of the components should optimize metrics such as area, total wire length and congestion. If a floorplan, also referred to as a placement, does not perform well on these metrics, the integrated circuit chip that is generated based on the floor plan will perform poorly. For example, the integrated circuitAttorney Docket No.: 45288-0612WO1

[0013] chip could fail to function, could consume an excessive amount of power, could have an unacceptable latency, or have any of a variety of other undesirable properties that are caused by sub-optimal placement of components on the chip.

[0014] Some conventional floorplanning solutions employ manual processes that require heavy7human involvement. Because of the enormous space of potential node placement combinations, conventional automated approaches have been unable to reliably generate high-quality floorplans without consuming an excessive amount of computational power and wall clock time, requiring human expert involvement, or both. By using a multimodal neural network to provide guidance to a policy model, the system described in this specification can quickly generate a high-quality7floorplan.

[0015] Furthermore, an integrated circuit chip which is produced using the system described in this specification may have reduced power consumption compared to one produced by a conventional method. It may also have increased computing power for a given surface area, or from another point of view be produced using fewer resources for a given amount of computing power. For example, the system can process a neural network input that includes an instruction to optimize performance metrics such as wire length and congestion, using the multimodal neural network to generate suggested regions. The generation of an integrated circuit chip placement to optimize performance metrics such as wire length, routing congestion and heat generation in the manner described herein provides a functional technical purpose in addressing specific technical constraints inherent in semiconductor physics and electronic design automation. The optimization is motivated at least in part by technical considerations relating to the physical properties of the integrated circuit, such as reducing power consumption and minimizing latency of the fabricated chip.

[0016] Some conventional systems for generating placements using a learned model require a large amount of training data that includes high-quality placements. Such training data is limited or infeasible to obtain. In addition, because some of these conventional systems are trained to map abstract representations to placements, these conventional systems may not generalize well at inference to unseen data.

[0017] The system described in this specification addresses this issue by generating suggested regions for placement using a multimodal neural network. The suggested regions can constrain the action space of a placement policy model, steering the placement process without requiring further training of the policy model or the multimodal neural netw ork. For example, the system can mask the policy’s output logits for each macro outside of theAttorney Docket No.: 45288-0612WO1

[0018] suggested region for the macro before sampling, constraining the policy model to place the macro within the area identified as promising by the multimodal neural network.

[0019] The system can incorporate spatial and visual reasoning capabilities of the multimodal neural network into the floorplanning process. For example, the system can include data representing previously generated placements in the neural network input for the multimodal neural network. The data representing a placement can include an image. In some examples, the image can depict each macro in a corresponding color. The system can use a colorcoding strategy’ to convey functional relationships between macros to the multimodal neural network.

[0020] In some implementations, the system described in this specification can generate high-quality chip floorplans at inference using in-context reinforcement learning. For example, at each of one or more iterations, the system can include data representing past suggestions, resulting placements, and performance metrics for resulting placements in the neural network input for the multimodal neural network. The multimodal neural network can thus iteratively suggest better regions by conditioning on at least previously generated placements, without requiring explicit optimization algorithms or gradient updates to the parameters of the multimodal neural network. Thus the system can use the adaptive guidance generated using the multimodal neural network to progressively improve the performance of the policy model and the quality of the generated floorplans over multiple iterations.

[0021] In addition, the system can use any appropriate policy model to generate the placements. The system is thus flexible, as the interface between the multimodal neural network and the policy model includes using suggested regions to constrain the action space of the policy model. In addition, the system can use the multimodal neural network to generate suggested regions for any appropriate grid size, e.g., the same grid size as the policy’ model is configured to operate at, or a different grid size.

[0022] The details of one or more embodiments of the subject matter described in this specification are set forth in the accompanying drawings and the description below. Other features, aspects, and advantages of the subject matter will become apparent from the description, the drawings, and the claims.

[0023] BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 shows an example placement generation system.Attorney Docket No.: 45288-0612WO1

[0024] FIG. 2 is a flow diagram of an example process for generating a set of integrated circuit chip placements.

[0025] FIG. 3 is a flow diagram of an example process for determining one or more respective sets of suggested regions on the surface of an integrated circuit chip.

[0026] FIG. 4 is a flow diagram of an example process for selecting one or more initial integrated circuit chip placements.

[0027] FIG. 5 is a flow diagram of an example process for generating one or more initial integrated circuit chip placements.

[0028] FIG. 6 shows another example process for generating a set of integrated circuit chip placements.

[0029] FIG. 7 shows the performance of an example chip placement system.

[0030] FIGS. 8A-8D show an example of a neural network input provided to a multimodal neural network.

[0031] Like reference numbers and designations in the various drawings indicate like elements.

[0032] DETAILED DESCRIPTION FIG. 1 shows an example placement generation system 100. The placement generation system 100 is an example of a system implemented as computer programs on one or more computers in one or more locations in which the systems, components, and techniques described below are implemented.

[0033] The system 100 receives netlist data 102 for a computer chip, i.e.. a very large-scale integration (VLSI) chip, that is to be manufactured and that includes a plurality of integrated circuit components, e.g., memory components and components that implement logic functions transistors, resistors, capacitors, and so on. The plurality of integrated circuit components may be different depending on the desired function of the chip. For example, the chip can be a special-purpose chip, i.e., an application-specific integrated circuit (ASIC), for machine learning computations, video processing, cryptography, or another computeintensive function. In some cases, the computer chip can be a portion of a larger computer chip, e.g., a portion of an ASIC that includes a certain subset of the components of the ASIC.

[0034] The netlist data 102 is data describing the connectivity’ of the integrated circuit components of the integrated circuit chip. In particular, the netlist data specifies a connectivity on the integrated circuit chip among a plurality' of nodes that each correspond toAttorney Docket No.: 45288-0612WO1

[0035] one or more of a plurality of integrated circuit components of the integrated circuit chip. That is. each node corresponds to a respective proper subset of the integrated circuit components, and those subsets do not overlap. In other words, the netlist data identifies, for each of the plurality of nodes, which other nodes (if any) the node needs to be connected to by one or more wires in the manufactured integrated circuit chip. In some cases, the integrated circuit components have already been clustered in clusters, e.g., by an external system or by using an existing clustering technique, and each node in the netlist data represents a different one of the clusters.

[0036] The integrated circuit components can include macro components. The nodes can include macro nodes that represent macro components. In some examples, the integrated circuit components can include standard cell components. The nodes can include standard cell nodes that represent standard cell components.

[0037] As one example, the netlist data 102 can identity' two types of nodes: nodes that represent macro components and nodes that represent standard cell components.

[0038] Macro components are large blocks of IC components, e.g., static random-access memory (SRAM) or other memory blocks, that are represented as a single node in the netlist. For example, the nodes representing macro components can include nodes that each represent a corresponding instance of an SRAM. As another example, the nodes representing macro components can include hard macros that are made up of a fixed number of standard cells, e.g., a macro that is made up of a fixed number of instances of a register file. As another example, the nodes representing macro components can include one or more nodes that each represent a phase-locked loop (PLL) circuit to be placed on the chip. As yet another example, the nodes representing macro components can include one or more nodes that each represent a sensor to be placed on the chip.

[0039] Standard cell components are a group of transistor and interconnect structures, e.g., a group that provides a Boolean logic function (e.g., AND, OR, XOR, XNOR, inverters) or a group that provides a storage function (e.g., flip-flop or latch).

[0040] In some implementations, nodes in the netlist data represent a single standard cell component. In some other implementations, nodes in the netlist data represent already clustered standard cell components.

[0041] Generally, the placement 152 assigns each node to a grid square in an A x B grid overlaid over the surface of the chip, where A and B are integers. In some cases, some or all of the nodes may be larger than a single grid square. In these cases, assigning a node to a grid square (or, more generally, to a location or position on the surface of the chip) refers toAttorney Docket No.: 45288-0612WO1

[0042] assigning a given position on the node, e.g., the center of the node, to a given position within the location or position, e.g.. the center of the grid square.

[0043] In some implementations, the values of A and B are provided as inputs to the system 100.

[0044] In some other implementations, the system 100 generates the values of A and B. For example, the system 100 can treat choosing the optimal number of rows and columns as a bin-packing problem and rank different combinations of rows and columns by the amount of wasted space they incur on the surface of the chip. The system 100 can then select the combination that results in the least amount of wasted space as the values for A and B.

[0045] As another example, the system 100 can process an input derived from the netlist data, data characterizing the surface of the integrated circuit chip, or both using a grid generation machine learning model that is configured to process the input to generate an output that defines how to divide the surface of the integrated circuit chip into the A and B grid.

[0046] The system 100 generates, as output, a chip placement 152 that places some or all of the nodes in the netlist data 102 at a respective position on the surface of the computer chip. That is, the chip placement 152 identifies a respective position on the surface of the computer chip for some or all of the nodes in the netlist data 102 and, therefore, for the integrated circuit components that are represented by the node. For convenience, this specification will refer to the placement of a given set of components that are represented by a node as placing the node that represents the components.

[0047] The system 100 can generate a respective set of one or more initial integrated circuit chip placements at each of one or more iterations in a sequence of multiple iterations.

[0048] For example, at each of one or more initial or intermediate iterations in the sequence, the system can obtain an initial integrated circuit chip placement that places some or all of the macro nodes in the netlist data 102 at a respective position on the surface of the integrated circuit chip without using suggested regions generated using the multimodal neural network. That is. the initial integrated circuit chip placement identifies a respective position on the surface of the integrated circuit chip for some or all of the macro nodes in the netlist data 102 and, therefore, for the integrated circuit components that are represented by the node.

[0049] For example, the system can generate the initial integrated circuit chip placement, e.g., using a policy model, an analytical heuristic method, or another model configured to generate integrated circuit chip placements. As another example, the system can obtain theAttorney Docket No.: 45288-0612WO1

[0050] initial integrated circuit chip placement, e.g., from a user. For example, the initial integrated circuit chip placement can be created manually by a user, or by allowing a user to modify placements generated by other neural networks.

[0051] At each of one or more other iterations in the sequence, e.g., that are subsequent to the one or more initial iterations or in between the one or more intermediate iterations, the system can determine one or more sets of suggested regions 112a-n on the surface of the integrated circuit chip. Each set can include a suggested region for one or more of the macro nodes in the netlist data. The system can use a multimodal neural network 1 10 to generate each set.

[0052] At each of the other iterations in the sequence, the system can generate, for each set 112a-n and using a policy model 120, one or more respective initial integrated circuit chip placements 122a-k. The policy model 120 places the macro nodes in the netlist of nodes at respective locations on the surface of the integrated circuit chip.

[0053] The system can generate the initial integrated circuit chip placements 122a-k based on at least the suggested regions 112a-n. For example, the system can constrain the action space of the policy model 120 based on the suggested regions 112a-n. An example process for generating the initial chip placements 122a-k are described in further detail below with reference to FIG. 5.

[0054] In some examples, to generate each set of suggested regions, the system can process a neural network input 108 that includes at least data representing one or more previously obtained initial integrated circuit chip placements 122a-k using the multimodal neural network. The one or more previously obtained initial circuit chip placements can include initial integrated circuit chip placements obtained at one or more initial or intermediate iterations, or initial integrated circuit chip placements generated at one or more other iterations, or both.

[0055] The multimodal neural network 110 can have any appropriate neural network architecture for processing a neural network input that includes text, images, or both, to generate a neural network output that includes text, images, or both. In particular, the multimodal neural network 110 can have any of a variety of processing layers (e.g., feedforward layers, convolutional layers, attention layers, graph processing layers, etc.) in any appropriate number (e.g., 5 layers, or 10 layers, or 50 layers), and connected in any appropriate configuration (e.g., as a directed graph of layers.

[0056] Prior to using the multi-modal neural network 110 to generate network outputs, the multi-modal neural network 110 is pre-trained e.g., by the system 100 or by one or more other systems.Attorney Docket No.: 45288-0612WO1

[0057] In particular, the system 100 or the other system(s) pre-trains the neural network 110 on any appropriate objective function for a multi-modal generation task, e.g.. a task that requires predicting, given a current sequence of tokens, the next token that follows the current sequence in the training data, a contrastive learning task that aligns image and text features in a joint embedding space, etc.

[0058] In some implementations, the system 100 or another system can further train the multi-modal neural network, as described further in detail below with reference to FIG. 2.

[0059] Examples of multi-modal generative neural networks include vision language model neural networks and multi-modal language model neural networks. For example, examples of such neural networks include those described in Comanici, Gheorghe, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025), Gemma Team, et al. Gemma 3 Technical Report arXiv preprint arXiv:2503.19786 (2025), and PaliGemma.

[0060] Example multi-modal generative neural networks are described in detail further below.

[0061] The policy model 120 can be any appropriate policy model that is configured to generate a score distribution over multiple positions on the surface of the integrated circuit chip. The policy model 120 can have been trained to generate initial integrated circuit chip placements that optimize for criteria such as a degree of overlap of the macro nodes in the placement, wire length of wires between nodes on the surface of the chip in the initial integrated circuit chip placement, routing congestion measuring available wiring resources in a given region on the chip compared to wires that run through the given region in the initial integrated circuit chip placement, a cost for the initial integrated circuit chip placement, power consumption for the chip, heat generation for the chip, area of the initial integrated circuit chip placement, or timing of the initial integrated circuit chip placement. In some examples, the system 100 can provide a policy model input as input to the policy model 120 to generate the initial chip placements 122a-k. In some examples, the policy model input can include at least data representing the netlist data 102. In some examples, the policy model input can include at least text data that describes semantic guidance for the policy model 120.

[0062] In some examples, the policy model 120 can have a neural network architecture. The policy model 120 can have any appropriate neural network architecture for performing chip placement. In particular, the policy model 120 can have any of a variety of processing layers (e.g., feed-forward layers, convolutional layers, attention layers, graph processing layers, etc.) in any appropriate number (e.g., 5 layers, or 10 layers, or 50 layers), and connected in anyAttorney Docket No.: 45288-0612WO1

[0063] appropriate configuration (e.g., as a directed graph of layers. Some example suitable policy models are described in further detail in Mirhoseini. A., et al., Chip Placement with Deep Reinforcement Learning, arXiv preprint arXiv:2004.10746 (2020); Mirhoseini, A., et al., A graph placement methodology for fast chip design, Nature, 594(7862):207-212, 2021;

[0064] GENERATING INTEGRATED CIRCUIT PLACEMENTS USING NEURAL NETWORKS CONDITIONED ON TARGET REWARD VALUES, U.S. Patent Publication No.

[0065] 20260057164A1; and GENERATING INTEGRATED CIRCUIT PLACEMENTS USING NEURAL NETWORKS, U.S. Patent Publication No. 20250124207A1, which are herein incorporated by reference in their entirety.

[0066] Although this specification describes placing nodes using a policy model, in general the system can use any appropriate technique to place nodes according to the suggested regions. For example, the system can use the suggested regions to guide an analytical placer. As an example, the system can formulate the placement of the integrated circuit components as a nonlinear optimization problem that minimizes wirelength subject to density constraints. To incorporate the spatial guidance from the multimodal neural network, the system can interpret the suggested regions as target locations and apply them as anchor constraints within the optimization objective. For example, the system can convert the bounding box of a suggested region into an anchor point, such as the bottom-left comer of the region. In some examples, the system can add a quadratic anchor penalty to the objective function, which is calculated as the sum of the squared distances between each macro node's actual position and its corresponding anchor point. In some examples, the system can apply an anchor weight to the penalty to control the strength of the guidance. The system can use the anchor weight to pull the macro nodes toward the multimodal neural network's target locations.

[0067] In some examples, the netlist data 102 can include standard cell nodes representing standard cell components. The system 100 can generate an integrated circuit chip placement 152 that places some or all of the macro nodes and the standard cell nodes at a respective position on the surface of the integrated circuit chip. For example, the system 100 can place the standard cell nodes while keeping the position of the macro nodes fixed.

[0068] In some implementations, the system can use a policy model, e.g., the same policy model as the policy model 120 or a different policy model, to place the standard cell nodes. For example, the system can use the policy model 120 to place the macro nodes and the standard cell nodes.

[0069] In some implementations, to place the standard cell nodes, the system clusters the standard cells into a set of standard cell clusters (or obtains data identifying already generatedAttorney Docket No.: 45288-0612WO1

[0070] clusters) and then places each cluster of standard cells at a respective position on the surface of the partially placed integrated circuit chip using a graph placement technique. As a particular example, the system can cluster the standard cells using a partitioning technique that is based on the normalized minimum cut objective. An example of such a technique is hMETIS, which is described in Kary pis, G. and Kumar, V. A hypergraph partitioning package. In HMETIS, 1998. In some examples, to enable faster computation of the performance metrics during the evolutionary search, the system groups the standard cell nodes into a fixed number of clusters with a specific target cell area utilization. In some examples, the system can split larger groups of standard cells. In some of these implementations, the system leaves the standard cells unclustered when determining the one or more performance metrics for the integrated circuit chip placement.

[0071] In some other implementations, the system does not cluster the standard cells and directly places each standard cell at a respective position on the surface of the partially placed integrated circuit chip using the graph placement technique.

[0072] The graph placement technique can be any appropriate technique for placing nodes of a graph. For example, the system can use a force based technique, i.e., a force-directed technique. In particular, when using a force based technique, the system represents the netlist as a system of springs that apply force to each node, according to the weight x distance formula, causing tightly connected nodes to be attracted to one another. Optionally, the system also introduces a repulsive force between overlapping nodes to reduce placement density. After applying all forces, the system moves nodes in the direction of the force vector. To reduce oscillations, the system can set a maximum distance for each move. Using force-directed techniques to place nodes is described in more detail in Shahookar, K. and Mazumder. P. Vlsi cell placement techniques. ACM Comput. Surv., 23(2): 143220. June 1991. ISSN 0360-0300. doi: 10.1145 / 103724.103725.

[0073] The system determines, for each of the initial chip placements 122a-k, one or more performance metrics that measure a quality of the initial integrated circuit chip placement. For example, the system can process each of the initial chip placements 122a-k using a quality engine 130.

[0074] The quality engine 130 is configured to determine, for each placement, one or more performance metrics 132a-k. Each performance metric can measure, for example, wire length of wires between nodes on the surface of the chip in the initial integrated circuit chip placement, a cost for the initial integrated circuit chip placement, or routing congestionAttorney Docket No.: 45288-0612WO1

[0075] measuring available wiring resources in a given region on the chip compared to wires that run through the given region in the initial integrated circuit placement.

[0076] The system includes the performance metrics 132a-k in the neural network input 108 that is provided as input to the multimodal neural network 110 for one or more subsequent iterations. The system thus creates a feedback loop for evolutionary' improvement of the quality of chip placements.

[0077] In some implementations, the system can maintain a buffer, e.g., a database for storing initial integrated circuit chip placements. For example, the system can maintain a population of previously generated initial integrated circuit chip placements in the buffer. The population of previously generated initial integrated circuit chip placements can include, e.g., initial integrated circuit chip placements generated using suggested regions generated using the multimodal neural network, initial integrated circuit chip placements generated without using suggested regions generated using the multimodal neural network. In some examples, the system can maintain other initial integrated circuit chip placements, e.g., initial integrated circuit chip placements obtained from a user.

[0078] In some examples, the system can maintain the one or more performance metrics for each initial integrated circuit chip placement in the buffer. The system can select initial integrated circuit chip placements to include in the neural network input at subsequent iterations from the buffer, as described below with reference to FIG. 4.

[0079] In some implementations, the system 100 uses an initial chip placement from the initial chip placements 122a-k at a particular iteration, e.g., at a final iteration or when a termination criterion is met, as the chip placement 152.

[0080] In some other implementations, the system 100 provides an initial chip placement from the initial chip placements 122a-k at a particular iteration as input to a legalization engine that adjusts the initial placement to generate the chip placement 152.

[0081] In particular, the legalization engine can generate a legalized integrated circuit chip placement by applying a greedy legalization algorithm to the initial integrated circuit chip placement. For example, the engine can perform a greedy legalization step to snap macros, standard cells, or both, onto the nearest legal position while honoring a set of minimum spacing constraints.

[0082] Optionally, the engine can further refine the legalized placement or can refine the initial placement directly without generating the legalized placement, e.g., by performing simulated annealing or by providing the legalized placement or the initial placement to an electronic design automation (EDA) software tool for evaluation and fine-tuning.Attorney Docket No.: 45288-0612WO1

[0083] Optionally, the system 100 or an external system can then fabricate (produce) a chip (integrated circuit) according to the chip placement 152. Such an integrated circuit may exhibit improved performance, e.g., have one or more of lower power consumption, lower latency, or smaller surface area, than one designed using a conventional design process, and / or be producible using fewer resources. The fabrication may use any know n technique. In some cases, fabricating the chip according to the final placement can include presenting data identifying the placement to a user to allow the user to modify the chip placement 152 before fabrication or providing the chip placement 152 to an electronic design automation (EDA) for fine-tuning before fabrication. Providing or outputting the chip placement to an EDA tool enables further evaluation and fine-tuning, such as using a legalization engine to resolve macro node overlaps and ensure nodes are validly positioned according to minimum spacing constraints. In this way, fabrication of an integrated circuit chip that exhibits reduced power consumption, lower latency, or a smaller surface area can be facilitated.

[0084] The system 100 can receive the netlist data 102 in any of a variety of ways.

[0085] For example, the system 100 can receive the netlist data 102 as an upload from a remote user of the system over a data communication network, e.g., using an application programming interface (API) made available by' the system 100. In some cases, the system 100 can then provide the chip placement 152 to the remote user through the API provided by the system 100, e.g., for use in fabricating a chip according to the chip placement 152.

[0086] As another example, the system 100 can be part of an electronic design automation (EDA) software tool and can receive the netlist data 102 from a user of the tool or from another component of the tool. In this example, the system 100 can provide the chip placement 152 for evaluation by another component of the EDA software tool before the computer chip is fabricated.

[0087] In some cases, the multimodal neural network 110 is configured to perform a multimodal processing task that requires processing multi-modal data. In general, multi-modal data is a combination of two or more different types of data, e.g., two or more of audio data, image data, video data, text data, or graph data. As one example the multi-modal data may comprise audio-visual data, comprising a combination of pixels of an image or of video and audio data representing values of a digitized audio waveform. As another example the multimodal data may comprise a combination of i) text data representing text in a natural language and ii) pixels of an image or of video or audio data representing values of an audio waveform. Optionally, but not necessarily, the different types of data may represent the same orAttorney Docket No.: 45288-0612WO1

[0088] overlapping objects using the different modalities (types), and when processing multi-modal data, the data may be mapped into a common embedding space.

[0089] In some cases, the task is a multi-modal task that requires processing both text and image inputs. That is, the target output to be generated by the computer vision neural network for a given image depends on one or more outputs generated by the text processing neural network for one or more corresponding text inputs (and vice versa). Examples of such tasks include open-vocabulary image classification, open- vocabulary object detection, image captioning, text-based image search, image-based retrieval, and so on.

[0090] Generally, the multimodal neural network 110 is configured to process a neural network input (“prompt’') to generate a data item. The data item can include data in any of a variety of modalities, e.g., text data, image data, video data, or audio data. Generally, the data item represents a response to the neural network input which may be, e.g. a “prompt” for the multimodal neural network 110. For example, the neural network input can characterize one or more desired properties for the generated data item.

[0091] In some implementations the multimodal neural network 110 generates an output token sequence from an input token sequence including the neural network input. The multimodal neural network 110 may then be configured to process the input token sequence to generate for each position in the output token sequence, a respective score for each token in a vocabulary of output tokens, that is used to select an output token for the output token sequence.

[0092] In some implementations the tokens can represent text, e.g., words, wordpieces or characters, in a natural or computer language. For example, text may be received, e.g., as a series of encoded characters, e.g. UTF-8 encoded characters; such “characters” can include Chinese and other similar characters, as well as logograms, syllabograms and the like. A text encoder, i.e. a tokenizer, can process a sequence of text to represent the text as a series of text tokens from a vocabulary of text tokens, e.g. that each represent words, wordpieces or characters in a natural or computer language. The computer language may be any formal language used to communicate with a computer, e.g. a markup language, or a command or configuration language, or a data exchange language such as JSON, or a programming language. The tokenizer can, e g., implement BPE (Byte Pair Encoding) or Wordpiece tokenization. Optionally the text can be obtained from audio data representing speech; the output tokens may be converted into audio data that represent speech corresponding to the text.Attorney Docket No.: 45288-0612WO1

[0093] Also, or instead the tokens may represent an image. For example, a set (sequence) of input or output tokens can represent an image. Each image token may comprise a patch encoding of values of the pixels in a different region of an image that maps a set of values of the pixels to a respective image token. The patch encoder may comprise a neural network, e.g. having one or more (self-)attention layers, such as a Transformer neural network.

[0094] Also, or instead the tokens may represent an audio waveform. For example, a set (sequence) of input or output tokens can represent audio data representing a waveform e.g. instantaneous audio amplitude values or time-frequency audio data. Each image token may comprise a segment encoding of the audio waveform in a different time segment of the audio that maps a set of values representing the audio waveform to a respective image token. The segment encoder may comprise a neural network, e.g. having one or more (self-)attention layers, such as a Transformer neural network.

[0095] In a multimodal system audio data or an image may be flagged by a start-of-audio token or start-of-image token.

[0096] In general, the multimodal neural network 110 can have any appropriate architecture to enable it to perform these machine learning tasks, e.g., to perform a generative task by processing the neural netw ork input to generate the data item.

[0097] In some implementations, the neural network can be configured as an auto-regressive generative neural network.

[0098] A neural network is referred to as an auto-regressive generative neural network when the neural network auto-regressively generates an output sequence of tokens as the output.

[0099] More specifically, the auto-regressively generated output is created by generating each particular token in the output sequence conditioned on a current input sequence that includes an input sequence included in the input and any tokens that precede the particular token in the output sequence, i.e., the tokens that have already been generated for any previous positions in the output sequence that precede the particular position of the particular token.

[0100] As a particular example, the neural network can have any of a variety of Transformerbased neural network architectures, e.g.. encoder-only Transformer architectures, encoderdecoder Transformer architectures, decoder-only Transformer architectures, mixture-of-experts (MoE) Transformer architectures, other attention-based architectures, and so on.

[0101] Examples of such Transformer-based neural network architectures include those described in Colin Raffel, et al., Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683, 2019; Daniel Adiwardana. et al..Attorney Docket No.: 45288-0612WO1

[0102] Towards a human-like open-domain chatbot. CoRR, abs / 2001.09977. 2020; Aakanksha Chowdhery, et al., PaLM: Scaling Language Modeling with Pathways, arXiv preprint arXiv:2204.02311; Rohan Anil, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023; Gemini Team, et al., Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023); Comanici, Gheorghe, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025); and Gemma Team, et al. Gemma 3 technical report. arXiv preprint arXiv:2503.19786 (2025).

[0103] In general, a vision language model neural netw ork can be a neural network that has been trained so that, given a text prompt that includes a sequence of tokens in a natural language, the neural network can generate the next token in the sequence. This process can be repeated to extend the text prompt one token at a time to generate a natural language output, i.e., to generate the natural language output auto-regressively token by token. At each time “time step,’' the language model neural network processes the current sequence to generate a probability distribution over a vocabulary of tokens. The next token can then be selected using the probability distribution, e.g., by sampling from the distribution using nucleus sampling or another sampling technique or by selecting the highest-probability token. The tokens in the vocabulary' can include any of a variety of tokens, e.g., some combination of words, sub-words, characters, punctuation and other symbols, and numbers. In general, the language model neural network is trained on a corpus of text made up of tokens from the vocabulary (and optionally other tokens that can be mapped to a designated out-of-vocabulary token), to predict the next token in a sequence of tokens from the training data. The vision language model neural network can comprise an autoregressive Transformer neural network.

[0104] A vision language model neural network can be made to perform a particular task by providing a natural language description of the desired response as a neural netw ork input or “prompt” (input sequence). In some cases, the prompt may be a few -shot prompt where a few, e.g., 1 to 10, examples of a query and an example output are provided in the text prior to the actual query.

[0105] Instead or in addition, a vision language model neural network may be “fine-tuned” to perform a particular task, by obtaining a pre-trained language model neural network trained on a large corpus of examples and then further training part of all of the language model neural network on a relatively small number of examples particular to the type of task that is to be performed.Attorney Docket No.: 45288-0612WO1

[0106] The vision language model neural network may be a large language model neural network, e.g., one that has greater than 1 billion, 10 billion or 100 billion trained parameters. The vision language model neural network may have been trained on greater than 10 billion, 100 billion or 1000 billion words or tokens representing words or other tokens.

[0107] FIG. 2 is a flow diagram of an example process 200 for generating a set of integrated circuit chip placements. For convenience, the process 200 will be described as being performed by a system of one or more computers located in one or more locations. For example, a placement generation system, e.g., the placement generation system 100 of FIG.

[0108] 1, appropriately programmed, can perform the process 200.

[0109] The system obtains netlist data for an integrated circuit (IC) chip (step 202). The netlist data specifies a connectivity on the integrated circuit chip between multiple nodes that each correspond to one or more of multiple integrated circuit components of the integrated circuit chip. The multiple nodes include multiple macro nodes that each represent a respective macro component of the integrated circuit chip.

[0110] The system generates a set of one or more integrated circuit chip placements. Each integrated circuit chip placement places each macro node in the netlist data at a respective position on a surface of the integrated circuit chip.

[0111] As an example, the respective positions can include grid squares from a grid overlaid over the surface of the integrated circuit chip.

[0112] The system generates the set of integrated circuit chip placements by generating a respective set of one or more initial integrated circuit chip placements at each of a sequence of iterations (step 206). Each respective initial integrated circuit chip placement places each macro node in the netlist data at a respective position on the surface of the integrated circuit chip.

[0113] At each of one or more initial iterations, the system can generate, using the policy model, an initial integrated circuit chip placement for the initial iteration. That is, the system can generate the respective initial integrated circuit chip placement for the initial iteration without using any suggested regions generated by the multimodal neural network.

[0114] Generating an initial integrated circuit chip placement is described in further detail below with reference to FIG. 5.

[0115] Alternatively or in addition, at each of one or more initial iterations, the system can obtain an initial integrated circuit chip placement for the initial iteration, e.g.. using another policy model, an analytical heuristic method, or another model configured to generateAttorney Docket No.: 45288-0612WO1

[0116] integrated circuit chip placements, or receiving the initial integrated circuit chip placement from a user.

[0117] At each of one or more particular iterations in the sequence of iterations, the system performs steps 208-212.

[0118] The system determines, using a multimodal neural network, one or more respective sets of suggested regions on the surface of the integrated circuit chip (step 208).

[0119] For example, the system can generate a neural network input for the particular iteration, and sample from the multimodal neural network given the neural network input for the particular iteration, to determine the one or more respective sets of suggested regions. An example process for determining the respective sets of suggested regions is described further below with reference to FIG. 3.

[0120] The system generates, for each set of suggested regions and using a policy model, one or more respective initial integrated circuit chip placements (step 210). For example, the system can generate the one or more respective initial integrated circuit chip placements based on the netlist data and the set of suggested regions.

[0121] An example process for generating respective initial integrated circuit chip placements is described below in further detail with reference to FIG. 5.

[0122] In some implementations, the system modifies each of the respective initial integrated circuit chip placements, e.g., to ensure the placement is physically viable. For example, the system can use a legalization engine to resolve overlaps and to guarantee the nodes are validly positioned.

[0123] The system determines, for each respective initial integrated circuit chip placement, one or more performance metrics (step 212). Each performance metric measures a quality of the initial integrated circuit chip placement.

[0124] As an example, a performance metric can measure wire length of wires between nodes on the surface of the chip in the initial integrated circuit chip placement. For example, the wire length can include a half-perimeter wirelength (HPWL) calculated by summing a half-perimeter of a smallest axis-aligned bounding box enclosing pins connected by each net across all nets, e.g., connections between nodes, in the netlist data.

[0125] Alternatively or in addition, a performance metric can measure a cost for the initial integrated circuit chip placement. The cost can measure, for example, performance, power, or area (PPA) of the integrated circuit chip.Attorney Docket No.: 45288-0612WO1

[0126] Alternatively or in addition, a performance metric can measure routing congestion, which measures available wiring resources in a given region on the chip compared to wires that run through the given region in the initial integrated circuit placement.

[0127] In some implementations, at each of one or more intermediate iterations that are after the first particular iteration in the sequence, the system can generate the respective set of initial integrated circuit chip placements for the intermediate iteration by generating, using the policy model, one or more respective initial integrated circuit chip placements for the intermediate iteration based on the one or more sets of suggested regions generated at a most recent particular iteration. That is, the system can generate the respective initial integrated circuit chip placements for the intermediate iteration without using the multimodal neural network to determine suggested regions for the intermediate iteration. The system can thus perform multiple rollouts of the policy model using the same high-level spatial guidance from the multimodal neural network over a set interval of intermediate iterations. By querying the multimodal neural network at a set interval rather than at every7iteration in the sequence, the system can reduce computational overhead and latency associated with multimodal neural network inference while still benefiting from the strategic spatial constraints. The system can thus balance the high-level reasoning of the multimodal neural network with the placement capabilities of the policy model, enabling the system to explore variations within promising subregions before evolving the next set of suggested regions.

[0128] As an example, the process can include E iterations, where e is the current iteration. The process can include a query interval of K. The system can include a buffer H. For each episode e from 1 to E, the system initializes an empty placement Pe. The system queries the multimodal neural network to determine the sets of suggested regions only at the particular iterations where the current iteration e modulo K is zero. For intermediate iterations, the system relies on the previously generated suggestions, thereby reducing computational overhead. Furthermore, during one or more initial iterations at the beginning of the sequence (e.g., where e is less than K and e modulo K is not zero), the system assigns macro node positions by sampling from the unconstrained policy model to generate unguided initial integrated circuit chip placements. The system calculates the performance metrics for these unguided placements and adds them to the buffer H, populating the buffer H with a baseline visual and textual history for the multimodal neural network to evaluate at the first particular iteration (e.g., at iteration K).

[0129] Alternatively or in addition, at each of one or more intermediate iterations, the system can obtain an initial integrated circuit chip placement for the intermediate iteration, e.g.,Attorney Docket No.: 45288-0612WO1

[0130] using another policy model, an analytical heuristic method, or another model configured to generate integrated circuit chip placements, or receiving the initial integrated circuit chip placement from a user.

[0131] In some implementations, the system or another training system can train the multimodal neural network on training examples that each include a training initial integrated circuit chip placement. In some implementations, the system or another training system can train the multimodal neural network on training examples that each include data representing the netlist data. For example, the multimodal neural network can have been pre-trained on a next token prediction task, and further trained using training examples using placement data. The system can thus specialize the capabilities of the multimodal neural network for Electronic Design Automation (EDA) tasks, improving multimodal neural network’s domain understanding and inference efficiency.

[0132] In some implementations, the policy model can have been trained on multiple training examples to generate integrated circuit chip placements that optimize for one or more criteria. The one or more criteria can measure, for example, one or more of: a degree of overlap of the macro nodes in the initial integrated circuit chip placement, wire length of wires between nodes on the surface of the chip in the initial integrated circuit chip placement, routing congestion measuring available wiring resources in a given region on the chip compared to wires that run through the given region in the initial integrated circuit chip placement, a cost for the initial integrated circuit chip placement, pow er consumption for the chip, heat generation for the chip, area of the initial integrated circuit chip placement, or timing of the initial integrated circuit chip placement. As a particular example, the policy model can have been trained using reinforcement learning techniques across a dataset of training integrated circuit chip placements to optimize for wire length, e.g., calculated as the Half-Perimeter Wirelength (HPWL).

[0133] In some implementations, for one or more initial integrated circuit chip placements of the respective sets of initial integrated circuit chip placements, the system can generate a respective training example that includes the initial integrated circuit chip placement. The system can train the policy model using the respective training examples. The system can perform online fine-tuning of the policy7model using reinforcement learning techniques using new7integrated circuit chip placements generated during inference with an unseen chip. The system can perform any appropriate downstream task using the set of one or more integrated circuit chip placements.Attorney Docket No.: 45288-0612WO1

[0134] For example, the system can select an initial integrated circuit chip placement from the respective sets of initial integrated circuit chip placements. As an example, the system can select the initial integrated circuit chip placement from the respective set of initial integrated circuit chip placements for the final iteration. Alternatively or in addition, the system can select the initial integrated circuit chip placement from the respective sets of initial integrated circuit chip placements based on the performance metrics for each initial integrated circuit chip placement. For example, the system can select the initial integrated circuit chip placement with the best performance according to the performance metrics.

[0135] As another example, the system can, for one or more of the set of initial integrated circuit chip placements, fabricate an integrated circuit chip according to the initial integrated circuit chip placement

[0136] As another example, the system can output data specifying the initial integrated circuit chip placement for one or more of the respective initial integrated circuit chip placements.

[0137] For example, the system can output the initial integrated circuit chip placement to a component of an electronic design automation (EDA) software tool.

[0138] As another example, the system can provide the data specifying the initial integrated circuit chip placement for presentation to a user on a user device.

[0139] As another example, the system can provide the initial integrated circuit chip placement for fabrication of an integrated circuit chip according to the initial integrated circuit chip placement.

[0140] In some implementations, each integrated circuit chip placement places only each macro node in the netlist data at a respective position on the surface of the integrated circuit chip.

[0141] In some implementations, the multiple nodes include multiple standard cell nodes that each represent a respective standard cell component of the integrated circuit chip, and each integrated circuit chip placement places each macro node in the netlist data at a respective position on the surface of the integrated circuit chip and each standard cell node at a respective position on the surface of the integrated circuit chip. For example, for one or more of the respective initial integrated circuit chip placements, the system can place each of the standard cell nodes at a respective position on the surface of the integrated circuit chip that includes the macro components represented by the macro nodes placed according to the respective initial integrated circuit chip placement.Attorney Docket No.: 45288-0612WO1

[0142] The system can perform any appropriate dow nstream task using the set of one or more integrated circuit chip placements that places each macro node in the netlist data at a respective position on the surface of the integrated circuit chip and each standard cell node at a respective position on the surface of the integrated circuit chip.

[0143] For example, the system can select an initial integrated circuit chip placement from the respective sets of initial integrated circuit chip placements. As an example, the system can select the initial integrated circuit chip placement from the respective set of initial integrated circuit chip placements for the final iteration. Alternatively or in addition, the system can select the initial integrated circuit chip placement from the respective sets of initial integrated circuit chip placements based on the performance metrics for each initial integrated circuit chip placement. For example, the system can select the initial integrated circuit chip placement with the best performance according to the performance metrics.

[0144] As another example, the system can, for one or more of the set of initial integrated circuit chip placements, fabricate an integrated circuit chip according to the initial integrated circuit chip placement

[0145] As another example, the system can output data specifying the initial integrated circuit chip placement for one or more of the respective initial integrated circuit chip placements.

[0146] For example, the system can output the initial integrated circuit chip placement to a component of an electronic design automation (EDA) software tool.

[0147] As another example, the system can provide the data specifying the initial integrated circuit chip placement for presentation to a user on a user device.

[0148] As another example, the system can provide the initial integrated circuit chip placement for fabrication of an integrated circuit chip according to the initial integrated circuit chip placement.

[0149] FIG. 3 is a flow diagram of an example process 300 for determining one or more respective sets of suggested regions on the surface of an integrated circuit chip. For convenience, the process 300 will be described as being performed by a system of one or more computers located in one or more locations. For example, a placement generation system, e.g., the placement generation system 100 of FIG. 1, appropriately programmed, can perform the process 300.

[0150] The system performs the process 300 as part of step 208 at each of one or more particular iterations as described with reference to FIG. 2.

[0151] The system generates a neural network input for the particular iteration (step 302).Attorney Docket No.: 45288-0612WO1

[0152] The neural network input for the particular iteration includes data representing one or more initial integrated circuit chip placements generated at one or more preceding iterations in the sequence.

[0153] The data representing the one or more initial integrated circuit chip placements can include, for each initial integrated circuit chip placement, a respective set of coordinates for each macro node. The set of coordinates for each macro node can identify the respective position of the macro node on the surface of the chip in the initial integrated circuit chip placement.

[0154] For example, the respective set of coordinates can specify the bottom-left comer and the top-right comer of each macro node on the discrete grid representing the surface of the integrated circuit chip.

[0155] Alternatively or in addition, the data representing the one or more initial integrated circuit chip placements can include, for each initial integrated circuit chip placement, an image depicting each macro node at the respective position of the macro node on the surface of the integrated circuit chip in the initial integrated circuit chip placement.

[0156] For example, the image can depict each macro node as a rectangle at its corresponding grid coordinates. The image can provide a visual representation of the chip layout, allowing the multimodal neural network to perceive spatial relationships, available canvas space, and placement patterns from previous initial integrated circuit chip placements.

[0157] In some implementations, the image can depict each macro node in a corresponding color. For example, the image can represent the macro nodes using a color coding strategy to convey functional relationships or connectivity information, e.g., derived from the netlist data. For example, macro nodes that share a high number of nets can be assigned similar colors, providing a visual signal for the multimodal neural network to suggest optimal regions. By assigning similar colors in this way, the approach descnbed herein can encode functional connectivity and topological relationships into a multi-channel image format optimized for visual reasoning by the multimodal neural network.

[0158] In some of these implementations, the system can determine the corresponding color for each macro node. For example, the system can generate a respective embedded representation for each macro node in the netlist data. As an example, the system can generate a connectivity graph using the netlist data. For example, the netlist data can be structured as a hypergraph where vertices represent nodes. Each hyperedge, also referred to as a net, connects a subset of related nodes together. Each hyperedge identifies which nodesAttorney Docket No.: 45288-0612WO1

[0159] need to be connected to each other by one or more wires in the manufactured integrated circuit chip.

[0160] The system can represent indirect connections through nets in the hypergraph as edges between macro nodes in the connectivity graph. The connectivity' graph includes vertices and edges. For example, the vertices can represent the individual macro nodes, and an undirected, weighted edge connects any two macro nodes that share one or more nets. The weight of the edge is directly proportional to the number of shared nets. The system can thus flatten the netlisf s hypergraph structure into a connectivity graph.

[0161] The system can process the connectivity7graph to generate the respective embedded representation for each macro node in the netlist data. For example, the system can use a spring-based graph layout algorithm to embed the graph into a low-dimensional space.

[0162] Example suitable spring-based graph layout algorithms are described in Hagberg, A. et al., Exploring netw ork structure, dynamics, and function using network, Los Alamos National Laboratory (2008); and Shahookar, K. and Mazumder, P. Vlsi cell placement techniques. ACM Comput. Surv., 23(2): 143220. June 1991. ISSN 0360-0300. doi:

[0163] 10.1145 / 103724.103725. The system can thus position macro nodes with stronger connections closer to one another in the embedded space.

[0164] The system can generate clusters for the macro nodes using the respective embedded representations. For example, the system can apply any appropriate clustering algorithm, e.g., k-means clustering, hierarchical clustering, etc., to the respective embedded representations to generate the clusters. In some examples, to determine the optimal number of clusters k, the system can iterate through a predefined range of values of k (e.g., from 2 to 30). For each value of k, the system can evaluate the clusters using a score that measures cluster separation. The system can select the k value that has the highest score.

[0165] The system can assign a respective color to each cluster. For each cluster, the system can assign the respective color for the cluster to each macro node of the cluster. In some examples, the system can apply a default color to any isolated macros that do not share nets and are excluded from the connectivity7graph.

[0166] In some examples, the one or more previously generated initial integrated circuit chip placements included in the neural network input can include placements obtained at the one or more initial or intermediate iterations, e.g., placements generated without multimodal neural network guidance or externally obtained baseline placements. Alternatively or in addition, the previously generated initial integrated circuit chip placements can include placements generated at one or more preceding particular iterations, e.g., placementsAttorney Docket No.: 45288-0612WO1

[0167] generated using previous sets of suggested regions. That is, the multimodal neural network can evaluate a history of both unguided placement attempts and its own previously guided placement attempts.

[0168] In some implementations, the system can include data representing all initial integrated circuit chip placements generated at the one or more preceding iterations in the sequence. In some other implementations, the system can select the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations. Some example processes for selecting the one or more initial integrated circuit chip placements to include in the neural network input for the particular iteration are described below with reference to FIG. 4.

[0169] In some implementations, the neural network input for the particular iteration includes one or more respective performance metrics for each of the one or more initial integrated circuit chip placements. For example, the one or more respective performance metrics can have been generated at one or more preceding iterations as described above with reference to FIG. 2.

[0170] In some implementations, the neural network input for the particular iteration includes an instruction to suggest bounding box regions for placing each of one or more macro nodes on the surface of the integrated circuit chip.

[0171] As an example, the instruction can include one or more of: data representing the netlist data, a description of a size of the surface of the integrated circuit chip, a description of a coordinate system for the surface of the integrated circuit chip, a dimension for each of the one or more macro components, an orientation for each of the one or more macro components, a color for each of the one or more macro nodes, data representing a current placement, a request to describe the suggested regions, a request to explain the suggested regions, a request to describe a placement strategy, a request to generate the suggested regions according to a macro node order, a request to analyze the one or more initial integrated circuit chip placements included in the neural network input, a description of an output format, one or more output examples for representing each suggested region, code representing the integrated circuit, or a description of one or more performance metrics.

[0172] For example, the data representing the netlist data can include, for each macro node in the netlist data, one or more of: an identifier for the macro node, a corresponding color for the macro node, or a dimension for the macro node.

[0173] As another example, a description of a size of the surface of the integrated circuit chip can include the grid resolution, e.g., 84x84.Attorney Docket No.: 45288-0612WO1

[0174] As another example, a description of a coordinate system for the surface of the integrated circuit chip can define boundaries of the grid, e.g., that the origin (0.0) is at the bottom-left comer and the top-right comer is (84,84).

[0175] As another example, a dimension for each of the one or more macro components can include the width and height of each macro node.

[0176] As another example, an orientation for each of the one or more macro components can include constraints such as not changing the orientation of macro nodes or not rotating the macro nodes.

[0177] As another example, a color for each of the one or more macro nodes can include a color, e.g., hex color code, indicating the connectivity group for the macro node.

[0178] As another example, the data representing the current placement can include one or more of: a respective set of coordinates for each of one or more macro nodes that identifies a respective current position of the macro node on the surface of the chip in the current placement, or an image depicting one or more macro nodes at the respective current position of the macro node on the surface of the integrated circuit chip in the current placement. In some examples, the image can depict one or more macro nodes for clear identification.

[0179] As another example, a request to describe the suggested regions can include a request to detail the regions using precise relative spatial relationships that align with an overall strategy.

[0180] As another example, a request to explain the suggested regions can include a request to provide justification for the placement, such as explaining how different color groups are organized and why the spatial relationships make sense to minimize wirelength and cost.

[0181] As another example, a request to describe a placement strategy can include a request for a description of the overall chip floorplan, or a request for optimization principles ranked by their importance based on historical evidence.

[0182] As another example, a request to generate the suggested regions according to a macro node order can include a request to suggest regions in decreasing order of size, e.g., largest first, to avoid overlapping region suggestions.

[0183] As another example, a request to analyze the one or more initial integrated circuit chip placements included in the neural network input can include a request to perform placement history analysis, e.g., examining color group position analytics, macro-level spatial relationships, critical edge alignments, adjacency relationships, or overall canvas utilization insights.Attorney Docket No.: 45288-0612WO1

[0184] As another example, a description of an output format can include all coordinates to be integers, ensuring regions have non-zero width and height, or formatting suggestions as ■‘MACRO NAME (W x H): (xl,yl) and (x2,y2)”.

[0185] As another example, one or more output examples for representing each suggested region can include text examples of the expected level of detail, e.g., demonstrating how to describe perfect shared boundaries (e.g., "RSTs right edge (x=42) precisely aligns with JKL's left edge") or specific unit channels between macros.

[0186] As another example, code representing the integrated circuit can include netlist properties, node connections, or code for the circuit function.

[0187] As another example, a description of one or more performance metrics can describe desired attributes, e.g., lower wirelength, lower cost, higher reward, etc.

[0188] In some implementations, the instruction provided to the multimodal neural network can include a prompt to control the balance of exploration and exploitation when generating suggested regions. For example, the instruction can include a greedy prompt that instructs the multimodal neural network to use previous placements as a strong prior and make only minor, low-risk refinements to the highest-performing historical arrangements. A greedy prompt can reduce computational overhead. Alternatively, the instruction can include an exploratory' prompt that instructs the network to use previous examples as context only and favor novel, diverse, and deliberate departures from prior placement patterns. As another example, the instruction can include a balanced default prompt that achieves optimal performance by balancing minor refinements with exploratory hypothesis generation without over-constraining the search space. An explanatory' prompt can increase the search space to avoid local optima.

[0189] The system samples from the multimodal neural network given the neural network input for the particular iteration (step 304). The system can sample from the multimodal neural network to generate one or more neural network outputs that each specify a respective set of suggested regions.

[0190] For example, to generate each neural network output, the system can process the neural network input using the multimodal neural network using a different random seed. As another example, the system can set sampling hyperparameters (e.g., temperature, Top-k, Top-p, etc.) to generate neural network outputs for the neural network input.

[0191] For example, each of the one or more neural network outputs can include data representing, for each suggested region in the respective set of suggested regions, a respective bounding box identifying the suggested region. In some examples, the data representing theAttorney Docket No.: 45288-0612WO1

[0192] respective bounding box identifying the suggested region can include a set of coordinates specifying the respective bounding boxes or an image depicting the respective bounding boxes. For example, as described above, the neural network input can include a request to define bounding boxes using the bottom-left and top-right. In some examples, the neural network input can include a request to suggest regions in decreasing order of size (largest first), which ensures that overlapping region suggestions are not generated.

[0193] In some implementations, the neural network output can include, for each suggested region in the respective set of suggested regions, an explanation for the suggested region. For example, the explanation can include a description of the overall chip floorplan, how different color groups are organized and how the overall strategy will minimize wirelength and cost, or relative spatial relationships for each macro node's region.

[0194] In some examples where the neural netw ork input includes coordinate data and performance metrics, the multimodal neural network can reason about how specific geometric arrangements of macro nodes correlate with physical design outcomes like wirelength and routing congestion, improving the quality of the neural network outputs.

[0195] An example of a neural network input is described below with reference to FIGS. 8A-8D.

[0196] FIG. 4 is a flow' diagram of an example process 400 for selecting one or more initial integrated circuit chip placements. For convenience, the process 400 will be described as being performed by a system of one or more computers located in one or more locations. For example, aplacement generation system, e.g., the placement generation system 100 of FIG.

[0197] 1, appropriately programmed, can perform the process 400.

[0198] In some implementations, the system can perform the process 400 as part of step 302 described above with reference to FIG. 3 to select one or more initial integrated circuit chip placements for inclusion in the neural network input for the particular iteration.

[0199] The system clusters each initial integrated circuit chip placement generated at each preceding iteration into one or more clusters (step 402).

[0200] For example, the system can obtain, for each initial integrated circuit chip placement, a corresponding set of coordinates for each macro node that identifies the respective position of the macro node on the surface of the chip in the initial integrated circuit chip placement. For example, the system can represent each placement in the population as a coordinate vector, e.g., by concatenating the x, y coordinates of all macro nodes.

[0201] The system can cluster the initial integrated circuit chip placements into the one or more clusters based on the corresponding sets of coordinates. For example, the system canAttorney Docket No.: 45288-0612WO1

[0202] apply K-means clustering on the coordinate vectors to form C clusters. In some examples, C can be based on the desired context length, or number of examples to be included in the network input.

[0203] The system selects a cluster based on a ranking for the clusters (step 404).

[0204] For example, the system can identify a respective feature value for each cluster. As an example, the respective feature value can include a minimum wire length of wires between nodes on the surface of the chip in the initial integrated circuit chip placement over each initial integrated circuit chip placement in the cluster.

[0205] The system can determine the ranking for the clusters based on the respective feature value for each cluster. For example, the system can determine the ranking based on the highest or lowest feature value in each cluster, e.g., based on the best global Half-Perimeter Wirelength (HPWL) in each cluster.

[0206] The system can select the cluster according to a softmax distribution over the ranking. For example, the system can sample a cluster with rank i with a probability proportional to ' T / T -y - , where is a temperature hyperparameter.

[0207] The system selects, using the selected cluster, one or more initial integrated circuit chip placements (step 406).

[0208] As an example, the system can select, using the selected cluster, one or more initial integrated circuit chip placements based on one or more performance metrics. As a particular example, the system can select, for each of one or more of the performance metrics, one or more initial integrated circuit chip placements from the selected cluster that each have a highest value for the performance metric. The system can thus provide the multimodal neural network with a set of C geometrically similar examples that represent variations on a single high-performing design pattern.

[0209] In some implementations, the system can select one or more second clusters that are closest to the selected cluster, e.g., based on centroid distance. For example, in some cases the selected cluster includes fewer than the required context length of initial integrated circuit chip placements. The system can select, for each of the one or more selected second clusters and for each of one or more of the performance metrics, one or more initial integrated circuit chip placements from the selected second cluster that each have a highest value for the performance metric.

[0210] In some other implementations, the system can select the one or more initial integrated circuit chip placements from the initial integrated circuit chip placementsAttorney Docket No.: 45288-0612WO1

[0211] generated at one or more immediately preceding iterations. The system can thus select the initial integrated circuit chip placements according to a First-In. First-Out (FIFO) approach, acting as a pure evolutionary search baseline where the multimodal neural network observes a temporally ordered sequence of recent placement attempts.

[0212] In some other implementations, the system can select the one or more initial integrated circuit chip placements by randomly sampling the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations. The system can thus provide a baseline that makes no assumptions about which placement attempts are most valuable for the evolutionary search.

[0213] In some other implementations, the system can select the one or more initial integrated circuit chip placements based on the one or more performance metrics for the initial integrated circuit chip placements generated at preceding iterations. By selecting the placements with the highest quality (e.g., the lowest global Half-Perimeter Wirelength), the system can thus expose the multimodal neural network to successful placement attempts that will encourage the multimodal neural network to replicate those patterns.

[0214] In some other implementations, the system can select the one or more initial integrated circuit chip placements by clustering each initial integrated circuit chip placement generated at each preceding iteration into one or more clusters. For example, the system can obtain, for each initial integrated circuit chip placement, a corresponding set of coordinates for each macro node that identifies the respective position of the macro node on the surface of the chip in the initial integrated circuit chip placement. The system can cluster the initial integrated circuit chip placements into the one or more clusters based on the corresponding sets of coordinates.

[0215] The system can select, from each cluster, one or more initial integrated circuit chip placements. For example, the system can select, from each cluster, the one or more initial integrated circuit chip placements based on the one or more performance metrics. As a particular example, the system can select, from each cluster and for each of one or more of the performance metrics, one or more initial integrated circuit chip placements from the cluster that each have a highest value for the performance metric. The system can thus promote geometric diversity across the selected examples while still favoring high-quality designs within each of those diverse clusters.

[0216] Thus, in some examples, the system can group geometrically similar floorplans together to either maximize geometric diversity or to focus on a single promising local regionAttorney Docket No.: 45288-0612WO1

[0217] for including initial integrated circuit chip placements in the model input for the multimodal neural network.

[0218] FIG. 5 is a flow diagram of an example process 500 for generating one or more initial integrated circuit chip placements. For convenience, the process 500 will be described as being performed by a system of one or more computers located in one or more locations. For example, a placement generation system, e.g., the placement generation system 100 of FIG.

[0219] 1, appropriately programmed, can perform the process 500.

[0220] The system can perform the process 500 as part of step 210 described above with reference to FIG. 2.

[0221] The system can perform the steps 502-508 for each macro node in the netlist data, at a corresponding time step of multiple time steps.

[0222] The system generates a policy model input for the corresponding time step (step 502). In some implementations, the policy model is configured to generate the initial integrated circuit chip placement autoregressively by placing the macro nodes over the multiple time steps.

[0223] As a particular example, the policy model can be configured to place each macro node at a position at a different corresponding time step according to a macro node order. The macro node order orders the macro nodes, with each node that is before any given macro node in the macro node order being placed before the given macro node. For example, the macro nodes can be ordered sequentially by descending macro area, or by connectivity and then size, ensuring that the largest and most highly connected macro nodes are placed first.

[0224] In some implementations, the policy model input for the corresponding time step can include one or more of: a target value of a reward function that measures a quality7of the initial integrated circuit chip placement, data characterizing respective positions on the surface of the chip of any macro nodes that are before a particular macro node to be placed at the corresponding time step in the macro node order, data characterizing the particular macro node to be placed at the corresponding time step, data characterizing the connectivity7between two or more of the multiple nodes specified by the netlist data, data representing the netlist data, or metadata for the netlist data.

[0225] In some other implementations, the policy model input for the corresponding time step can include one or more of: data representing the netlist data, data characterizing respective positions on the surface of the chip of any macro nodes that are before a particular macro node to be placed at the corresponding time step in the macro node order, data characterizing the particular macro node to be placed at the corresponding time step, dataAttorney Docket No.: 45288-0612WO1

[0226] characterizing the initial integrated circuit chip placement at one or more time steps that are before the corresponding time step, a target value of a reward function that measures a quality of the initial integrated circuit chip placement, or respective target values of the reward function for each initial integrated circuit chip placement at one or more time steps that are before the corresponding time step.

[0227] As an example, the policy model input can include graph embeddings of the netlist, along with embeddings of states, actions, or retums-to-go. In some examples, the target retum-to-go values can be configured to ambitious targets, e.g., using the negative value of wirelength as the reward. The system can thus encourage the policy model to generate high-quality placements with low wirelengths.

[0228] As another example, the data characterizing the connectivity between two or more of the multiple nodes specified by the netlist data can characterize for, some or all of the nodes, one or more other of the nodes to which that node is connected according to the netlist. For example, the policy model input can represent each connection between any two nodes as an edge connecting the two nodes.

[0229] In some examples, the reward function measures certain characteristics of the generated placements that, when optimized, result in a chip that is manufactured using the generated placement exhibiting good performance, e.g., in terms of one or more of power consumption, heat generation, or timing performance.

[0230] In some examples, the reward function can be based on the one or more performance metrics. For example, the system can determine a reward for a complete initial integrated circuit chip placement based on the final estimated wirelength performance metric. Because the ultimate objective of the placement task is to minimize the wirelength, the system can set the target retums-to-go for the policy’ model by using the negative value of the wirelength as the reward. By optimizing for this reward, the system can thus use the policy model to generate placements that score well on the performance metrics.

[0231] Generally, the target value for the reward function specifies a value that should be achieved by a high-quality placement for the current netlist of nodes.

[0232] In some cases, the target value is the same for all netlists that are placed by the system. As one example, the target value can be received as input by the system. As another example, the target value can be determined as a mean, median, or other aggregation measure of reward values for a subset of placements for netlists of nodes in training data used to train the neural network. As one example, the subset can include placements that have been determined to be high-quality placements, e.g., by a user of the system.Attorney Docket No.: 45288-0612WO1

[0233] In some other cases, the target value depends on certain criteria for the placement or for the fabricated chip, e.g., constraints on surface area, power consumption, heat generation and so on. For example, the target value can be determined based on a respective baseline reward value for a baseline placement for the current netlist of nodes, e.g., one determined manually by a human user or by a different, default placer. As one example, the system can determine the target value by selecting a value that is “better’" by a threshold amount than the baseline reward value.

[0234] The system samples from the policy model to generate one or more score distributions (step 504). For example, the system samples from the policy model given the policy model input for the corresponding time step to generate one or more score distributions over multiple positions on the surface of the integrated circuit chip.

[0235] For example, to generate each score distribution, the system can process policy model input using the policy model using a different random seed. As another example, the system can set sampling hyperparameters (e.g., temperature, Top-k, Top-p, etc.) to generate score distributions for the policy model input.

[0236] Each score distribution can include a distribution of logits, each corresponding to one of the multiple positions. For example, each logit corresponds to a grid square of a grid that discretizes the surface of the computer chip. Each logit corresponds to a grid square where the current macro node’s reference point, e.g., bottom left comer, can be placed.

[0237] The system performs the steps 506-508 for each of the one or more score distributions when the set includes a valid suggested region for the macro node.

[0238] For example, the system can determine that the set includes a valid suggested region for the macro node if the set includes a suggested region that does not overlap w ith any macro nodes that have already been placed on the initial integrated circuit chip placement. For example, because the policy model places macro nodes sequentially (e.g., autoregressively) one by one, a region initially suggested by the multimodal neural network might become invalid for a macro node at a later time step if previously placed macro nodes already occupy or overlap with the suggested region.

[0239] When the set includes a valid suggested region for the macro node, the system modifies the score distribution based on the suggested region (step 506).

[0240] For example, the system can apply a mask to one or more logits of the score distribution that correspond to positions that are not within the suggested region. For example, the system can mask, e.g., zero out, the logits for any positions that fall outside of the set of suggested regions specified by the neural network output. The system can thusAttorney Docket No.: 45288-0612WO1

[0241] constrain the policy model to place the macro within the area identified as promising by the multimodal neural network, while still allowing the policy model to retain control over the exact placement coordinates within that region. Furthermore, by using a multimodal neural network to generate suggested regions that constrain the action space of a policy model, the system provides adaptive guidance to steer the placement process without requiring explicit training updates. This interaction optimizes performance metrics such as wire length, routing congestion, and heat generation, thereby addressing specific technical constraints in electronic design automation. Consequently, this approach can fulfill afunctional technical purpose by generating high-quality floorplans that reduce power consumption and increase computing pow er for a given surface area while conserving computational resources.

[0242] The system assigns the macro node to be placed to a position from the plurality of positions using the modified score distribution (step 508).

[0243] As an example, the system can assign the macro node to the position having the highest score in the modified score distribution.

[0244] As another example, the system can sample a position from the modified score distribution. The system can assign the macro node to the sampled position. For example, the system can sample the final position representing a reference point of the node (e.g., the bottom-left comer) from the modified probability distribution. Because the probabilities for all positions outside the suggested regions are zeroed out by the mask before sampling, the system prevents the policy model from selecting a position outside the multimodal neural network's suggested regions. This guided sampling process is formally denoted as

[0245]

[0246] ! ?' ' , illustrating that while the action space is constrained by the valid suggestion ‘ ’’ for the macro node mt, the policy model still retains full control over •p<

[0247] selecting the exact discrete placement coordinates * within that suggested boundary.

[0248] For each of the one or more score distributions, when the set does not include a valid suggested region for the macro node mt, the system can assign the macro node to be placed to a position from the plurality of positions using the score distribution without modifying the score distribution. For example, the system can sample from the unconstrained policy model

[0249]

[0250] As an example, the system can assign the macro node to the position having the highest score in the score distribution.Attorney Docket No.: 45288-0612WO1

[0251] As another example, the system can sample a position from the score distribution. The system can assign the macro node to the sampled position.

[0252] In some implementations, when the set does not include a valid suggested region for the macro node, the system can generate, using the policy model, multiple respective initial integrated circuit chip placements based on the netlist data and any remaining macro nodes from the netlist data to be placed. For example, in some cases the set does not include a valid suggested region for the macro node due to already-placed macro nodes overlapping the suggested region, or due to output token limits of the multi-modal neural network.

[0253] For example, for each of the multiple respective initial integrated circuit chip placements, for each remaining macro node, at a corresponding remaining time step of one or more remaining time steps, the system can generate a remaining policy model input for the corresponding remaining time step. The system can sample from the policy model given the remaining policy model input for the corresponding time step to generate multiple score distributions over the multiple positions on the surface of the integrated circuit chip. For each of the multiple score distributions, the system can assign the remaining macro node to be placed to a position from the multiple positions using the score distribution. The system can thus generate multiple initial integrated circuit chip placements by alternating between multimodal neural network guided rollouts, e.g., constrained by the suggested regions, and unguided rollouts, e.g., sampling from the policy model without guidance from the multimodal neural network.

[0254] The system can thus stochastically roll out copies of the current partial placement for each of the multiple respective initial integrated circuit chip placements using the policy model. The system can increase data coverage for scenarios with invalid suggestions, ensuring the multimodal neural network has a diverse population of initial integrated circuit chip placements to evaluate in future iterations.

[0255] For example, in these instances where a valid suggestion is not available, the system can generate multiple alternative rollouts to increase data coverage. For example, the system can stochastically roll out a predefined number M of copies of the current partial placement Pe' to the end of the placement sequence using the unconstrained policy model. The system calculates the performance metric for each of the M completed alternative placements and adds these pairs to the buffer H. Upon assigning the position ptfor the macro node mt, the system updates the current placement for iteration e, denoted as PePeU {(mt,pt )} . After allAttorney Docket No.: 45288-0612WO1

[0256] macro nodes in the netlist are placed, the system calculates the final performance metric for the completed placement Peand adds the pair to the buffer H.

[0257] FIG. 6 shows another example process 600 for generating a set of integrated circuit chip placements. For convenience, the process 600 will be described as being performed by a system of one or more computers located in one or more locations. For example, a placement generation system, e.g., the placement generation system 100 of FIG. 1, appropriately programmed, can perform the process 600.

[0258] At 602, the system generates a neural network input that includes data representing previously generated initial integrated circuit chip placements retrieved from a buffer.

[0259] At 604, the system processes the neural network input using a multimodal neural network to determine one or more sets of suggested regions for the macro nodes on the surface of the chip.

[0260] At 606, the system uses the suggested regions to provide guidance to a policy model during the placement process performed by the policy model. When a valid suggested region exists for a macro node, the system modifies the policy model's score distribution based on the suggested region, effectively constraining the action space of the policy model. If a valid suggested region is not available, such as when a suggested region overlaps with already-placed macro nodes, the system assigns the macro node's position by sampling from the original, unconstrained policy model. The system continues to use this guided policy model to sequentially place each macro node until it finishes generating a complete initial integrated circuit chip placement.

[0261] At 608, after the placement is finished, the system determines one or more performance metrics to evaluate the quality7of the placement.

[0262] At 610, the system adds the completed placement and the one or more performance metrics into the buffer.

[0263] At 602 for a subsequent iteration, the system then selects placements from the updated buffer to provide historical context to the multimodal neural network in subsequent iterations, creating an evolutionary in-context reinforcement learning feedback loop that continuously improves placement quality.

[0264] FIG. 7 shows the performance of an example chip placement system. For example, FIG. 7 shows the performance of the placement generation system 100 described above with reference to FIG. 1, denoted as “VeoPlace”, across four benchmark integrated circuits. Each of the graphs 710, 720. 730, and 740 show the performance of the chip placement system in terms of a performance metric, e.g., the normalized HPWL ratio, where a lower value isAttorney Docket No.: 45288-0612WO1

[0265] beter, versus the number of iterations in the sequence of iterations over which the system generates a respective set of one or more initial integrated circuit chip placements

[0266] The graphs 710, 720, 730, and 740 demonstrates that the chip placement system described in this specification, denoted by the light gray line, achieves lower wirelengths and beter scaling behavior than the baseline standalone policy model, denoted by the dark gray line, as iterations increase.

[0267] The system thus provides a technological improvement in superior scaling behavior. The system leverages in-context reinforcement learning by providing the multimodal neural network with data representing one or more previously generated initial integrated circuit chip placements so the multimodal neural network can analyze past spatial atempts and iteratively suggest beter regions. Furthermore, by using bounding boxes to constrain the action space of the policy model, the system efficiently steers the optimization process without requiring further training of the policy model or the multimodal neural network. The system can thus generate high-quality7chip placements, e.g., that have reduced power consumption and increased computing power for a given surface area, without conserving computational power and wall clock time.

[0268] FIGS. 8A-8D show an example of a neural network input provided to a multimodal neural network.

[0269] As shown in FIG. 8A, the example neural network input can include high-level task instructions for guiding the placement policy and tables of macro properties (e.g., macro names, corresponding colors, and dimensions).

[0270] As shown in FIG. 8B, the example neural network input can include placement rules (e.g., canvas size, coordinate system definitions, and region constraints) and placement quality metrics (e.g., an instruction indicating that lower wirelength and cost are preferred).

[0271] As shown in FIG. 8C, the example neural network input can include a visual and textual history of one or more previously generated initial integrated circuit chip placements alongside output formatting constraints. For example, the history7can include the exact coordinates from previous placement episodes, the final performance metrics for those episodes, and an image of the canvas where the macro nodes are depicted in their corresponding colors and outlined.

[0272] As shown in FIG. 8D, the example neural network input can include reasoning templates and information defining the current canvas state. For example, the input can prompt the multimodal neural network to conduct a detailed placement history analysis, analyze geometric formations and critical edge alignments, and articulate a holistic placementAttorney Docket No.: 45288-0612WO1

[0273] strategy' before outputting the suggested regions. By processing this combined text and image input, the multimodal neural network can extract structural insights and formulate specific placement strategies.

[0274] In this specification, the term "configured" is used in relation to computing systems and environments, as well as computer program components. A computing system or environment is considered "configured" to perform specific operations or actions when it possesses the necessary software, firmware, hardware, or a combination thereof, enabling it to carry' out those operations or actions during operation. For instance, configuring a system might involve installing a software library' with specific algorithms, updating firmware with new instructions for handling data, or adding a hardware component for enhanced processing capabilities. Similarly, one or more computer programs are "configured" to perform particular operations or actions when they contain instructions that, upon execution by a computing device or hardware, cause the device to perform those intended operations or actions.

[0275] The embodiments and functional operations described in this specification can be implemented in various forms, including digital electronic circuitry, software, firmware, computer hardware (encompassing the disclosed structures and their structural equivalents), or any combination thereof. The subject matter can be realized as one or more computer programs, essentially modules of computer program instructions encoded on a tangible non-transitory storage medium for execution by or to control the operation of a computing device or hardware. The storage medium can be a storage device such as a hard drive or solid-state drive (SSD), a storage medium, a random or serial access memory device, or a combination of these. Additionally or alternatively, the program instructions can be encoded on a transmitted signal, such as a machine-generated electrical, optical, or electromagnetic signal, designed to carry information for transmission to a receiving device or system for execution by a computing device or hardware. Furthermore, implementations may leverage emerging technologies like quantum computing or neuromorphic computing for specific applications, and may be deployed in distributed or cloud-based environments where components reside on different machines or within a cloud infrastructure.

[0276] The term "computing device or hardware" refers to the physical components involved in data processing and encompasses all types of devices and machines used for this purpose. Examples include processors or processing units, computers, multiple processors or computers working together, graphics processing units (GPUs), tensor processing units (TPUs), and specialized processing hardware such as field-programmable gate arraysAttorney Docket No.: 45288-0612WO1

[0277] (FPGAs) or application-specific integrated circuits (ASICs). In addition to hardware, a computing device or hardware may also include code that creates an execution environment for computer programs. This code can take the form of processor firmware, a protocol stack, a database management system, an operating system, or a combination of these elements. Embodiments may particularly benefit from utilizing the parallel processing capabilities of GPUs, in a General-Purpose computing on Graphics Processing Units (GPGPU) context, where code specifically designed for GPU execution, often called kernels or shaders, is employed. Similarly, TPUs excel at running optimized tensor operations crucial for many machine learning algorithms. By leveraging these accelerators and their specialized programming models, the system can achieve significant speedups and efficiency gains for tasks involving artificial intelligence and machine learning, particularly in areas such as computer vision, natural language processing, and robotics.

[0278] A computer program, also referred to as software, an application, a module, a script, code, or simply a program, can be written in any programming language, including compiled or interpreted languages, and declarative or procedural languages. It can be deployed in various forms, such as a standalone program, a module, a component, a subroutine, or any other unit suitable for use within a computing environment. A program may or may not correspond to a single file in a file system and can be stored in various ways. This includes being embedded within a file containing other programs or data (e.g., scripts within a markup language document), residing in a dedicated file, or distributed across multiple coordinated files (e.g., files storing modules, subprograms, or code segments). A computer program can be executed on a single computer or across multiple computers, whether located at a single site or distributed across multiple sites and interconnected through a data communication network. The specific implementation of the computer programs may involve a combination of traditional programming languages and specialized languages or libraries designed for GPGPU programming or TPU utilization, depending on the chosen hardware platform and desired performance characteristics.

[0279] In this specification, the term "engine" broadly refers to a software-based system, subsystem, or process designed to perform one or more specific functions. An engine is typically implemented as one or more software modules or components installed on one or more computers, which can be located at a single site or distributed across multiple locations. In some instances, one or more dedicated computers may be used for a particular engine, while in other cases, multiple engines may operate concurrently on the same one or more computers. Examples of engine functions within the context of Al and machine learningAttorney Docket No.: 45288-0612WO1

[0280] could include data pre-processing and cleaning, feature engineering and extraction, model training and optimization, inference and prediction generation, and post-processing of results. The specific design and implementation of engines will depend on the overall architecture and the distribution of computational tasks across various hardware components, including CPUs, GPUs, TPUs, and other specialized processors.

[0281] The processes and logic flows described in this specification can be executed by one or more programmable computers running one or more computer programs to perform functions by operating on input data and generating output. Additionally, graphics processing units (GPUs) and tensor processing units (TPUs) can be utilized to enable concurrent execution of aspects of these processes and logic flows, significantly accelerating performance. This approach offers significant advantages for computationally intensive tasks often found in Al and machine learning applications, such as matrix multiplications, convolutions, and other operations that exhibit a high degree of parallelism. By leveraging the parallel processing capabilities of GPUs and TPUs, significant speedups and efficiency¬ gains compared to relying solely on CPUs can be achieved. Alternatively or in combination with programmable computers and specialized processors, these processes and logic flows can also be implemented using specialized processing hardware, such as field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs), for even greater performance or energy efficiency in specific use cases.

[0282] Computers capable of executing a computer program can be based on general-purpose microprocessors, special-purpose microprocessors, or a combination of both. They can also utilize any other type of central processing unit (CPU). Additionally, graphics processing units (GPUs), tensor processing units (TPUs), and other machine learning accelerators can be employed to enhance performance, particularly for tasks involving artificial intelligence and machine learning. These accelerators often work in conjunction with CPUs, handling specialized computations while the CPU manages overall system operations and other tasks. Typically, a CPU receives instructions and data from read-only memory7(ROM), random access memory (RAM), or both. The elements of a computer include a CPU for executing instructions and one or more memory devices for storing instructions and data. The specific configuration of processing units and memory will depend on factors like the complexity7of the Al model, the volume of data being processed, and the desired performance and latency requirements. Embodiments can be implemented on a wide range of computing platforms, from small embedded devices with limited resources to large-scale data center systems withAttorney Docket No.: 45288-0612WO1

[0283] high-performance computing capabilities. The system may include storage devices like hard drives, SSDs, or flash memory for persistent data storage.

[0284] Computer-readable media suitable for storing computer program instructions and data encompass all forms of non-volatile memory, media, and memory devices. Examples include semiconductor memory7devices such as read-only memory7(ROM), solid-state drives (SSDs), and flash memory devices; hard disk drives (HDDs); optical media; and optical discs such as CDs. DVDs, and Blu-ray discs. The specific type of computer-readable media used will depend on factors such as the size of the data, access speed requirements, cost considerations, and the desired level of portability7or permanence.

[0285] To facilitate user interaction, embodiments of the subject matter described in this specification can be implemented on a computing device equipped with a display device, such as a liquid crystal display (LCD) or an organic light-emitting diode (OLED) display, for presenting information to the user. Input can be provided by the user through various means, including a keyboard), touchscreens, voice commands, gesture recognition, or other input modalities depending on the specific device and application. Additional input methods can include acoustic, speech, or tactile input, while feedback to the user can take the form of visual, auditory7, or tactile feedback. Furthermore, computers can interact wi th users by7exchanging documents with a user's device or application. This can involve sending web content or data in response to requests or sending and receiving text messages or other forms of messages through mobile devices or messaging platforms. The selection of input and output modalities will depend on the specific application and the desired form of user interaction.

[0286] Machine learning models can be implemented and deployed using machine learning frameworks, such as TensorFlow or JAX. These frameworks offer comprehensive tools and libraries that facilitate the development, training, and deployment of machine learning models.

[0287] Embodiments of the subject matter described in this specification can be implemented within a computing system comprising one or more components, depending on the specific application and requirements. These may include a back-end component, such as a back-end server or cloud-based infrastructure; an optional middleware component, such as a middleware server or application programming interface (API), to facilitate communication and data exchange; and a front-end component, such as a client device with a user interface, a web browser, or an app, through which a user can interact with the implemented subject matter. For instance, the described functionality could be implemented solely on a clientAttorney Docket No.: 45288-0612WO1

[0288] device (e g., for on-device machine learning) or deployed as a combination of front-end and back-end components for more complex applications. These components, when present, can be interconnected using any form or medium of digital data communication, such as a communication network like a local area network (LAN) or a wide area network (WAN) including the Internet. The specific system architecture and choice of components will depend on factors such as the scale of the application, the need for real-time processing, data security requirements, and the desired user experience.

[0289] The computing system can include clients and servers that may be geographically separated and interact through a communication network. The specific ty pe of network, such as a local area network (LAN), a wide area network (WAN), or the Internet, will depend on the reach and scale of the application. The client-server relationship is established through computer programs running on the respective computers and designed to communicate with each other using appropriate protocols. These protocols may include HTTP, TCP / IP, or other specialized protocols depending on the nature of the data being exchanged and the security requirements of the system. In certain embodiments, a server transmits data or instructions to a user's device, such as a computer, smartphone, or tablet, acting as a client. The client device can then process the received information, display results to the user, and potentially send data or feedback back to the server for further processing or storage. This allows for dynamic interactions between the user and the system, enabling a wide range of applications and functionalities.

[0290] While this specification contains many specific implementation details, these should not be construed as limitations on the scope of any invention or on the scope of what may be claimed, but rather as descriptions of features that may be specific to particular embodiments of particular inventions. Certain features that are described in this specification in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially be claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.

[0291] Similarly, while operations are depicted in the drawings and recited in the claims in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations beAttorney Docket No.: 45288-0612WO1

[0292] performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system modules and components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0293] Particular embodiments of the subject matter have been described. Other embodiments are within the scope of the following claims. For example, the actions recited in the claims can be performed in a different order and still achieve desirable results. As one example, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some cases, multitasking and parallel processing may be advantageous.

[0294] Aspects of the present disclosure may be as set out in the following clauses:

[0295] Clause 1. A method performed by one or more computers, the method comprising: obtaining netlist data for an integrated circuit chip, wherein the netlist data specifies a connectivity on the integrated circuit chip between a plurality of nodes that each correspond to one or more of a plurality of integrated circuit components of the integrated circuit chip, wherein the plurality of nodes comprise a plurality of macro nodes that each represent a respective macro component of the integrated circuit chip; and

[0296] generating a set of one or more integrated circuit chip placements for the integrated circuit chip, wherein each integrated circuit chip placement places each macro node in the netlist data at a respective position on a surface of the integrated circuit chip, comprising:

[0297] generating a respective set of one or more initial integrated circuit chip placements at each of a sequence of iterations, comprising, at each of one or more particular iterations in the sequence:

[0298] determining, using a multimodal neural network, one or more respective sets of suggested regions on the surface of the integrated circuit chip, wherein each respective set of suggested regions comprises a respective suggested region on the surface of the integrated circuit chip for each of one or more of the macro nodes in the netlist data;

[0299] generating, for each set of suggested regions and using a policy model, one or more respective initial integrated circuit chip placements based on the netlist data and the set of suggested regions; and

[0300] determining, for each respective initial integrated circuit chipAttorney Docket No.: 45288-0612WO1

[0301] placement, one or more performance metrics that each measure a quality of the initial integrated circuit chip placement.

[0302] Clause 2. The method of clause 1, wherein each respective initial integrated circuit chip placement places each macro node in the netlist data at a respective position on the surface of the integrated circuit chip.

[0303] Clause 3. The method of any one of clauses 1-2, wherein determining, using a multimodal neural network, one or more respective sets of suggested regions on the surface of the integrated circuit chip comprises:

[0304] generating a neural network input for the particular iteration that comprises data representing one or more initial integrated circuit chip placements generated at one or more preceding iterations in the sequence; and

[0305] sampling from the multimodal neural network given the neural network input for the particular iteration to generate one or more neural network outputs that each specify a respective set of suggested regions.

[0306] Clause 4. The method of clause 3, wherein the neural network input for the particular iteration comprises one or more respective performance metrics for each of the one or more initial integrated circuit chip placements.

[0307] Clause 5. The method of any preceding clause, wherein each of the one or more respective performance metrics measures one or more of: w ire length of wires between nodes on the surface of the chip in the initial integrated circuit chip placement, a cost for the initial integrated circuit chip placement, or routing congestion measuring available wiring resources in a given region on the chip compared to wires that run through the given region in the initial integrated circuit placement.

[0308] Clause 6. The method of any one of clauses 3-5, w herein the data representing one or more initial integrated circuit chip placements comprises, for each initial integrated circuit chip placement, one or more of: a respective set of coordinates for each macro node that identifies the respective position of the macro node on the surface of the chip in the initial integrated circuit chip placement, or an image depicting each macro node at the respective position of the macro node on the surface of the integrated circuit chip in the initial integrated circuit chip placement.

[0309] Clause 7. The method of clause 6, wherein the image depicts each macro node in a corresponding color.Attorney Docket No.: 45288-0612WO1

[0310] Clause 8. The method of clause 7, further comprising determining the corresponding color for each macro node, comprising:

[0311] generating a respective embedded representation for each macro node in the netlist data;

[0312] generating a plurality of clusters for the macro nodes using the respective embedded representations;

[0313] assigning a respective color to each cluster; and

[0314] for each cluster, assigning the respective color for the cluster to each macro node of the cluster.

[0315] Clause 9. The method of clause 8, wherein generating a respective embedded representation for each macro node in the netlist data comprises:

[0316] generating a connectivity graph using the netlist data; and

[0317] processing the connectivity’ graph to generate the respective embedded representation for each macro node in the netlist data.

[0318] Clause 10. The method of any one of clauses 3-9, wherein generating the neural network input for the particular iteration comprises selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations.

[0319] Clause 11. The method of clause 10, wherein selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations comprises selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at one or more immediately preceding iterations.

[0320] Clause 12. The method of clause 10. wherein selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations comprises randomly sampling the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations.

[0321] Clause 13. The method of clause 10. wherein selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations comprises selecting the one or more initial integrated circuit chip placements based on the one or more performance metrics for the initial integrated circuit chip placements generated at preceding iterations.Attorney Docket No.: 45288-0612WO1

[0322] Clause 14. The method of clause 10, wherein selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations comprises:

[0323] clustering each initial integrated circuit chip placement generated at each preceding iteration into one or more clusters; and

[0324] selecting, from each cluster, one or more initial integrated circuit chip placements. Clause 15. The method of clause 14. wherein clustering each initial integrated circuit chip placement generated at each preceding iteration into one or more clusters comprises:

[0325] obtaining, for each initial integrated circuit chip placement, a corresponding set of coordinates for each macro node that identifies the respective position of the macro node on the surface of the chip in the initial integrated circuit chip placement; and

[0326] clustering the initial integrated circuit chip placements into the one or more clusters based on the corresponding sets of coordinates.

[0327] Clause 16. The method of any one of clauses 14-15, wherein selecting, from each cluster, one or more initial integrated circuit chip placements comprises selecting, from each cluster, the one or more initial integrated circuit chip placements based on the one or more performance metrics.

[0328] Clause 17. The method of clause 16, wherein selecting, from each cluster, the one or more initial integrated circuit chip placements based on the one or more performance metrics comprises:

[0329] selecting, from each cluster and for each of one or more of the performance metrics, one or more initial integrated circuit chip placements from the cluster that each have a highest value for the performance metric.

[0330] Clause 18. The method of clause 10. wherein selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations comprises:

[0331] clustering each initial integrated circuit chip placement generated at each preceding iteration into one or more clusters;

[0332] selecting a cluster based on a ranking for the clusters; and

[0333] selecting, using the selected cluster, one or more initial integrated circuit chip placements.

[0334] Clause 19. The method of clause 18, wherein clustering each initial integrated circuit chip placement generated at each preceding iteration into one or more clusters comprises:

[0335] obtaining, for each initial integrated circuit chip placement, a corresponding set ofAttorney Docket No.: 45288-0612WO1

[0336] coordinates for each macro node that identifies the respective position of the macro node on the surface of the chip in the initial integrated circuit chip placement; and

[0337] clustering the initial integrated circuit chip placements into the one or more clusters based on the corresponding sets of coordinates.

[0338] Clause 20. The method of any one of clauses 18-19, wherein selecting a cluster based on a ranking for the clusters comprises:

[0339] identifying a respective feature value for each cluster; and

[0340] determining the ranking for the clusters based on the respective feature value for each cluster.

[0341] Clause 21. The method of any one of clauses 18-20, wherein selecting a cluster based on a ranking for the clusters comprises selecting a cluster according to a softmax distribution over the ranking.

[0342] Clause 22. The method of any one of clauses 20-21, wherein the respective feature value comprises a minimum wire length of wires between nodes on the surface of the chip in the initial integrated circuit chip placement over each initial integrated circuit chip placement in the cluster.

[0343] Clause 23. The method of any one of clauses 18-22, wherein selecting, using the selected cluster, one or more initial integrated circuit chip placements comprises selecting, using the selected cluster, one or more initial integrated circuit chip placements based on one or more performance metrics.

[0344] Clause 24. The method of clause 23, wherein selecting, using the selected cluster, one or more initial integrated circuit chip placements based on one or more performance metrics comprises selecting, for each of one or more of the performance metrics, one or more initial integrated circuit chip placements from the selected cluster that each have a highest value for the performance metric.

[0345] Clause 25. The method of any one of clauses 23-24, further comprising:

[0346] selecting one or more second clusters that are closest to the selected cluster; and selecting, for each of the one or more selected second clusters and for each of one or more of the performance metrics, one or more initial integrated circuit chip placements from the selected second cluster that each have a highest value for the performance metric.

[0347] Clause 26. The method of any of clauses 3-25, wherein the neural network input for the particular iteration comprises an instruction to suggest bounding box regions for placing each of one or more macro nodes on the surface of the integrated circuit chip.Attorney Docket No.: 45288-0612WO1

[0348] Clause 27. The method of clause 26, wherein the instruction comprises one or more of: data representing the netlist data, a description of a size of the surface of the integrated circuit chip, a description of a coordinate system for the surface of the integrated circuit chip, a dimension for each of the one or more macro components, an orientation for each of the one or more macro components, a color for each of the one or more macro nodes, data representing a current placement, a request to describe the suggested regions, a request to explain the suggested regions, a request to describe a placement strategy, a request to generate the suggested regions according to a macro node order, a request to analyze the one or more initial integrated circuit chip placements included in the neural network input, a description of an output format, one or more output examples for representing each suggested region, code representing the integrated circuit, or a description of one or more performance metrics.

[0349] Clause 28. The method of clause 27, wherein the data representing a current placement comprises one or more of: a respective set of coordinates for each of one or more macro nodes that identifies a respective current position of the macro node on the surface of the chip in the current placement, or an image depicting one or more macro nodes at the respective current position of the macro node on the surface of the integrated circuit chip in the current placement.

[0350] Clause 29. The method of any one of clauses 27-28, wherein the data representing the netlist data comprises, for each macro node in the netlist data, one or more of: an identifier for the macro node, a corresponding color for the macro node, or a dimension for the macro node.

[0351] Clause 30. The method of any of clauses 3-29, wherein each of the one or more neural network outputs comprises data representing, for each suggested region in the respective set of suggested regions, a respective bounding box identifying the suggested region.

[0352] Clause 31. The method of clause 30, wherein the data representing, for each suggested region in the respective set of suggested regions, a respective bounding box identifying the suggested region comprises one or more of: a set of coordinates specifying the respective bounding boxes, or an image depicting the respective bounding boxes.

[0353] Clause 32. The method of any of clauses 3-31, wherein the neural network output comprises, for each suggested region in the respective set of suggested regions, an explanation for the suggested region.

[0354] Clause 33. The method of any preceding clause, wherein generating, for each set of suggested regions and using a policy model, one or more respective initial integrated circuitAttorney Docket No.: 45288-0612WO1

[0355] chip placements based on the netlist data and the set of suggested regions comprises, for each set of suggested regions:

[0356] for each macro node in the netlist data, at a corresponding time step of a plurality of time steps:

[0357] generating a policy model input for the corresponding time step; sampling from the policy model given the policy model input for the corresponding time step to generate one or more score distributions over a plurality of positions on the surface of the integrated circuit chip;

[0358] for each of the one or more score distributions:

[0359] when the set includes a valid suggested region for the macro node, modifying the score distribution based on the suggested region; and

[0360] assigning the macro node to be placed to a position from the plurality of positions using the modified score distribution.

[0361] Clause 34. The method of clause 33, wherein the score distribution comprises a distribution of logits, each corresponding to one of the plurality of positions, and wherein modifying the score distribution based on the suggested region comprises applying a mask to one or more logits of the score distribution that correspond to positions that are not within the suggested region.

[0362] Clause 35. The method of any one of clauses 33-34, wherein assigning the macro node to be placed to a position from the plurality of positions using the modified score distribution comprises:

[0363] assigning the macro node to the position having the highest score in the modified score distribution.

[0364] Clause 36. The method of any one of clauses 33-34. wherein assigning the macro node to be placed to a position from the plurality' of positions using the modified score distribution comprises:

[0365] sampling a position from the modified score distribution, and

[0366] assigning the macro node to the sampled position.

[0367] Clause 37. The method of clause 33. further comprising, for each of the one or more score distributions, when the set does not include a valid suggested region for the macro node, assigning the macro node to be placed to a position from the plurality of positions using the score distribution without modify ing the score distribution.

[0368] Clause 38. The method of clause 37. wherein assigning the macro node to be placed to a position from the plurality of positions using the score distribution comprises:Attorney Docket No.: 45288-0612WO1

[0369] assigning the macro node to the position having the highest score in the score distribution.

[0370] Clause 39. The method of clause 37, wherein assigning the macro node to be placed to a position from the plurality of positions using the score distribution comprises:

[0371] sampling a position from the score distribution, and

[0372] assigning the macro node to the sampled position.

[0373] Clause 40. The method of any one of clauses 37-39. further comprising, for each of the one or more score distributions, generating, using the policy model, a plurality of respective initial integrated circuit chip placements based on the netlist data and any remaining macro nodes from the netlist data to be placed.

[0374] Clause 41. The method of clause 40. wherein generating, using the policy model, a plurality of respective initial integrated circuit chip placements based on the netlist data and any remaining macro nodes from the netlist data to be placed comprises:

[0375] for each of the plurality' of respective initial integrated circuit chip placements:

[0376] for each remaining macro node, at a corresponding remaining time step of one or more remaining time steps:

[0377] generating a remaining policy model input for the corresponding remaining time step;

[0378] sampling from the policy model given the remaining policy model input for the corresponding time step to generate a plurality of score distributions over the plurality of positions on the surface of the integrated circuit chip; and

[0379] for each of the plurality7of score distributions:

[0380] assigning the remaining macro node to be placed to a position from the plurality of positions using the score distribution.

[0381] Clause 42. The method of any preceding clause, wherein the policy model is configured to generate the initial integrated circuit chip placement autoregressively by placing the macro nodes over the plurality7of time steps.

[0382] Clause 43. The method of clause 42, wherein the policy model is configured to place each macro node at a position at a different corresponding time step according to a macro node order.

[0383] Clause 44. The method of clause 43, wherein the policy model input for the corresponding time step comprises one or more of: a target value of a reward function that measures a quality of the initial integrated circuit chip placement, data characterizingAttorney Docket No.: 45288-0612WO1

[0384] respective positions on the surface of the chip of any macro nodes that are before a particular macro node to be placed at the corresponding time step in the macro node order, data characterizing the particular macro node to be placed at the corresponding time step, data characterizing the connectivity7between two or more of the plurality of nodes specified by the netlist data, data representing the netlist data, or metadata for the netlist data.

[0385] Clause 45. The method of clause 43, wherein the policy model input for the corresponding time step comprises one or more of: data representing the netlist data, data characterizing respective positions on the surface of the chip of any macro nodes that are before a particular macro node to be placed at the corresponding time step in the macro node order, data characterizing the particular macro node to be placed at the corresponding time step, data characterizing the initial integrated circuit chip placement at one or more time steps that are before the corresponding time step, a target value of a reward function that measures a quality of the initial integrated circuit chip placement, or respective target values of the reward function for each initial integrated circuit chip placement at one or more time steps that are before the corresponding time step.

[0386] Clause 46. The method of any preceding clause, wherein generating a respective set of one or more initial integrated circuit chip placements at each of a sequence of iterations comprises:

[0387] at each of one or more initial iterations that precede a first particular iteration in the sequence:

[0388] generating, using the policy model, one respective initial integrated circuit chip placement for the initial iteration without using any suggested regions generated by the multimodal neural network.

[0389] Clause 47. The method of any preceding clause, wherein generating a respective set of one or more initial integrated circuit chip placements at each of a sequence of iterations comprises:

[0390] at each of one or more intermediate iterations that are after a first particular iteration in the sequence:

[0391] generating, using the policy7model, one or more respective initial integrated circuit chip placements for the intermediate iteration based on the one or more sets of suggested regions generated at a most recent particular iteration.

[0392] Clause 48. The method of any preceding clause, wherein the policy model has been trained on a plurality of training examples to generate integrated circuit chip placements that optimize for one or more criteria.Attorney Docket No.: 45288-0612WO1

[0393] Clause 49. The method of clause 48, wherein the one or more criteria measure one or more of: a degree of overlap of the macro nodes in the initial integrated circuit chip placement, wire length of wires between nodes on the surface of the chip in the initial integrated circuit chip placement, routing congestion measuring available wiring resources in a given region on the chip compared to wires that run through the given region in the initial integrated circuit chip placement, a cost for the initial integrated circuit chip placement, power consumption for the chip, heat generation for the chip, area of the initial integrated circuit chip placement, or timing of the initial integrated circuit chip placement.

[0394] Clause 50. The method of any preceding clause, further comprising, for one or more initial integrated circuit chip placements of the respective sets of initial integrated circuit chip placements, generating a respective training example comprising the initial integrated circuit chip placement.

[0395] Clause 51. The method of clause 50, further comprising training the policy model using the respective training examples.

[0396] Clause 52. The method of any preceding clause, wherein the multimodal neural network comprises a visual language model neural network.

[0397] Clause 53. The method of any preceding clause, further comprising training the multimodal neural network on training examples that each comprise a training initial integrated circuit chip placement.

[0398] Clause 54. The method of any preceding clause, further comprising training the multimodal neural network on training examples that each comprise data representing the netlist data.

[0399] Clause 55. The method of any preceding clause, wherein the respective positions comprise grid squares from a grid overlaid over the surface of the integrated circuit chip. Clause 56. The method of any preceding clause, further comprising selecting an initial integrated circuit chip placement from the respective sets of initial integrated circuit chip placements.

[0400] Clause 57. The method of any preceding clause, further comprising, for one or more of the respective initial integrated circuit chip placements:

[0401] outputting data specifying the initial integrated circuit chip placement.

[0402] Clause 58. The method of clause 57, wherein outputting data specifying the initial integrated circuit chip placement comprises outputting the initial integrated circuit chip placement to a component of an electronic design automation (EDA) software tool.Attorney Docket No.: 45288-0612WO1

[0403] Clause 59. The method of clause 57 or clause 58, wherein outputting data specifying the initial integrated circuit chip placement comprises providing the data specifying the initial integrated circuit chip placement for presentation to a user on a user device.

[0404] Clause 60. The method of any one of clauses 57-59, wherein outputting data specifying the initial integrated circuit chip placement comprises providing the initial integrated circuit chip placement for fabrication of an integrated circuit chip according to the initial integrated circuit chip placement.

[0405] Clause 61. The method of any preceding clause, further comprising, for one or more of the set of initial integrated circuit chip placements:

[0406] fabricating an integrated circuit chip according to the initial integrated circuit chip placement.

[0407] Clause 62. The method of any preceding clause, wherein each integrated circuit chip placement places only each macro node in the netlist data at a respective position on the surface of the integrated circuit chip.

[0408] Clause 63. The method of any preceding clause, wherein the plurality of the nodes comprise a plurality of standard cell nodes that each represent a respective standard cell component of the integrated circuit chip.

[0409] Clause 64. The method of clause 63, wherein each integrated circuit chip placement places each macro node in the netlist data at a respective position on the surface of the integrated circuit chip and each standard cell node at a respective position on the surface of the integrated circuit chip.

[0410] Clause 65. The method of clause 63, wherein generating a set of one or more integrated circuit chip placements for the integrated circuit chip further comprises, for one or more of the respective initial integrated circuit chip placements, placing each of the standard cell nodes at a respective position on the surface of the integrated circuit chip that includes the macro components represented by the macro nodes placed according to the respective initial integrated circuit chip placement.

[0411] Clause 66. The method of any one of clauses 64-65, further comprising, for one or more of the set of integrated circuit chip placements:

[0412] outputting data specifying the integrated circuit chip placement.

[0413] Clause 67. The method of clause 66, wherein outputting data specifying the integrated circuit chip placement comprises outputting the integrated circuit chip placement to a component of an electronic design automation (EDA) software tool.Attorney Docket No.: 45288-0612WO1

[0414] Clause 68. The method of clause 66 or clause 67, wherein outputting data specifying the integrated circuit chip placement comprises providing the data specifying the integrated circuit chip placement for presentation to a user on a user device.

[0415] Clause 69. The method of any one of clauses 66-68, wherein outputting data specifying the integrated circuit chip placement comprises providing the integrated circuit chip placement for fabrication of an integrated circuit chip according to the integrated circuit chip placement.

[0416] Clause 70. The method of any preceding clause, further comprising, for one or more of the set of integrated circuit chip placements:

[0417] fabricating an integrated circuit chip according to the integrated circuit chip placement.

[0418] Clause 71. An integrated circuit chip having an integrated circuit chip placement generated by performing the operations of the respective method of any preceding clause. Clause 72. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform the operations of the respective method of any one of clauses 1-70. Clause 73. One or more computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform the operations of the respective method of any one of clauses 1-70.

[0419] Clause 74. A computer program product comprising instructions that when executed by one or more processors cause the one or more processors to perform the operations of the respective method of any one of clauses 1-70.

[0420] What is claimed is:

Claims

Attorney Docket No.: 45288-0612WO1CLAIMS1. A method performed by one or more computers, the method comprising:obtaining netlist data for an integrated circuit chip, wherein the netlist data specifies a connectivity on the integrated circuit chip between a plurality of nodes that each correspond to one or more of a plurality7of integrated circuit components of the integrated circuit chip, wherein the plurality of nodes comprise a plurality of macro nodes that each represent a respective macro component of the integrated circuit chip; andgenerating a set of one or more integrated circuit chip placements for the integrated circuit chip, wherein each integrated circuit chip placement places each macro node in the netlist data at a respective position on a surface of the integrated circuit chip, comprising:generating a respective set of one or more initial integrated circuit chip placements at each of a sequence of iterations, comprising, at each of one or more particular iterations in the sequence:determining, using a multimodal neural network, one or more respective sets of suggested regions on the surface of the integrated circuit chip, wherein each respective set of suggested regions comprises a respective suggested region on the surface of the integrated circuit chip for each of one or more of the macro nodes in the netlist data;generating, for each set of suggested regions and using a policy model, one or more respective initial integrated circuit chip placements based on the netlist data and the set of suggested regions; anddetermining, for each respective initial integrated circuit chip placement, one or more performance metrics that each measure a quality of the initial integrated circuit chip placement.

2. The method of claim 1, wherein each respective initial integrated circuit chip placement places each macro node in the netlist data at a respective position on the surface of the integrated circuit chip.

3. The method of any one of claims 1-2, wherein determining, using a multimodal neural network, one or more respective sets of suggested regions on the surface of the integrated circuit chip comprises:generating a neural network input for the particular iteration that comprises data representing one or more initial integrated circuit chip placements generated at one or moreAttorney Docket No.: 45288-0612WO1preceding iterations in the sequence; andsampling from the multimodal neural network given the neural network input for the particular iteration to generate one or more neural network outputs that each specify a respective set of suggested regions.

4. The method of claim 3, wherein the neural network input for the particular iteration comprises one or more respective performance metrics for each of the one or more initial integrated circuit chip placements.

5. The method of any preceding claim, wherein each of the one or more respective performance metrics measures one or more of: wire length of wires between nodes on the surface of the chip in the initial integrated circuit chip placement, a cost for the initial integrated circuit chip placement, or routing congestion measuring available wiring resources in a given region on the chip compared to wires that run through the given region in the initial integrated circuit placement.

6. The method of any one of claims 3-5, wherein the data representing one or more initial integrated circuit chip placements comprises, for each initial integrated circuit chip placement, one or more of: a respective set of coordinates for each macro node that identifies the respective position of the macro node on the surface of the chip in the initial integrated circuit chip placement, or an image depicting each macro node at the respective position of the macro node on the surface of the integrated circuit chip in the initial integrated circuit chip placement.

7. The method of claim 6, wherein the image depicts each macro node in a corresponding color.

8. The method of claim 7, further comprising determining the corresponding color for each macro node, comprising:generating a respective embedded representation for each macro node in the netlist data;generating a plurality'- of clusters for the macro nodes using the respective embedded representations;assigning a respective color to each cluster; andAttorney Docket No.: 45288-0612WO1for each cluster, assigning the respective color for the cluster to each macro node of the cluster.

9. The method of claim 8, wherein generating a respective embedded representation for each macro node in the netlist data comprises:generating a connectivity graph using the netlist data; andprocessing the connectivity graph to generate the respective embedded representation for each macro node in the netlist data.

10. The method of any one of claims 3-9, wherein generating the neural network input for the particular iteration comprises selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations.

11. The method of claim 10, wherein selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations comprises selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at one or more immediately preceding iterations.

12. The method of claim 10, wherein selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations comprises randomly sampling the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations.

13. The method of claim 10, wherein selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations comprises selecting the one or more initial integrated circuit chip placements based on the one or more performance metrics for the initial integrated circuit chip placements generated at preceding iterations.

14. The method of claim 10, wherein selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at precedingAttorney Docket No.: 45288-0612WO1iterations comprises:clustering each initial integrated circuit chip placement generated at each preceding iteration into one or more clusters; andselecting, from each cluster, one or more initial integrated circuit chip placements.

15. The method of claim 14, wherein clustering each initial integrated circuit chip placement generated at each preceding iteration into one or more clusters comprises:obtaining, for each initial integrated circuit chip placement, a corresponding set of coordinates for each macro node that identifies the respective position of the macro node on the surface of the chip in the initial integrated circuit chip placement; andclustering the initial integrated circuit chip placements into the one or more clusters based on the corresponding sets of coordinates.

16. The method of any one of claims 14-15, wherein selecting, from each cluster, one or more initial integrated circuit chip placements comprises selecting, from each cluster, the one or more initial integrated circuit chip placements based on the one or more performance metrics.

17. The method of claim 16, wherein selecting, from each cluster, the one or more initial integrated circuit chip placements based on the one or more performance metrics comprises:selecting, from each cluster and for each of one or more of the performance metrics, one or more initial integrated circuit chip placements from the cluster that each have a highest value for the performance metric.

18. The method of claim 10, wherein selecting the one or more initial integrated circuit chip placements from the initial integrated circuit chip placements generated at preceding iterations comprises:clustering each initial integrated circuit chip placement generated at each preceding iteration into one or more clusters;selecting a cluster based on a ranking for the clusters; andselecting, using the selected cluster, one or more initial integrated circuit chip placements.Attorney Docket No.: 45288-0612WO119. The method of claim 18, wherein clustering each initial integrated circuit chip placement generated at each preceding iteration into one or more clusters comprises:obtaining, for each initial integrated circuit chip placement, a corresponding set of coordinates for each macro node that identifies the respective position of the macro node on the surface of the chip in the initial integrated circuit chip placement; andclustering the initial integrated circuit chip placements into the one or more clusters based on the corresponding sets of coordinates.

20. The method of any one of claims 18-19, wherein selecting a cluster based on a ranking for the clusters comprises:identifying a respective feature value for each cluster; anddetermining the ranking for the clusters based on the respective feature value for each cluster.

21. The method of any one of claims 18-20, wherein selecting a cluster based on a ranking for the clusters comprises selecting a cluster according to a softmax distribution over the ranking.

22. The method of any one of claims 20-21, wherein the respective feature value comprises a minimum wire length of wires between nodes on the surface of the chip in the initial integrated circuit chip placement over each initial integrated circuit chip placement in the cluster.

23. The method of any one of claims 18-22, wherein selecting, using the selected cluster, one or more initial integrated circuit chip placements comprises selecting, using the selected cluster, one or more initial integrated circuit chip placements based on one or more performance metrics.

24. The method of claim 23, wherein selecting, using the selected cluster, one or more initial integrated circuit chip placements based on one or more performance metrics comprises selecting, for each of one or more of the performance metrics, one or more initial integrated circuit chip placements from the selected cluster that each have a highest value for the performance metric.Attorney Docket No.: 45288-0612WO125. The method of any one of claims 23-24, further comprising:selecting one or more second clusters that are closest to the selected cluster; and selecting, for each of the one or more selected second clusters and for each of one or more of the performance metrics, one or more initial integrated circuit chip placements from the selected second cluster that each have a highest value for the performance metric.

26. The method of any of claims 3-25, wherein the neural network input for the particular iteration comprises an instruction to suggest bounding box regions for placing each of one or more macro nodes on the surface of the integrated circuit chip.

27. The method of claim 26, wherein the instruction comprises one or more of: data representing the netlist data, a description of a size of the surface of the integrated circuit chip, a description of a coordinate system for the surface of the integrated circuit chip, a dimension for each of the one or more macro components, an orientation for each of the one or more macro components, a color for each of the one or more macro nodes, data representing a current placement, a request to describe the suggested regions, a request to explain the suggested regions, a request to describe a placement strategy7, a request to generate the suggested regions according to a macro node order, a request to analyze the one or more initial integrated circuit chip placements included in the neural network input, a description of an output format, one or more output examples for representing each suggested region, code representing the integrated circuit, or a description of one or more performance metrics.

28. The method of claim 27, wherein the data representing a current placement comprises one or more of: a respective set of coordinates for each of one or more macro nodes that identifies a respective current position of the macro node on the surface of the chip in the current placement, or an image depicting one or more macro nodes at the respective current position of the macro node on the surface of the integrated circuit chip in the current placement.

29. The method of any one of claims 27-28, wherein the data representing the netlist data comprises, for each macro node in the netlist data, one or more of: an identifier for the macro node, a corresponding color for the macro node, or a dimension for the macro node.Attorney Docket No.: 45288-0612WO130. The method of any of claims 3-29. wherein each of the one or more neural network outputs comprises data representing, for each suggested region in the respective set of suggested regions, a respective bounding box identifying the suggested region.

31. The method of claim 30, wherein the data representing, for each suggested region in the respective set of suggested regions, a respective bounding box identifying the suggested region comprises one or more of: a set of coordinates specifying the respective bounding boxes, or an image depicting the respective bounding boxes.

32. The method of any of claims 3-31. wherein the neural network output comprises, for each suggested region in the respective set of suggested regions, an explanation for the suggested region.

33. The method of any preceding claim, wherein generating, for each set of suggested regions and using a policy model, one or more respective initial integrated circuit chip placements based on the netlist data and the set of suggested regions comprises, for each set of suggested regions:for each macro node in the netlist data, at a corresponding time step of a plurality of time steps:generating a policy model input for the corresponding time step; sampling from the policy model given the policy model input for the corresponding time step to generate one or more score distributions over a plurality of positions on the surface of the integrated circuit chip;for each of the one or more score distributions:when the set includes a valid suggested region for the macro node, modifying the score distribution based on the suggested region; andassigning the macro node to be placed to a position from the plurality of positions using the modified score distribution.

34. The method of claim 33, wherein the score distribution comprises a distribution of logits, each corresponding to one of the plurality of positions, and wherein modifying the score distribution based on the suggested region comprises applying a mask to one or moreAttorney Docket No.: 45288-0612WO1logits of the score distribution that correspond to positions that are not within the suggested region.

35. The method of any one of claims 33-34, wherein assigning the macro node to be placed to a position from the plurality of positions using the modified score distribution comprises:assigning the macro node to the position having the highest score in the modified score distribution.

36. The method of any one of claims 33-34, wherein assigning the macro node to be placed to a position from the plurality of positions using the modified score distribution comprises:sampling a position from the modified score distribution, andassigning the macro node to the sampled position.

37. The method of claim 33, further comprising, for each of the one or more score distributions, when the set does not include a valid suggested region for the macro node, assigning the macro node to be placed to a position from the plurality7of positions using the score distribution without modifying the score distribution.

38. The method of claim 37, wherein assigning the macro node to be placed to a position from the plurality of positions using the score distribution comprises:assigning the macro node to the position having the highest score in the score distribution.

39. The method of claim 37, wherein assigning the macro node to be placed to a position from the plurality7of positions using the score distribution comprises:sampling a position from the score distribution, andassigning the macro node to the sampled position.

40. The method of any one of claims 37-39, further comprising, for each of the one or more score distributions, generating, using the policy model, a plurality of respective initial integrated circuit chip placements based on the netlist data and any remaining macro nodes from the netlist data to be placed.Attorney Docket No.: 45288-0612WO141. The method of claim 40, wherein generating, using the policy model, a plurality of respective initial integrated circuit chip placements based on the netlist data and any remaining macro nodes from the netlist data to be placed comprises:for each of the plurality of respective initial integrated circuit chip placements:for each remaining macro node, at a corresponding remaining time step of one or more remaining time steps:generating a remaining policy model input for the corresponding remaining time step;sampling from the policy model given the remaining policy model input for the corresponding time step to generate a plurality of score distributions over the plurality of positions on the surface of the integrated circuit chip; andfor each of the plurality7of score distributions:assigning the remaining macro node to be placed to a position from the plurality of positions using the score distribution.

42. The method of any preceding claim, wherein the policy model is configured to generate the initial integrated circuit chip placement autoregressively by placing the macro nodes over the plurality7of time steps.

43. The method of claim 42, wherein the policy model is configured to place each macro node at a position at a different corresponding time step according to a macro node order.

44. The method of claim 43, wherein the policy model input for the corresponding time step comprises one or more of: a target value of a reward function that measures a quality of the initial integrated circuit chip placement, data characterizing respective positions on the surface of the chip of any macro nodes that are before a particular macro node to be placed at the corresponding time step in the macro node order, data characterizing the particular macro node to be placed at the corresponding time step, data characterizing the connectivity between two or more of the plurality of nodes specified by the netlist data, data representing the netlist data, or metadata for the netlist data.

45. The method of claim 43, wherein the policy model input for the corresponding time step comprises one or more of: data representing the netlist data, data characterizingAttorney Docket No.: 45288-0612WO1respective positions on the surface of the chip of any macro nodes that are before a particular macro node to be placed at the corresponding time step in the macro node order, data characterizing the particular macro node to be placed at the corresponding time step, data characterizing the initial integrated circuit chip placement at one or more time steps that are before the corresponding time step, a target value of a reward function that measures a quality of the initial integrated circuit chip placement, or respective target values of the reward function for each initial integrated circuit chip placement at one or more time steps that are before the corresponding time step.

46. The method of any preceding claim, wherein generating a respective set of one or more initial integrated circuit chip placements at each of a sequence of iterations comprises:at each of one or more initial iterations that precede a first particular iteration in the sequence:generating, using the policy model, one respective initial integrated circuit chip placement for the initial iteration without using any suggested regions generated by the multimodal neural network.

47. The method of any preceding claim, wherein generating a respective set of one or more initial integrated circuit chip placements at each of a sequence of iterations comprises:at each of one or more intermediate iterations that are after a first particular iteration in the sequence:generating, using the policy model, one or more respective initial integrated circuit chip placements for the intermediate iteration based on the one or more sets of suggested regions generated at a most recent particular iteration.

48. The method of any preceding claim, wherein the policy model has been trained on a plurality7of training examples to generate integrated circuit chip placements that optimize for one or more criteria.

49. The method of claim 48, wherein the one or more criteria measure one or more of: a degree of overlap of the macro nodes in the initial integrated circuit chip placement, wire length of wires between nodes on the surface of the chip in the initial integrated circuit chip placement, routing congestion measuring available wiring resources in a given region on the chip compared to wires that run through the given region in the initial integrated circuit chipAttorney Docket No.: 45288-0612WO1placement, a cost for the initial integrated circuit chip placement, power consumption for the chip, heat generation for the chip, area of the initial integrated circuit chip placement, or timing of the initial integrated circuit chip placement.

50. The method of any preceding claim, further comprising, for one or more initial integrated circuit chip placements of the respective sets of initial integrated circuit chip placements, generating a respective training example comprising the initial integrated circuit chip placement.

51. The method of claim 50, further comprising training the policy model using the respective training examples.

52. The method of any preceding claim, wherein the multimodal neural network comprises a visual language model neural network.

53. The method of any preceding claim, further comprising training the multimodal neural network on training examples that each comprise a training initial integrated circuit chip placement.

54. The method of any preceding claim, further comprising training the multimodal neural network on training examples that each comprise data representing the netlist data.

55. The method of any preceding claim, wherein the respective positions comprise grid squares from a grid overlaid over the surface of the integrated circuit chip.

56. The method of any preceding claim, further comprising selecting an initial integrated circuit chip placement from the respective sets of initial integrated circuit chip placements.

57. The method of any preceding claim, further comprising, for one or more of the respective initial integrated circuit chip placements:outputting data specifying the initial integrated circuit chip placement.Attorney Docket No.: 45288-0612WO158. The method of claim 57, wherein outputting data specifying the initial integrated circuit chip placement comprises outputting the initial integrated circuit chip placement to a component of an electronic design automation (EDA) software tool.

59. The method of claim 57 or claim 58, wherein outputting data specifying the initial integrated circuit chip placement comprises providing the data specifying the initial integrated circuit chip placement for presentation to a user on a user device.

60. The method of any one of claims 57-59, wherein outputting data specifying the initial integrated circuit chip placement comprises providing the initial integrated circuit chip placement for fabrication of an integrated circuit chip according to the initial integrated circuit chip placement.

61. The method of any preceding claim, further comprising, for one or more of the set of initial integrated circuit chip placements:fabricating an integrated circuit chip according to the initial integrated circuit chip placement.

62. The method of any preceding claim, wherein each integrated circuit chip placement places only each macro node in the netlist data at a respective position on the surface of the integrated circuit chip.

63. The method of any preceding claim, wherein the plurality of the nodes comprise a plurality of standard cell nodes that each represent a respective standard cell component of the integrated circuit chip.

64. The method of claim 63, wherein each integrated circuit chip placement places each macro node in the netlist data at a respective position on the surface of the integrated circuit chip and each standard cell node at a respective position on the surface of the integrated circuit chip.

65. The method of claim 63, wherein generating a set of one or more integrated circuit chip placements for the integrated circuit chip further comprises, for one or more of the respective initial integrated circuit chip placements, placing each of the standard cell nodes atAttorney Docket No.: 45288-0612WO1a respective position on the surface of the integrated circuit chip that includes the macro components represented by the macro nodes placed according to the respective initial integrated circuit chip placement.

66. The method of any one of claims 64-65, further comprising, for one or more of the set of integrated circuit chip placements:outputting data specifying the integrated circuit chip placement.

67. The method of claim 66, wherein outputting data specifying the integrated circuit chip placement comprises outputting the integrated circuit chip placement to a component of an electronic design automation (EDA) software tool.

68. The method of claim 66 or claim 67, wherein outputting data specifying the integrated circuit chip placement comprises providing the data specifying the integrated circuit chip placement for presentation to a user on a user device.

69. The method of any one of claims 66-68, wherein outputting data specifying the integrated circuit chip placement comprises providing the integrated circuit chip placement for fabrication of an integrated circuit chip according to the integrated circuit chip placement.

70. The method of any preceding claim, further comprising, for one or more of the set of integrated circuit chip placements:fabricating an integrated circuit chip according to the integrated circuit chip placement.

71. An integrated circuit chip having an integrated circuit chip placement generated by performing the operations of the respective method of any preceding claim.

72. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform the operations of the respective method of any one of claims 1-70.

73. One or more computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform the operations of theAttorney Docket No.: 45288-0612WO1respective method of any one of claims 1-70.

74. A computer program product comprising instructions that when executed by one or more processors cause the one or more processors to perform the operations of the respective method of any one of claims 1-70.