Volltext-Downloads (blau) und Frontdoor-Views (grau)

Differentiable Neural Network Surrogate Models for gm/ID-based Analog IC Sizing Optimization

  • Analog integrated circuit sizing still relies heavily on human expert knowledge as previous automation approaches have not found wide-spread acceptance in industry. One strand, the optimization-based automation, is often discarded due to inflated constraining setups, infeasible results or excessive run times. To address these deficits, this work proposes a alternative optimization flow featuring a designer’s intuition for feasible design spaces through integration of expert knowledge based on the gm/ID-method. Moreover, the extensive run times of simulation-based optimization flows are overcome by incorporating computationally efficient machine learning methods. Neural network surrogate models predicting eleven performance parameters increase the evaluation speed by 3 400× on average compared to a simulator. Additionally, they enable the use of optimization algorithms dependent on automatic differentiation, that would otherwise be unavailable in this field. First, an up to 4× more efficient way for sampling training data based on the aforementioned space is detailed. After presenting the architecture and training effort regarding the surrogate models, they are employed as part of the objective function for sizing three operational amplifiers with three different optimization algorithms. Additionally, the benefits of using the gm/ID-method become evident when considering technology migration, as previously found solutions may be reused for other technologies.

Download full text files

  • 4684.pdf
    eng

Export metadata

Additional Services

Search Google Scholar

Statistics

frontdoor_oas
Metadaten
Author of HS ReutlingenUhlmann, Yannick; Moldenhauer, Till; Scheible, Jürgen
DOI:https://doi.org/10.1109/MLCAD58807.2023.10299834
Erschienen in:2023 ACM/IEEE Workshop on Machine Learning for CAD Organization
Publisher:IEEE
Place of publication:Piscataway, NJ
Document Type:Conference proceeding
Language:English
Publication year:2023
Page Number:6
DDC classes:620 Ingenieurwissenschaften und Maschinenbau
Open access?:Nein
Licence (German):License Logo  In Copyright - Urheberrechtlich geschützt