Level 3 · Signoff Reasoning

Expert MMMC Interview Questions

Reason through ECO side effects, correlation gaps, parasitic coverage, PBA/GBA, SI, and signoff acceptance.

0 of 10 marked complete

What to practise at this level

Treat every result as scoped evidence. Compare environments, validate coverage, and cross-check every physical change across all required views.

  1. 01 A setup ECO improves WNS by 100 ps in one view but creates hold violations in two other views. What actually happened?Expert · Nothing is wrong with the timing engine. The physical change you made — a resize, a buffer insertion, whatever it was — is global. It affected every view that touches that cell or net, not just the setup view you were optimizing.
  2. 02 Why might ICC2 report clean timing while PrimeTime signoff on the same design shows failures?Expert · This is almost never "the signoff tool is being pessimistic." It's much more likely a correlation gap — a mismatch somewhere between what ICC2 analyzed and what PrimeTime analyzed: netlist revision, SDC, libraries, operating condition, parasitics, SI settings, OCV/AOCV/POCV treatment, CPPR, or whether one side used graph-based analysis and the other used path-based.
  3. 03 Timing changed dramatically between two runs, but the netlist is provably identical. What do you compare first?Expert · The run environment, before you ever suspect tool instability. Compare SDC, library versions, parasitic extraction/revision, derating and SI settings, clock definitions, and tool version — in that order — looking for the first thing that actually differs.
  4. 04 SPEF loaded without a fatal error, but one critical net reports almost zero net delay. Does that prove parasitic annotation is correct?Expert · No. A successful file read only proves the parser accepted the file — it says nothing about whether every net was correctly matched, mapped, and actually consumed by the delay calculation.
  5. 05 Why can a change in RC alone also change cell delay, even when the cell itself is untouched?Expert · Because a cell's delay depends on its output slew, and output slew depends on the load it drives — which includes the interconnect. Change the RC, and you change the effective load and the receiver's input slew, which changes the delay the next cell reports, even though that next cell never physically changed.
  6. 06 Why might the globally worst-WNS scenario not be enough, by itself, to decide which views can be removed from a signoff plan?Expert · Because WNS is only one lens on risk. A view can be required for unique endpoints, a specific check type, a distinct optimization stage, or regulatory/product coverage reasons even if it has never once been the single worst number in the run.
  7. 07 Graph-based analysis (GBA) and path-based analysis (PBA) give significantly different slack on the same critical path. What's going on, and which one do you trust?Expert · GBA propagates conservative, graph-level worst-case assumptions efficiently across the whole design; PBA recalculates a specific path with path-specific context, removing some of that built-in pessimism. Neither one is universally "more correct" — you compare them under aligned settings and understand what each is actually measuring before drawing a conclusion.
  8. 08 Why can enabling signal integrity (SI) analysis change which corner or view is actually the worst one?Expert · Because SI introduces a new dependency that plain nominal timing doesn't capture: coupling between neighboring nets, which depends on aggressor switching alignment, direction, and effective loading — and that dependency doesn't scale uniformly with the same PVT or RC assumptions that make a corner look "worst" without SI.
  9. 09 A view is named FUNC_SS_RCMAX, but its actual configured objects point to an FF library and a different RC condition. What does the timing engine use?Expert · The configured object associations — not the human-readable name. The engine doesn't parse the name string for meaning; it uses whatever library, RC corner, and constraint mode are actually attached to that view object.
  10. 10 An engineer tells you "timing is clean." As the signoff lead, what do you ask before accepting that statement?Expert · Which modes, which corners, which setup views, which hold views, and were all the views required by the signoff plan actually active. Which libraries and RC/parasitic data were used, and how was variation (OCV/AOCV/POCV) and SI handled. Whether unconstrained paths were reviewed, exceptions validated, extraction is current, and whether any waivers are hiding underneath the headline result.