AgenticECO is an automated framework designed to perform post-route engineering change orders (ECO) on 3D integrated circuits. It addresses the challenge of fixing bond-level physical defects in 3D-IC designs without triggering the excessive router churn associated with traditional, manual, or full-reroute methods. By using an evidence-gated, tool-using agent, the system generates deterministic repair programs that allow for precise, attributable fixes that pass rigorous signoff gates.
The Problem: 3D-IC Repair
As chip design moves into three dimensions, hybrid bonding creates vertical connections that can suffer from spacing violations. These defects often occur after the design has already been routed, meaning the physical layout is essentially frozen. Traditional repair methods—such as full-reroute or manual editing—often entangle the repair with unrelated changes to the design, making it impossible to determine if a specific edit actually resolved the issue. Furthermore, because these defects are cross-tier, they have no direct equivalent in 2D design, making standard 2D ECO tools insufficient.
How AgenticECO Works
The framework operates through a closed-loop, four-stage process: querying evidence, generating a repair program, executing a registered edit, and verifying the results.
Evidence-Gated Orchestration: Instead of analyzing the entire chip database at once, the system uses "specialist agents" to gather scoped, typed evidence. This evidence is used to synthesize a deterministic repair program.
EcoRoute Executor: To ensure the repair is attributable to the specific edit, the system uses a minimal-disturbance routing layer. It only reroutes nets incident to the modified instances, keeping the rest of the design frozen.
Independent Verification: A separate, frozen verifier evaluates the results against strict criteria, including routing completion, design-rule checks (DRC), timing, and structural equivalence. The model itself does not perform the verification, ensuring that the acceptance gates remain objective.
Adaptive Control: If an attempt fails, the orchestrator uses the failure artifacts to decide whether to try a different candidate, gather more evidence, or revise the entire repair program.
Experimental Results
In tests across nine natural post-route defect cases, AgenticECO cleared seven cases, compared to two for both full-reroute and stock repair methods. The successful repairs achieved a mean disturbance of 0.66% to the design, with zero clock nets touched. When subjected to a cross-backbone rerun, the system cleared all nine cases. The researchers note that the system successfully overturned incorrect infeasibility verdicts produced by a single-agent baseline and ensured that every accepted result passed all structural and timing gates.
Key Considerations
The framework is designed to operate under a fixed contract that defines the rules, edit vocabulary, and acceptance gates. While the system automates the repair process, it relies on the ability to isolate edits to prevent unrelated routing churn. The authors emphasize that the system is intended for post-route 3D-IC ECO, where the primary goal is to resolve localized violations while preserving the integrity of the existing, valid implementation. The code, environment, and audit artifacts are provided to support the reproducibility of these results.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!