Algorithmic Challenges in 3D Electronic Design Automation

23 concepts21 papers1 proposal

As traditional two-dimensional (2D) integrated circuits hit a "wall" of interconnect delay and power consumption, stacking components vertically into Three-Dimensional Integrated Circuits (3D-ICs) has emerged as the primary path forward for semiconductor scaling. By replacing long planar wires with short vertical interconnects, 3D integration promises massive reductions in latency and footprint.

However, moving from a 2D plane to a 3D volume fundamentally breaks the mathematical and algorithmic foundations of existing Electronic Design Automation (EDA) tools. Traditional physical design workflows—developed over decades for planar silicon substrates—cannot scale or adapt to the vertical dimension without severe re-engineering. The core challenge is not simply adding a $Z$-axis to existing equations, but managing an exponential explosion in the optimization search space alongside tightly coupled physical and thermal constraints.

The Search Space Explosion in 3D Floorplanning

In traditional 2D physical design, floorplanning involves arranging rectangular circuit blocks on a flat plane such that they do not overlap while minimizing total area and wirelength. When transitioning to 3D, this problem transforms into a volumetric packing puzzle.

As Renshen Wang et al. demonstrate in On the complexity of graph cuboidal dual problems for 3-D floorplanning of integrated circuit design, even basic formulations of 3D floorplanning are NP-complete. The complexity arises because 2D topological representations do not readily map to 3D cuboidal structures. Adding the vertical dimension introduces an astronomically large search space. According to Guojie Luo et al. in Advances and Challenges in 3D Physical Design, and further supported by Chengrui Gao et al. in Expertise Can Be Helpful for Reinforcement Learning-based Macro Placement, automated frameworks designed for planar layouts face severe scalability issues because they cannot efficiently navigate the expanded combinatorial possibilities of multi-layer environments.

This explosion is present at every level of the design hierarchy, down to individual transistors. For instance, when synthesizing standard cells using advanced vertical architectures, Daeyeal Lee et al. show in Many-Tier Vertical Gate-All-Around Nanowire FET Standard Cell Synthesis for Advanced Technology Nodes that finding optimal component placement and signal routing within a 3D grid becomes too complex for standard heuristics, requiring heavy logic-based solvers to prevent spatial constraint violations.

Multi-Tier Placement and Heterogeneous Discontinuities

Placement algorithms must determine the exact coordinates of millions of components. In a 3D-IC, placement requires a two-fold decision: partitioning components into discrete vertical tiers and placing them optimally within those tiers.

This introduces two major algorithmic hurdles:

  1. Multi-Scale Optimization: As the authors of Advances and Challenges in 3D Physical Design point out, current tools struggle to simultaneously optimize the placement of millions of tiny standard cells alongside massive functional macro blocks across multiple vertical tiers. The scale mismatch makes uniform mathematical modeling highly inefficient.
  2. Heterogeneous Technology Nodes: Modern 3D-ICs frequently leverage face-to-face (F2F) bonding to stack entirely different technologies (e.g., a memory tier optimized for density on top of a logic tier optimized for speed). In Analytical Die-to-Die 3-D Placement With Bistratal Wirelength Model and GPU Acceleration, Peiyu Liao et al. explain that traditional analytical placement methods rely on smooth, differentiable wirelength approximations. When tiers have different physical grids and design rules, these mathematical models break down, failing to accurately capture wirelength and partitioning constraints simultaneously.

This multi-layer placement challenge is not confined to conventional silicon. In emerging architectures like superconducting quantum computers, Junyao Zhang et al. note in qGDP: Quantum Legalization and Detailed Placement for Superconducting Quantum Computers that planar placement frameworks fail when qubits and resonators must be distributed across distinct physical layers to mitigate crosstalk.

The Routing and Interconnect Bottleneck

Routing algorithms find the physical paths for wires connecting the placed components. In 2D, routing models assume signals travel along a few planar metal layers. In 3D, routing must handle vertical interconnects—such as Through-Silicon Vias (TSVs) or F2F micro-bumps—which occupy significant physical space and have distinct electrical properties.

Traditional 2D Routing:          3D Multi-Tier Routing:
   [Cell A] ------> [Cell B]        Tier 2: [Cell A] --------+
                                                             | (Vertical Interconnect / TSV)
                                                             v
                                    Tier 1:                  +------> [Cell B]

Standard routing formulations, like the Rectilinear Steiner Minimum Tree (RSMT) problem, are fundamentally planar. As Andrew B. Kahng et al. observe in NN-Steiner: A Mixed Neural-Algorithmic Approach for the Rectilinear Steiner Minimum Tree Problem, extending these frameworks to 3D space is mathematically necessary but computationally daunting because the router must balance lateral wire lengths with the discrete locations of vertical vias.

Manufacturing constraints compound this difficulty. To make advanced nodes printable, designers use unidirectional routing, where wires on a given layer only run in one direction (e.g., horizontal on Layer 1, vertical on Layer 2). The author, Xiaoqing Xu, argues in Toward Unidirectional Routing Closure in Advanced Technology Nodes that while these constraints work in 2D, managing concurrent multi-layer routing in 3D creates severe bottlenecks where one-way constraints on different levels intersect and compete for vertical connectivity. Consequently, generative layout tools must be completely re-architected; Yunheng Shen et al. emphasize in DiffPattern: Layout Pattern Generation via Discrete Diffusion that single-layer pattern generation cannot model the strict vertical alignment and interdependencies required by 3D architectures.

The Physics-Design Loop: Thermal and Material Constraints

What prevents 3D EDA from being treated as a pure geometric optimization problem is the underlying physics. Stacking active silicon layers traps heat. In a 2D chip, heat escapes easily through the substrate to a heatsink. In a 3D chip, inner layers are insulated by surrounding logic and dielectric materials, leading to severe thermal hotspots that degrade performance and reliability.

This creates a tight coupling between architectural decisions, physical layout, and thermal management:

Because standard commercial EDA tools lack native support for 3D physical synthesis, vertical floorplanning, and automated tier partitioning, engineers are often forced into highly manual, iterative, and sub-optimal design workflows (Nassim Abderrahmane et al. in NimbleAI: Towards Neuromorphic Sensing-Processing 3D-integrated Chips; Giulia Beanato et al. in Configurable Low-Latency Interconnect for Multi-core Clusters).

To unlock the full potential of 3D integration, the research direction points toward a unified paradigm shift. As argued by Ozgur Sinanoglu et al. in Large-Scale 3D Chips: Challenges and Solutions for Design Automation, Testing, and Trustworthy Integration, the industry requires standardized, holistic design workflows that natively integrate floorplanning, routing, and thermal management into a single, concurrent 3D optimization engine.

Key Foundational Work

On the complexity of graph cuboidal dual problems for 3-D floorplanning of integrated circuit design (2009) — contributed 2 concept elements to this direction.
Fine grain 3D integration for microarchitecture design through cube packing exploration (2007) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_heterogeneous_3d_eda_challenges_4b428438,
  title = {heterogeneous 3D EDA challenges},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/heterogeneous-3d-eda-challenges/rdr_4b4284387bda4b28ab9fd4baa324e624},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_heterogeneous_3d_eda_challenges_4b428438_v1,
  title = {heterogeneous 3D EDA challenges},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/heterogeneous-3d-eda-challenges/rdr_4b4284387bda4b28ab9fd4baa324e624/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake