Skip to content

B04 Increment 29 — Realistic-Density Routed-Block External Validity

Increment 28 demonstrated strong work avoidance on sparse layouts, including a real Sky130 standard-cell motif placed sparsely inside synthetic large domains. That is useful but insufficient for an industrial promotion decision.

Increment 29 attacks the exact failure mode named in the industry-value protocol:

scaling gain exists only on synthetic sparse layouts.

Pinned public fixture

Dataset:

SJTU-YONGFU-RESEARCH-GRP/PDB-Physical-Design-Database
commit 9e1e3399b1b707f26fee853bce1ff91ab466ce24
layout/sky130hd/ibex/ibex.gds
layout/sky130hd/ibex/ibex.json

The public metadata identifies:

design: ibex
top cell: ibex_core
PDK: sky130hd
standard-cell count: 15515
core area: 300532 um^2
stdcell area: 173954 um^2
stdcell utilization: 0.578821

The GDS is an OpenROAD-generated routed physical design, not a single-cell placement invented for B04.

Deterministic crop rule

The benchmark uses nested center crops of the real routed block at 1 nm exact vector sampling:

4096^2
8192^2
16384^2

and a 32768^2 screen-only diagnostic.

The crop is exposed through ExactVectorCropSource, which translates local window/run queries into the parent GDS coordinate system. No parent or crop full raster is materialized by the vector path.

Primary question

Increment 29 is deliberately hostile to the current screening strategy:

Does exact-empty-context screening still avoid enough forward work
inside a ~58% utilized real routed block?

This is more important than obtaining another favorable speed curve.

At the largest fully executed crop define

[ a_{\rm real} = A_{\rm active}/A_{\rm crop}. ]

The decision rule is:

a_real <= 0.50     STRONG
0.50 < a_real <= .80  MODERATE
a_real > 0.80      WEAK

These thresholds are engineering stop/go thresholds, not mathematical constants.

If STRONG

Keep exact-empty screening and proceed to survivor-only continuous verifier integration.

If MODERATE

Retain empty screening but add a stronger theorem-safe screen for nonempty tiles before promoting the industrial claim.

If WEAK

Do not spend effort optimizing exact-empty screening. Redirect the next work-avoidance increment to nonempty-tile screening, for example:

cheap nominal field / geometric margin
-> certified no-threshold-band condition
-> skip continuous contour machinery

The intended invariant is not "empty geometry"; it is "provably irrelevant to the requested lithography metric".

Baselines and timing

At 4096^2 the benchmark runs:

dense_full
tiled_raster
b04_vector
b04_selective

At 8192^2:

b04_vector
b04_selective

At 16384^2:

b04_selective

The 32768^2 level performs screening only, to measure active-set discovery without risking billions of forward pixels if the real block is dense.

Every worker reports both:

parent_load_wall_s
execution_wall_s
end_to_end_wall_s

so a favorable post-parse result cannot hide an expensive real-GDS parser cost.

Claim firewall

This benchmark may support:

  • work-avoidance external validity on a real routed block crop;
  • real-GDS parser cost;
  • no-full-raster scaling for the vector path;
  • real-density active fraction;
  • common-size wall-time comparisons.

It does not support:

  • foundry-qualified wafer prediction;
  • entire-Ibex full-chip verification;
  • calibrated Hopkins speedup;
  • continuous EPE/CD/process-window certification at these scales.

Its main scientific value is the stop/go decision on the screening invariant.