Path Depth Sensitivity: Expanded Case Study

Extended case study

Signoff review: short-path hold margin, long-path setup margin, depth bucket WNS fails after an apparently minor change affecting Path Depth Sensitivity.

Background

Team had closure last week on full MMMC matrix. New run shows regression clustered in one path group.

Symptoms observed

• short-path hold margin, long-path setup margin, depth bucket WNS regression

• Disagreement between synth and STA WNS

• Hold violations after setup ECO

Investigation timeline

1. Hour 0: freeze netlist, SDC, SPEF, lib tags

2. Hour 1: diff constraints and run manifest

3. Hour 2: isolate corner/mode that fails

4. Hour 3: read worst path — cell vs net, clocks

5. Hour 4: classify root cause

6. Hour 5: minimal ECO or SDC fix

7. Hour 6: full regression + signoff memo

Root cause

Root cause tied to Path Depth Sensitivity: Short paths see different variation pessimism than long paths; depth-aware models prevent both optimism and excess margin.

Fix and validation

• Bounded ECO or audited SDC change

• Re-run path depth vs slack scatter, AOCV table dump, hold/setup split report

• Full MMMC + SI regression

Lessons learned

• Tag every run

• Diff constraints before blaming PD

• Setup and hold together

CASE STUDY — Path Depth Sensitivity
baseline WNS / regressed WNS / post-fix WNS

Sequence under stress

TIMING PATH — Path Depth Sensitivity

launch FF --logic stages--> capture FF
     |                           |
  launch clock                capture clock
     |                           |
  required: arrival < capture_edge - setup
  metric: short-path hold margin, long-path setup margin, depth bucket WNS

STA deep dive

Variation policy is negotiated margin, not a tool checkbox.

Concept diagram

VARIATION STACK

nominal delay + OCV/AOCV/POCV margin = signoff delay

Metric graph

DERATE IMPACT ON WNS

OCV only     WNS -120ps
AOCV         WNS -45ps
POCV         WNS -60ps (policy dependent)

Reports and artifacts

• derate table

• AOCV depth report

• POCV sigma

• hold/setup split

Mini case study

Switched to AOCV without re-baselining hold — short paths failed min analysis.

Debug branches

• Short paths for hold

• Policy approval for derate changes

Senior review question

Ask: what corner/mode/view proves this topic is closed or failing?

Key takeaways

• State corner, mode, view, and database tag with every slack claim.

• Run setup and hold plus MMMC regression after every ECO.

Common pitfalls

• Comparing STA runs with different SPEF or SDC tags.

• Broad false_path to green-wash violations.

• Setup-only ECO without hold check.

Principal STA review addendum

Short paths see different variation pessimism than long paths; depth-aware models prevent both optimism and excess margin.

Metric: short-path hold margin, long-path setup margin, depth bucket WNS