Small QoR moves have two families of causes. Only one of them is the RTL.
Source | Family | Answer |
|---|---|---|
Rerun variation | Noise | Same inputs, slightly different numbers; the band absorbs it. |
Tool or library drift | Noise | Unpinned revisions move numbers without touching the design. |
Constraint edit | Real | A new target redefines slack; re-baseline deliberately. |
RTL change | Real | The design moved; the regression names the metric. |
Flow script edit | Real | A stage change is a flow change; re-baseline deliberately. |
A band turns a wiggle into a verdict. The calc block replays three EXAMPLE reruns of an unchanged 4-bit adder against the 5% area band.
EXAMPLE reruns, unchanged inputs, course.lib area units: 51, 52, 53
mean = (51 + 52 + 53) / 3 = 156 / 3 = 52
range = 53 - 51 = 2 units = 2 / 52 = 3.85% of the mean
band: 5% of 52, so readings up to 54 stay noise (54 x 100 = 5400 < 5460)
verdict: 3.85% < 5%, so the spread is noise; no debugA suspect move is rerun with every revision pinned: same tool, same library file, same constraints, same script. Baseline and current rerun side by side, and the band judges.
Version-pinned libraries make this possible. The library file travels with the run as a recorded revision, never as a floating path.
No. The 3.85% spread sits inside the band, so it is noise.
Unproven. Repin the recorded library revision and rerun first; unpinned libraries are noise until pinned.