IR-Drop / EM Triage
Power integrity failures can turn a timing-clean chip into a field failure. This skill reads native IR-drop and EM reports, pinpoints hot regions, and proposes power-grid fixes. The native tool result against its original limit determines signoff.
When to use
- After initial power grid insertion
- After placement (static IR)
- After routing + SDF back-annotation (dynamic IR)
- Whenever tapeout checklist flags an EM-density violation
Inputs
- IR-drop report (static + dynamic)
- EM report (current density per metal layer)
- Power grid definition
- Switching activity file (VCD, SAIF, or propagation from simulation)
- Floorplan with macro/standard-cell placement
Workflow
- Hotspot map + budget pass/fail — DETERMINISTIC, enforced by
programs/ir_drop_budget_check.py. It computesbudget_mV = pct·Vdd·1000(the 5–10%-of-Vdd budget; 10% permissive default) and FAILs iff the worst-case measured drop ≥ budget. Do NOT re-derive the budget by hand — call the program. - Classify the measured cause. Inspect the native voltage/current map, grid connectivity, via arrays, strap pitch, metal width and switching activity at each hotspot. Separate sparse straps, narrow metal, weak via connections and local switching demand using the tool's actual data.
- Choose a physical fix. Add straps or vias, widen metal, add decoupling, or reduce local switching demand when the measured cause supports it. Rerun native IR and EM analysis against the same original budgets.
Native PDN implementation choices
Read the failing segment's layer, coordinates, current, cross-section and original limit from the native EM report. Distinguish a metal-width violation from a per-cut via violation before choosing a repair. A width-scaled utilization is only a prediction; rebuilding the grid can redistribute current.
For the native flow, use programs/flow_config.py --project <project> --route ic --set KEY=JSON to record the selected PDN_* setting. The program preserves
other choices in the existing plugin_output/declaration.json and emits the
configuration that consumes the move. This is an implementation choice, not a
new specification limit or a required declaration contract. Do not change the
user's prompt or generated L documents to express it.
Keep the original report and the rollback value. Resume through the public entry from PDN generation, then remeasure EM, IR, timing and physical verification with their original limits. A changed declaration alone does not invalidate a retained result or establish that the candidate ran.
Output format
power/ir_triage.md:- Hotspot table with coordinates, cell, drop %, fix
- EM violation table with metal layer, current density, fix
- P&R script snippets (strap insertion, decap placement)
- Estimated impact on routing congestion
Technical basis
Power integrity references: Tan & Roy "On-Chip Power Delivery and Management". Commercial signoff: Ansys RedHawk, Cadence Voltus, Synopsys PrimePower. Open-source: OpenROAD psm (PDN analysis).
Handoff
- PDN changes →
/flow-orchestrateto re-run PDN + route - Timing impact of sizing changes →
/sta-review - Decap placement →
/placement-optimize
Captured rule
Read the native EM report's layer, coordinates, current, cut count, area and
original current-density limit before choosing a repair. A small IR drop does
not clear an overloaded via array. For a via-array bottleneck, use the native
PDN geometry options through flow_config.py to widen the landing or provide
more connections, then regenerate PDN and rerun the affected routing, timing,
physical checks, IR and EM against their original limits. Preserve a proven
design requirement in its own input documents so Phase 1 carries it into L19;
do not make one measured width a global default or use a hand-set declaration
as proof of document ingestion.
An EM via-array bottleneck is not repaired by IR drop alone. Subservient's native Classic reference had only low IR drop but514 EM violations /664.63% utilization at the original100% limit. A wider landing alone still failed313/390.73%; retain that failure. Derive denser template-grid proposals from measured native EM, then let original native EM/DRC/routing/timing decide. Include native CTS/slew choices when clock power/fanout remains a constraint; do not waive fanout1 just because33 earlier rows PASS. The program-derived eligible complete candidate measured EM0/69.98%, DRC/LVS0 and44/44 native rows PASS on image0.4.76. The actual proposer/selector (not a hand setting or this lesson) captures the recovery.
Captured rule
Before changing a power grid for a via electromigration failure, read what the
report compared. A foundry states a via limit as a current per cut; a density
over the cut area against that same number is a different, far stricter rule.
The native report says which it used ("Vias judged per cut", and the basis
column of the per-segment rows). When a PDK's cut limits are per via, its cited
per-via table belongs in programs/pdk_registry.json (em_current_limits,
with source, temperature and cut size), and the same per-cut comparison then
serves every design on that PDK. Measured on SKY130 (the technology LEF says
"mA per via"): a hard macro's rail-to-strap crossings carried at most 0.109 mA
through 10 cuts. Read as a density that was 19 violations, worst 166.44 %; per
cut it is 3.7 % of the 0.29 mA limit, and the row passed with 0 violations,
worst 8.08 % (met1), on identical segment currents. The limit itself is
unchanged. Widening the straps does not address such a finding.
Captured rule
A connected, unloaded power grid can have a valid native PSM solution with solved supply voltages and zero load current. An empty load-current map does not invalidate that solution. Judge IR drop and electromigration from the native reports against the original limits. Do not invent a load to obtain a solution or waive an unmeasured result.
An unloaded-grid result does not establish loaded-grid performance. Include the design's actual power demand when evaluating loaded IR drop and EM.