Why does a clean graph-based timing report still need a path-based rerun before signoff?

Short Answer

Graph-based analysis (GBA) times every stage of every path using the single worst-case delay for that stage, so it never actually walks one real path start to finish. That worst case might come from a completely different path than the one being reported, which stacks extra pessimism onto the number PrimeTime prints. Path-based analysis (PBA) re-times the specific path in question using its own real delays, so signoff only trusts a report once the paths that look like they fail under GBA have been re-checked with PBA.

Technical Explanation

• GBA builds one timing graph for the whole design and, at each pin, keeps only the worst arrival and required time seen across every path that touches that pin.
• Because GBA mixes worst-case numbers from different paths at every stage, the "path" it reports is a worst-case composite that no signal actually travels.
• This composite is always pessimistic, never optimistic, so a path that fails setup under GBA might have real slack once every stage is timed along its own actual path.
• PBA re-times a single path with report_timing -pba_mode exhaustive (PT) or -pba_mode path (PT), using only the real delays that path's signal passes through.
• PBA typically recovers some slack because the worst-case-per-stage assumption is removed, but it never makes a genuinely failing path look clean, it only tightens the number.
• PBA on every path in a design costs far more compute than a graph-based sweep, so it is usually reserved for the paths GBA reports as violating or marginal.
• Signoff practice runs GBA everywhere first, then reruns PBA on the reported violators before deciding a fix is really needed.

Follow-up Question & Model Response

If PBA always shows less pessimism, why does signoff not just run PBA everywhere and skip GBA entirely?

Candidate Model Response: PBA is dramatically slower because it re-times each path independently instead of reusing one shared graph, so running it on every path in a multi-million-instance design would make signoff impractical. GBA gives a fast, guaranteed-pessimistic sweep of the whole design, and its output is exactly the list PBA needs to check next. Using PBA only on GBA's reported violators keeps the full-chip pass fast while still confirming the small number of paths that matter with real numbers. This two-pass approach is standard in PrimeTime signoff flows for exactly this reason.