bspdn-goal-driver
Drive BSPDN objective optimization toward target outcomes (about 8% dynamic-power reduction without area/timing regression, or 5%+ frequency uplift with non-worse power/area) using gated experiments and evidence tracking.
BSPDN Goal Driver
When to use
Use this skill when the user asks to push objective/performance toward explicit PPA goals, not just run exploratory experiments.
Goal contract
Target A:
- dynamic power reduction about
8%, - area/timing non-worse (
delta_area<=0,delta_wns>=0,delta_tns>=0).
Target B:
- power/area non-worse,
- frequency uplift
>=5%validated by period-sweep evidence.
Milestone policy (incremental, mandatory)
Do not pursue one-shot final target. Use staged gates:
- M0: evidence-valid baseline and repeatable A/B flow (
execution-contract PASS). - M1:
power <= -1%with non-worse area/timing. - M2:
power <= -3%with non-worse area/timing. - M3:
power <= -5%with non-worse area/timing. - M4: final target A (
~ -8%) or target B (>=5%frequency uplift with non-worse power/area).
Promotion to next milestone requires current milestone PASS + retrospective note.
Workflow
- Build/update goal tracker:
python3 scripts/debug/track_bspdn_goal_progress.py --campaign-md <ppa_md> --out-prefix <prefix>
- Select testcase tier:
- Tier-1: high backside-usage potential with route feasibility,
- Tier-2: mechanism-isolation testcase for fast iteration.
- Run pre-submit checks:
eda-preflight-reflect- route smoke (
smoke_targeted_backside_route_toycase.sh) for targeted net lists - execution policy lock (
vanilla_replaceprimary baseline)
- Execute A/B campaign with strict gates:
- shared baseline and route policy,
- explicit net-selection policy/version tag,
- collect route PPA and execution-contract artifacts.
- Decision:
GOonly when current milestone gate is satisfied with valid artifacts,- otherwise update hypotheses and next-step campaign.
Outputs
bspdn_goal_progress.summary.mdbspdn_goal_progress.detail.tsv- campaign plan (
goal_campaign_plan.md) - evidence-backed milestone decision note (
M0/M1/.../M4).
Reporting policy (user-facing)
- Do not generate slides for every process stage.
- Generate a PDF version-summary only when a new validated version reaches a reporting milestone:
- first reach
>=2%power reduction with non-worse area/timing, - then each higher validated step (for example
>=3%,>=5%, final target), - or Goal-B (
>=5%frequency uplift with non-worse power/area).
- first reach
- Each version-summary slide deck must include method delta, key conclusions, core PPA table, and next-step plan.
Hard rules
- No promotion claim without execution-contract PASS.
- Goal-B frequency claim must come from period sweep, not WNS proxy alone.
- Record exact selector/route-hint config for each compared run.