Multi-Corner STA — MMMC POCV/SOCV with LVF

Scope

This document records the multi-corner static-timing-analysis methodology across the open-tooling MVP (today, Sky130 SS/TT/FF x min/max RC) and the 2028 advanced-node target (100-200 corners typical at N3/N2, ML-pruned to 32-64). It is the human-readable companion to docs/evidence/pd/multi-corner-sta-evidence.yaml.

MVP (Sky130 + IHP SG13G2): 6 corners

Driver: scripts/run_multi_corner_sta.py calls OpenSTA (or OpenROAD's internal STA engine) once per corner. Outputs:

• {corner}.tcl script that read_liberty / read_verilog / read_sdc / read_spef and runs report_checks.
• {corner}.rpt four-line digest (setup/hold WNS + TNS).
• multi_corner_sta.json aggregate summary.

OpenSTA is not installed natively on every developer host. The container-friendly invocation is:

docker run --rm -v "$PWD":/work -w /work \
    -e PDK_ROOT=/work/external/pdks \
    ghcr.io/efabless/openlane2:2.4.0.dev1 \
    python3 scripts/run_multi_corner_sta.py \
        --run-dir pd/openlane/runs/<RUN_TAG> \
        --out-dir build/pd/multi_corner_sta/<RUN_TAG> \
        --pdk-root /work/external/pdks/volare/sky130/versions/0fe599b2afb6708d281543108caf8310912f54af

This image ships OpenSTA at /nix/store/.../opensta/bin/sta which is on PATH inside the container.

Acceptance (Stage 1):

• All 6 corners run to completion.
• TT_max setup_wns >= 0.
• FF_min hold_wns >= 0.
• SS_max setup_wns >= 0 OR the failing paths land in the architectural exception list under pd/constraints/.

Stage 3 (advanced node, BLOCKED on commercial EDA)

At N3/N2 the multi-corner space is fundamentally bigger:

Full Cartesian product is 100-200 corners. Running all of them is prohibitive. The industry-standard approach in 2025+ is:

• POCV/SOCV with LVF Liberty: statistical delay model rather than worst case + derate. Cuts pessimism by 5-15 %.
• ML corner pruning: train a regressor on a subset of corners to predict which corners are most likely to gate setup/hold/slew. Run the expensive PrimeTime path-based STA only on the top 32-64 predicted corners.
• Path-based vs graph-based: graph-based first pass for coverage, path-based second pass for the violators.

Tool stack: PrimeTime SI / Tempus for path-based, plus a thin ML predictor trained on historical reports. This is all BLOCKED on the commercial EDA gate (docs/evidence/pd/commercial-eda-gate.yaml).

Why exercise STA methodology now

The methodology-validation discipline matters:

• Driving multi-corner STA on Sky130 forces us to discover SDC bugs, exception leakage, and corner-skew bugs at a stage where one Liberty per corner is fast and free.
• The aggregate JSON schema (eliza.pd_multi_corner_sta.v1) is the same shape we will use at the advanced node. Only the corner count and the derate model change.

What unblocks the multi-corner-sta-evidence gate

For Stage 1:

• A full Sky130 release run produces all 6 corner reports.
• All 6 corners have setup_wns >= 0 (or documented exception).
• All 6 corners have hold_wns >= 0.

For Stage 3:

• Commercial-EDA gate unblocks.
• PrimeTime SI runs path-based STA on >= 32 N3/N2 corners.
• POCV/SOCV LVF Liberty in use.
• ML corner-pruning trained and validated against the full Cartesian subset for at least one block.