Skip to main content
Back to timeline
arXivSource publication:

Open-source agent platform OpenLayout takes over VLSI physical design, lifting atomic-task Pass@1 well above a standalone LLM

Synopsis

OpenLayout turns VLSI physical design from a scripted flow that senior engineers tune command by command into an automated loop run by specialized agents on commercial EDA tools. Its companion OpenLayout-Bench evaluates agent capability at TSMC 28nm: standalone Gemini 3.1-pro reaches only 18.9% Pass@1 on Tool A atomic tasks, while OpenLayout reaches 80.0% on the same metric, and on Tool B the standalone model reaches 19.3% versus 84.1% with OpenLayout. On the 10-design holistic flow suite, small and medium designs close timing, while NVDLA-large and Open910 retain residual DRC violations after Routeopt.

Source-provided article image: OpenLayout: An Open and Agentic-Driven Platform for Holistic VLSI Physical Design and Benchmarking
arXiv

Interpretation

Chip physical design no longer depends solely on senior engineers writing Tcl scripts, reading logs and retuning parameters by hand. OpenLayout hands the whole flow, from floorplanning and power planning through placement, clock tree synthesis and routing, to a team of specialized agents that read tool feedback, diagnose errors and iterate on parameters. Earlier automation either relied on predefined parameter spaces and hand-crafted objectives, or optimized a single stage; LLM-based approaches mainly used the model's own knowledge or retrieval-augmented generation to interpret tool feedback and struggled on complex designs and advanced technology nodes. On the OpenLayout-Bench atomic task suite, standalone Gemini 3.1-pro reaches 18.9% Pass@1 on Tool A, versus 80.0% with OpenLayout; on Tool B the standalone model reaches 19.3%, versus 84.1% with OpenLayout. The retrieval-augmented RAG-Gemini reaches 60.0% Pass@1 on Tool A, still below OpenLayout's 80.0%.

The platform organizes five agents through hub-and-spoke message passing: a Planning Agent decomposes user intent into a directed acyclic graph of subtasks and maintains global state, a Worker Agent sends Tcl commands to tools such as Cadence Innovus through a persistent interactive shell, an Analyzer Agent extracts timing, utilization, DRC and power metrics from verbose logs, a Debugger Agent performs root-cause analysis, and an Optimizer Agent tunes parameters along metric trajectories. Domain knowledge is written as auto-discovered skill packs named by technology node and tool, containing reference Tcl scripts, tuning strategies, debug decision matrices and cross-stage impact analyses, so switching process or toolchain does not require changing agent code. The holistic flow suite contains 10 open-source RTL designs, from a 176-cell elevator FSM to the 1,478,865-cell NVDLA-large; with Gemini 3.1-pro-preview as the backbone, OpenLayout completes the full flow through an average of 53 interactions with the EDA tools.

Full-flow results split by design scale: small and medium designs reach timing closure with zero or minimal DRC violations at the final Routeopt stage, while NVDLA-large and Open910 retain residual DRC violations after Routeopt and do not fully close timing in all cases. The retrieval-augmented baseline fails to complete most designs beyond the first two stages, with pin placement failures at floorplanning and missing power rings or unconnected stripes at power planning. On the 10-design holistic flow suite, OpenLayout runs with Gemini 3.1-pro-preview as the backbone; the retrieval-augmented method stalls at the first two stages on most designs.

Perspective

The evaluation covers the physical design flow from initialization to Routeopt, using Cadence Innovus 22, Synopsys ICC2 2022, PrimeTime 2022 and StarRC 2022 at the TSMC 28nm node; the holistic flow suite contains 10 open-source RTL designs. The platform does not yet support LVS verification, and DRC fixing remains incomplete for violations that require GUI-based interaction.

Full-flow results are reported only with Gemini 3.1-pro-preview as the backbone; NVDLA-large and Open910 retain residual DRC violations after Routeopt and do not fully close timing, which the authors partly attribute to the framework's current lack of more sophisticated ECO and DRC fixing capabilities. Evaluation is limited to TSMC 28nm and does not yet reflect performance at other process nodes. The source prose states 84.4% and 86.4% as pass rates, while in the table those two figures correspond to Pass@5, not the Pass@1 of the 18.9% standalone baseline; the source does not resolve this inconsistency.