Back to Eliza

PDN and current-budget local evidence

packages/chip/docs/pd/signoff/pdn-current/local-budget.md

2.0.12.3 KB
Original Source

PDN and current-budget local evidence

Status: draft_local_evidence Release use: prohibited_until_external_review

This budget is derived from local OpenLane artifacts only. It is useful for triage and first-article planning, but it is not a tapeout or board-fabrication release approval.

Selected local run

pd/openlane/runs/RUN_2026-05-18_05-41-42

Key local artifacts:

  • final/metrics.json
  • final/metrics.csv
  • 57-openroad-irdropreport/irdrop.rpt
  • 20-openroad-generatepdn/openroad-generatepdn.log
  • signoff-run.yaml
  • benchmarks/power/local-estimates/e1-npu-openlane-npu-estimates.json

Derived budget

RailBoard netVoltageLocal current basisLocal budget
VDDCORE+1V81.8 V5.554 mW / 1.8 V = 3.086 mA6.172 mA with 2x local margin
VDDIO+3V33.3 Vunavailableblocked pending IO/package/board load model

For first-article planning only, both rails should start with a 25 mA bench current limit until workload-calibrated power and board loads replace this local estimate.

The local estimate artifact also records the NPU architecture-model result (47.663 to 49.152 modeled INT8 TOPS) and the invalid arithmetic that would result from dividing it by OpenLane power (8581.556 to 8849.730 TOPS/W). That arithmetic is kept only as a blocker tripwire: the power and TOPS values come from different substrates and cannot support sustained efficiency claims.

Local PDN observations

  • OpenROAD reported all shapes connected on VPWR and VGND.
  • VPWR worst local IR drop is 87.6137 uV at nom_tt_025C_1v80.
  • VGND worst local bounce is 105.98 uV at nom_tt_025C_1v80.
  • Metrics report zero power-grid violations for VPWR and VGND.

Release blockers

  • Replace local metrics-derived power with vector/workload-calibrated post-route power across selected release corners.
  • Capture a measured sustained power/thermal run where TOPS, rail power, frequency, temperature, and throttle state share the same window.
  • Archive EM evidence and any foundry-required current density checks.
  • Derive VDDIO current from released IO pads, package model, and board loads.
  • Archive board regulator, fuse, thermal, and first-article current-limit review.
  • Tie this budget into released package, padframe, board, and thermal evidence.