AICircuit: A Multi-Level Dataset and Benchmark for AI-Driven Analog Integrated Circuit Design

TL;DR

AICircuit tackles the challenge of AI-assisted analog and RF circuit design by providing a multi-level benchmark that includes seven homogeneous circuit blocks and two 28 GHz heterogeneous transceiver systems, all generated via Cadence-based parameter sweeps. The authors explore an end-to-end pipeline where a design-spec-to-parameter mapping $m{y} \rightarrow \bm{x}$ is learned using models such as MLPs and Transformers, with performance evaluated through a Cadence simulator to yield relative errors $err_i$ on each metric. Key findings show that while DC power and some simple metrics are predicted with low error, more complex and nonlinear metrics (e.g., gain, phase noise, S-parameters) and heterogeneous-system interactions pose substantial challenges, indicating nonlinearity and load effects that simple block-wise models struggle to capture. The results also demonstrate that transferring learning from homogeneous blocks to heterogeneous systems is nontrivial, motivating system-aware modeling and richer datasets to advance AI-driven analog design and benchmarking. Overall, AICircuit provides a standardized resource to benchmark, compare, and improve ML methods for real-world analog/RF design tasks with practical simulator-backed evaluation.

Abstract

Analog and radio-frequency circuit design requires extensive exploration of both circuit topology and parameters to meet specific design criteria like power consumption and bandwidth. Designers must review state-of-the-art topology configurations in the literature and sweep various circuit parameters within each configuration. This design process is highly specialized and time-intensive, particularly as the number of circuit parameters increases and the circuit becomes more complex. Prior research has explored the potential of machine learning to enhance circuit design procedures. However, these studies primarily focus on simple circuits, overlooking the more practical and complex analog and radio-frequency systems. A major obstacle for bearing the power of machine learning in circuit design is the availability of a generic and diverse dataset, along with robust metrics, which are essential for thoroughly evaluating and improving machine learning algorithms in the analog and radio-frequency circuit domain. We present AICircuit, a comprehensive multi-level dataset and benchmark for developing and evaluating ML algorithms in analog and radio-frequency circuit design. AICircuit comprises seven commonly used basic circuits and two complex wireless transceiver systems composed of multiple circuit blocks, encompassing a wide array of design scenarios encountered in real-world applications. We extensively evaluate various ML algorithms on the dataset, revealing the potential of ML algorithms in learning the mapping from the design specifications to the desired circuit parameters.

Figures (25)

• Figure 1: The advances in analog and mm-wave circuit design. (a) The limits of transistor scaling predicted by Moore’s law; (b) Comparison between homogeneous and heterogeneous circuits.
• Figure 2: Conventional procedure of analog circuit design, which involves tremendous efforts to sweep in the parameter space to find the optimal design given design specifications. The design space contains possible parameters such as transistors, resistors, etc. Design specifications contain power consumption, bandwidth, etc.
• Figure 3: Procedure for creating datasets for commonly used analog circuits, including homogeneous and heterogeneous circuits.
• Figure 4: 28 GHz wireless transceiver circuits. (a) Transmitter architecture involving VCO and PA. Buffer used here to sustain system stability; (b) Receiver architecture comprising LNA, Mixer, and CVA. Low-Pass Filter deployed here to filter out the undesired high frequency components.
• Figure 5: An end-to-end model training and evaluation pipeline.