AI-Driven, Timing-Aware Power Optimization

AI-Driven, Timing-Aware Power Optimization. By Amit Sharma. Published on the eighth of October. No code repository is linked to it.

After a chip design has been routed, there is still power to recover by changing which cells are used. Those changes are made under constraints that are tightly coupled: timing, electrical and physical all at once. Swap a cell to save power and something else may break. The real value of any one change therefore depends on what it will cost to repair the side effects it causes, and that cost is not visible from the change alone.

The proposal is a framework in which a learned model drives both phases of the optimization. It ranks which power recovery edits to try, and it ranks the repair actions that follow them. What the model never does is decide whether a change is kept. Deterministic electronic design automation analyses govern every acceptance. The learning chooses where to look, and the established analyses remain the judge.

The framework works as a careful ledger. It records the baseline constraint margins first. For each recovery transaction it identifies the damage that transaction introduces in each analysis view. Violations that already existed are held at their baseline, rather than being worked on inside the power step. And repair spending is capped at a fraction of the saving that was measured, so fixing a change can never cost more than the change was worth.

The learned part is an ensemble of gradient-boosted models. Using features of the action and its timing context that are available before each trial, it predicts three things: the power change, the normalized reduction in damage, and the risk of a constraint regression. Calibrated, conservative estimates feed separate acquisition scores for recovery and for repair, and when the model is uncertain there is a deterministic fallback. Nothing is final while it is being tried. Recovery and repair stay speculative until physical implementation and fresh analysis establish two facts: the baseline is preserved, and the net saving is positive.

The paper is explicit about what it does and does not show. It specifies the contract, the learned policy, the transactional flow, and a pre-registered public evaluation protocol on open designs. Evidence from a companion study on public designs supports the contract. The learned policy itself, though, is proposed rather than tested. No trained-model results are reported, and no measurements on proprietary designs.

This reads as a specification laid down before the experiment. The division of labour is the core of it: a model that suggests and estimates, analyses that accept or reject, and a transaction that does not commit until saving and baseline are both confirmed. Whether the learned ranking pays off is left to the protocol the paper registers in advance.