TACS-RF Puts an AI Agent Behind the EDA Controls for mmWave Design

Introduction

The use of large language models in hardware design is moving beyond script generation. A paper on arXiv presents TACS-RF, or Tool-driven Agentic Circuit Synthesis of RF and mmWave Modules. The framework places a coding agent, identified as Claude Opus 4.8, in direct control of industrial electronic-design-automation tools, including Cadence Virtuoso/Spectre, Calibre, Cadence EMX, and Ansys HFSS.

The goal is not simply to ask a model for a circuit topology. Instead, the agent participates in a tool-based design loop that spans circuit implementation, layout-related processing, parasitic extraction, and electromagnetic verification. That makes the proposal relevant to analog and RF engineering, where performance depends on interactions between schematic choices and physical implementation.

Key points

• A tool-connected agent: TACS-RF gives the model access to mainstream EDA environments rather than limiting it to an abstract circuit-generation task.
• Broad sub-block coverage: In the reported demonstration, the agent semi-autonomously designed a 60 GHz push-push Colpitts VCO with second-harmonic extraction, multistage differential cascode power drivers, a push-push frequency doubler, a differential bowtie antenna, and the interstage matching networks.
• A demanding validation target: The resulting transmitter targets 240–251 GHz and uses a 90 nm SiGe BiCMOS process. The reported chain includes parasitic-extracted devices and EM-verified passive structures.
• Measured design targets in simulation and verification: The transmitter delivers +5.1 to +6.3 dBm across an 11 GHz varactor-only tuning range. Its phase noise is reported as -80 to -88 dBc/Hz at a 1 MHz offset, with 361 mW of DC power consumption.

Why it matters

RF and mmWave design is difficult to automate because the relevant constraints cross several domains. Device models, matching networks, layout parasitics, electromagnetic coupling, and simulator behavior all affect the final result. Many existing automation approaches focus on one optimization stage. TACS-RF instead explores whether an agent can coordinate several tools and sub-blocks as part of a more complete engineering workflow.

The result should not be read as evidence that a single prompt can produce a production-ready chip. The paper describes the process as semi-autonomous and discusses a division of labor between humans and the agent. Engineers still need to define specifications, constrain the search, check tool outputs, handle convergence issues, and judge whether a result is physically credible. These responsibilities are especially important at sub-THz frequencies, where small implementation details can materially change performance.

The broader contribution is a concrete step toward a “text-to-ASIC” workflow for difficult analog hardware. The open questions are equally important: how much manual intervention is required, whether the approach generalizes across processes and architectures, how reliably it handles failures, and whether post-silicon measurements confirm the design predictions. For now, TACS-RF is best viewed as an engineering demonstration of agent-assisted RF synthesis, not a replacement for RF designers.

Source: arXiv