Introduction
Analog circuit sizing—the process of determining optimal dimensions and parameters for transistors and components—remains a significant bottleneck in semiconductor design. Unlike digital design, which benefits from high levels of automation, analog design is largely a manual, iterative process. Engineers typically rely on SPICE simulators to verify designs, but the high computational cost of these simulations makes exhaustive search impractical. Existing automated tools often require tens of thousands of simulations or fail to scale to industrial-scale circuits containing hundreds of thousands of components. This paper presents DNN-Opt, a sample-efficient framework that leverages Reinforcement Learning (RL) architectures and data augmentation to automate and accelerate this process.
Methodology
DNN-Opt adopts a two-stage Deep Neural Network (DNN) architecture inspired by the Actor-Critic framework used in Reinforcement Learning. While circuit sizing is not a traditional Markov Decision Process (MDP), the authors demonstrate that this structure is highly effective for navigating complex design spaces.
The Actor-Critic Architecture
- The Critic Network: This network serves as a surrogate for the computationally expensive SPICE simulator. It learns to predict circuit performance based on input design parameters. To maximize sample efficiency, the framework employs pseudo-sample generation. By training the network on the differences between pairs of designs rather than just individual points, the system effectively squares the amount of training data available from a fixed number of simulations.
- The Actor Network: The Actor proposes design improvements ($\Delta x$) to minimize the objective function (such as power consumption) while ensuring all performance constraints (like gain and noise) are met.
Scaling to Industrial Circuits
To manage circuits with massive component counts, DNN-Opt utilizes Sensitivity Analysis. By perturbing variables to identify those with the most significant impact on performance, the search space can be pruned. This allows the optimizer to handle industrial designs, such as a Continuous-Time Linear Equalizer (CTLE) with 173,000 transistors, by focusing only on the critical devices.
Experimental Results
The framework was evaluated against traditional optimization methods, including Differential Evolution (DE) and Bayesian Optimization (BO), across various circuit topologies.
Efficiency and Performance
In tests on a Folded Cascode Operational Transconductance Amplifier (OTA), DNN-Opt demonstrated remarkable efficiency. It reached a feasible solution in only 132 simulations, representing a 24x speedup over Differential Evolution, which required 3,200 simulations. Furthermore, the final design produced by DNN-Opt consumed 43% less power than those generated by competing model-based methods.
Industrial Validation
DNN-Opt's scalability was proven on industrial-scale circuits. For a CTLE design with over 170k transistors, the framework met all design constraints using 4x fewer simulations than a commercial Simulated Annealing tool, while also achieving an 8% reduction in power consumption. In reliability tests for a Strong-Arm Latch Comparator, DNN-Opt was the only method to achieve a 100% success rate in finding feasible solutions across multiple trials.
Conclusion
DNN-Opt represents a significant advancement in automated analog circuit sizing. By combining the strengths of RL-inspired architectures with clever data augmentation through pseudo-sampling, it achieves 5–30x better sample efficiency than traditional methods. Its ability to scale to industrial-sized circuits with hundreds of thousands of transistors addresses a critical need in modern chip design. Future work may explore transfer learning to allow models trained on one topology to accelerate the optimization of new, different circuit designs.
Summary updated