TL;DR
This paper introduces a novel multi-objective optimization framework for common-centroid (CC) placement of analog transistors. By simultaneously considering crucial layout constraints like diffusion break, routing complexity, and layout-dependent effects, the proposed method explores a wider range of CC topologies and achieves superior post-layout simulation results compared to existing techniques.
What Problem Does This Paper Address?
In analog circuit design, process variations can significantly degrade circuit performance, particularly affecting the matching of critical components. Common-centroid (CC) layouts are a widely adopted technique to mitigate these variations by ensuring that matched devices share similar environmental conditions. However, designing optimal CC layouts is a complex task. Existing methods often struggle to collectively handle multiple, competing layout constraints such as maximizing device dispersion, minimizing routing complexity, and reducing layout-dependent effects (like Length of Diffusion and Well Proximity Effects). This often leads to suboptimal solutions or requires extensive manual tuning. The paper aims to bridge this gap by providing a unified optimization framework that systematically addresses these challenges.
Methods
The core of the proposed solution is a multi-objective optimization framework built upon a modified simulated annealing algorithm called AMOSA (Adaptive Multi-Objective Simulated Annealing). Unlike single-objective optimization, which requires manual weighting of objectives, AMOSA inherently handles multiple objectives by focusing on the concept of "domination" among solutions.
Key methodological contributions include:
- Unified Multi-Objective Framework: Integrating various layout constraints (degree of dispersion, routing cost, diffusion break, layout-dependent effects) into a single optimization problem.
- Enhanced AMOSA: Adapting AMOSA to handle multiple new solutions generated per iteration (especially from XY/180° transformations) and refining its probability calculation for accepting new solutions.
- Powerful Transformations for Topology Exploration: Employing XX/180° and XY/180° transformations to generate a significantly larger and more diverse set of CC topologies than previously possible. This allows for a more thorough exploration of the design space.
- Constraint Handling: Specifically addressing diffusion breaks and dummy transistor counts by setting upper bounds, ensuring practical and manufacturable layouts.
Datasets
Not discussed in the paper. The paper uses existing circuit configurations (current mirror, Cascode Differential Input Pair, Cascode Differential Load Pair) and relies on post-layout simulations using industry-standard tools (Cadence Virtuoso) for evaluation.
Key Insights and Ideas
- The Power of Multi-Objective Optimization: Directly tackling multiple competing constraints simultaneously, rather than sequentially or with manual weighting, leads to more balanced and ultimately better overall solutions.
- Topology Exploration is Crucial: The choice of CC topology has a significant impact on various performance metrics. The paper demonstrates that exploring a wider range of topologies through powerful transformations (like XY/180°) is key to finding superior placements.
- Interplay of Constraints: There are inherent trade-offs between different layout objectives. For instance, maximizing dispersion might increase routing complexity or area. The multi-objective approach effectively navigates these trade-offs.
- Diffusion Break Management: Limiting diffusion breaks and dummy counts is essential for practical layouts, and integrating these as optimization objectives alongside others yields better results than optimizing for dummy count alone.
Key Results
The proposed method demonstrates significant improvements across various metrics when compared to state-of-the-art techniques, particularly in post-layout simulations.
- Improved Mismatch: For current mirror circuits, the proposed method consistently achieves lower current mismatch values compared to a reference method [25]. For example, in CM:1, the mismatch improved from 2.38 to 3.39 (model) and 75 to 77 (router) for the current mirror circuit.
- Reduced Routing Cost: The routing cost, measured by RMST/MRST, is generally lower with the proposed method, indicating more efficient wire routing. For instance, in CDIP:1, the routing cost dropped from 124/186 (model/router) to 122/139.
- Better Layout-Dependent Effects (LDE): The LDE values are often reduced, especially for topologies where it was previously non-zero. For example, in CDIP:1, LDE improved from 0.21 to 0.01.
- Efficient Diffusion Break and Dummy Count: The method effectively generates solutions with zero diffusion breaks and minimal dummy counts, as seen in Tables 1 and 2.
- Wider Topology Exploration: The paper claims to explore more CC topologies than previous methods, leading to better overall performance.
Limitations and Future Work
- Runtime: While the paper mentions a runtime of about 1 minute for the largest test case, the computational cost of multi-objective optimization can still be a factor for very large circuits.
- Post-processing for Solution Selection: While the algorithm generates a set of non-dominated solutions, a post-processing step is used to select the "best" solution based on specific requirements. This selection process could potentially be further refined.
- Circuit-Level Optimization: The current work focuses on device-level placement. The authors acknowledge that for complex circuits, optimizing individual device groups with different topologies and then considering their interconnections (circuit-level optimization) might yield even better results. This is suggested as a direction for future work.
- Exploring More Transformations: While XX/180° and XY/180° are powerful, exploring other symmetry-preserving transformations could potentially uncover even more novel CC topologies.
Summary updated