AI-Driven, Timing-Aware Power Optimization

The gist The proposed framework makes post-route power optimization AI-driven in a specific sense: a learned, uncertainty-aware policy decides which power-recovery edits to try, which repairs to apply, in what order and when to stop, while deterministic EDA analysis remains the authority for accepting a new implementation.

Optimization Contract

The framework utilizes a baseline-relative optimization contract that couples recovery and repair through per-view damage sets, baseline targets, a budget tied to the measured saving, and complete transaction rollback (Page 1). This contract defines states as x0 the accepted implementation checkpoint, x1 the speculative recovery-only state and x2 the repaired candidate offered for commitment. The system tracks margins mj(x) for each constrained item j, where tj = min[0, mj(x0)] and dj(x) = max[0, tj −mj(x)−τj] (Page 3). The aggregate search objective is defined as D(x) = the set of items where dj(x) > 0, and Φ(x) = the sum of wj dj(x), with weights wj > 0 (Page 3). Baseline-relative feasibility requires mj(x2) ≥ tj − τj for every required check (Page 3).

Learned Policy and Ranking

The learned model drives both phases of the optimization, ranking recovery edits to try and repair actions that follow (Page 1). The architecture employs an ensemble of gradient-boosted models to predict power change, normalized damage reduction, and constraint-regression risk from action and timing-context features available before each trial (Page 1). The policy utilizes two acquisition scores: Arec(a) = −p+ −λ max(0,−g−)−η v+, and Arep(a) = max(0,g−)ε +max(0, p+)−η v+ (Page 4). Recovery favors the largest conservative power reduction with limited predicted damage, while repair favors deficit reduction per unit of predicted give-back (Page 4).

Transactional Flow and Budgeting

The transactional flow is exposed as a tool with a declared contract, which includes inputs, limits, a budget share and a verification report (Page 5). A provisional recovery saving S = P0 − P1 is used to define the budget B = βS and retention ratio R = S/C (Page 3). Each repair trial is charged by the measured total-power change from the current checkpoint, and cumulative spend is re-evaluated against x1 rather than reconstructed from historical per-cell savings (Page 3). The system stops recovery when the conservative expected saving of the best remaining candidate falls below a declared fraction of the analysis cost per trial, or at the edit or call limit (Page 4).

Evidence and Evaluation Protocol

The contract is supported by evidence from a companion study on public designs, which found that the retention of a conventional whole-design repair fell steeply as the ratio of pre-existing violators to recovery damage rose (Page 1). The paper specifies a preregistered public evaluation protocol with hypotheses and arms designed to separate the value of learning from the value of the contract (Page 1). Hypotheses include H2 (safety), stating that uncertainty-adjusted ranking lowers the protected-constraint regression rate and rollback fraction by at least one third relative to unadjusted ranking, at equal or better final power (Page 6). The evaluation compares existing flow, heuristic ranking, learned sizer/RL ECO policy, and learned ranking without uncertainty adjustment (Page 7).

Limitations and Context

The framework proves neither convergence nor optimality nor the existence of a repair within a chosen budget (Page 8). The learned surrogate can misrank actions when timing paths change, libraries differ or the activity profile shifts, and ensemble disagreement need not identify every error (Page 8). The framework proves neither convergence nor optimality nor the existence of a repair within a chosen budget (Page 8). This manuscript supplies the method, the companion study’s public evidence for the contract and the protocol; it contains no executed study of the proposed model, no trained weights and no measured comparison against the listed baselines (Page 7). The contribution is a specified methodology and an evaluation plan (Page 9).

The paper outlines an AI-driven framework for timing-aware power optimization that separates the decision of which edits to attempt from the deterministic EDA analysis that governs acceptance, using a contract based on baseline preservation and a measured saving budget. This approach leverages learned ranking policies to guide recovery and repair actions, while maintaining strict deterministic checks for every proposed change.

How it works

The framework operates through a sequential decision process where the learned model controls candidate priority, and EDA tools supply the measurements that control state transitions (Page 5). The process involves several key steps:

1. Candidate generation is performed by a deterministic generator that enumerates legal actions before learning is applied (Page 4). Recovery actions include lower-leakage substitutions and drive reductions, while repair actions involve restoration of a recorded recovery edit or localized buffer insertion (Page 4).

2. The learned ranking policy uses features such as electrical context (input slew, output load) and timing context (setup and hold margins by view) to rank these legal actions using acquisition scores like Arec(a) or Arep(a) (Page 4).

3. Trial labels are generated based on the trial outcome x' relative to the current state x, specifically yP = P(x′)−P(x), yG = Φ(x)−Φ(x′), and yV = 1 if a protected check regresses in x' (Page 4).

4. The system uses uncertainty-aware acquisition scores to rank actions, where p+ is an upper estimate of normalized power change, g− is a lower estimate of damage reduction, and v+ is a conservative estimate of regression risk (Page 4).

5. The transaction flow involves trial candidates one at a time in a saved branch; only physically verified results that preserve every baseline target within the repair budget become a new accepted checkpoint (Page 5).

Key Components

The framework is characterized by several specific components:

((

The contract defines states, margins, and baseline targets, along with normalized damage dj(x) and the aggregate search objective D(x) (Page 3).

((

Power accounting defines the provisional recovery saving S = P0 − P1 and sets the budget B = βS, C = P2 − P1, and R = S/C (Page 3).

((

The learned policy is designed as a ranking with uncertainty rather than end-to-end reinforcement learning, favoring sample-efficient supervised surrogates (Page 4).

((

A deterministic fallback mechanism is in place, triggered by high ensemble disagreement or failed calibration diagnostics, which ranks legal restoration actions by measured deficit coverage per unit of estimated give-back (Page 4).

The framework establishes a methodology and an evaluation plan rather than certifying the results of the learned policy (Page 9).

1. Bold ranking policy for recovery edits: The learned model drives both phases by predicting how much power an edit will save, how much damage it will cause or remove to generate acquisition scores like Arec(a) = −p+ −λ max(0,−g−)−η v+ and Arep(a) = max(0,g−)ε +max(0, p+)−η v+. This allows the system to prioritize actions based on a learned trade-off between power reduction and damage risk.

2. Robust uncertainty handling: The framework employs calibrated, conservative estimates and includes a deterministic fallback when the model is uncertain, ensuring that recovery and repair remain speculative until physical implementation and fresh analysis establish baseline preservation and positive net saving. This prevents speculative actions from being accepted without verification.

3. Baseline-relative constraint management: The system enforces a contract where the objective is not just to achieve timing cleanliness, but to maintain constraints relative to the original state, as described by Baseline-relative feasibility requires mj(x2) ≥ tj − τ j for every required check. This prevents accumulated constraint drift and ensures that baseline targets are enforced at every commit, as stated in Section III.

4. Damage budgeting and cost control: The system utilizes a budget tied to the measured saving, defining B = βS where S = P0 −P1, ensuring that repair spending is capped at a fraction of the measured saving, thereby preventing runaway costs by charging each trial by the measured total-power change from the current checkpoint.

5. Transactional flow orchestration: The system exposes Section V as a tool with a declared contract, allowing an orchestrating agent, including one based on a large language model, to schedule transactions and summarize damage by apparent cause for human review. This enables high-level management of the sequential decision-making process.

6. Uncertainty-aware ranking for repair: The system utilizes different acquisition scores for recovery (Arec(a)) versus repair (Arep(a) = max(0,g−)ε +max(0, p+)−η v+), allowing the model to prioritize actions based on whether the goal is power reduction or deficit reduction per unit of predicted give-back.