PPA Predict
Read this before the first number: this is an estimate
Every number this skill produces carries status ESTIMATED, and
docs/PPA_INTERFACES.md §2 is unambiguous about what that status may do:
| status | may enter a numeric comparison |
|---|---|
MEASURED |
yes |
ESTIMATED |
never in final PPA |
So a number from this skill may be used to choose between two RTL variants, to size a floorplan before it exists, or to decide whether a synthesis run is worth starting. It may not be used to answer whether a design met its target, it may not be carried into a PPA report, and it may not be compared against a number that came from an artefact. A prediction and a measurement look identical once they are two cells of the same table, which is why they are never allowed to become two cells of the same table.
The skill that measures is ppa-measure. It reads the actual native tool artifacts and reports their measured values
with scope and source paths. This skill has read no artefact —
that is the whole point of it, and it is also the whole limit of it.
This skill also produces no gate verdict. Whether an estimate is good enough to proceed is a caller's decision or a program's; it is not a line in this report.
Doctrine (v0.1.50): 把修法寫進工具,而非寫進 prompt。 Use native tool measurements when available and label estimates explicitly.
Provide fast, pre-synthesis estimates of Power, Performance (Fmax / critical-path delay), and Area for an RTL module. Acts as a Circuit Foundation Model surrogate for the synthesis tool — seconds instead of hours.
Prepare an explicit estimate
Read the actual design input, RTL structure and selected PDK characterization. Use a native synthesis cell count if one is already available, and state that scope separately from a purely pre-synthesis estimate. Name the area, delay, activity, capacitance and voltage assumptions behind the model.
Report ranges and units. Keep every predicted value labelled ESTIMATED, state what the model omits, and use native tool measurements before claiming a PPA requirement is met. No retired estimator program supplies a JSON result or exit code for this workflow.
When to use
Trigger when the user:
- Wants an area/timing ballpark before running DC/Genus/Yosys
- Is comparing two RTL variants and needs a quick "which is better"
- Asks "will this fit in X gates" or "can I close timing at Y MHz"
- Needs PPA feedback inside a generate-evaluate loop
Inputs to gather
- Target technology (e.g. an open PDK such as
gf180mcuD,sky130Aorihp-sg13g2; or a generic standard-cell assumption) - Target clock frequency (MHz)
- Target cell library (if known) or standard-cell assumption
- Optimization goal: area, speed, power, or balanced
Prediction workflow
- Structural analysis — count flops, adders, multipliers, memory bits, mux layers
- Critical-path estimate — identify the deepest combinational chain and estimate logic-level delay
- Gate-count estimate — map each construct to approximate NAND2 equivalents
- Power estimate — dynamic (activity × capacitance × V² × f) + leakage (gate count × node leakage)
- Confidence band — report each number as a range, not a point estimate
Output format
# PPA Prediction — <module>
Status: ESTIMATED (never enters final PPA)
Not a measurement: run /ppa-measure for numbers parsed from tool artefacts
Technology: <node>
Target frequency: <MHz>
Every row below is an ESTIMATED record. None of them has a source artefact,
because at this point in the flow no artefact exists.
| Metric | Status | Estimate | Range | Confidence |
|--------|--------|----------|-------|------------|
| Area (µm²) | ESTIMATED | ... | ... – ... | High/Med/Low |
| Gate count (NAND2 eq) | ESTIMATED | ... | ... – ... | ... |
| Fmax (MHz) | ESTIMATED | ... | ... – ... | ... |
| Dynamic power (mW @ target f) | ESTIMATED | ... | ... – ... | ... |
| Leakage (µW) | ESTIMATED | ... | ... – ... | ... |
## Summary
<one paragraph: what was estimated, from what, and how far it can be trusted>
## Critical path (estimated)
<signal A> → <logic> → <signal B> (~<n> logic levels)
## Bottleneck
<what is limiting Fmax or dominating area>
## Optimization suggestions
- <suggestion 1, e.g., pipeline the multiplier>
- <suggestion 2, e.g., share the adder>
Next: run /ppa-measure
Technical basis
Grounded in Circuit Foundation Models for pre-synthesis prediction — encoder-based CFMs like MasterRTL, CircuitEncoder, and graph-neural-network surrogates trained on synthesized netlists. Typical prediction error: 10–15% vs real synthesis for area, 15–20% for timing.
Do not
- Do not claim single-number precision — always give a range
- Do not replace real synthesis for sign-off; this is a pre-check only
- Do not extrapolate far outside the training distribution (e.g., very exotic architectures)
- Do not label any number here
MEASURED, and do not attach a source artefact hash to one. Nothing was parsed; a provenance field on an estimate is a claim that a reader can only discover is false by trying to re-derive it. - Do not attribute an estimate to a post-route or extracted stage. This skill runs before those stages exist, so a stage label from them is not an optimistic approximation — it is a different number's label on this number.
- Do not put an estimate and a measurement in the same table, the same chart, or
the same sentence with a comparison in it. If both are needed, they are two
documents: this one, and a
ppa-measurereport. - Do not answer whether the design met its target from these numbers. That
question is answered from
MEASUREDrecords by a deterministic program.
Preserve the ECO reserve
Keep unused spare cells, ECO gates and reserved pads through synthesis and
physical repair. Compare the native netlist or database before and after the
change. No unused reserve instance may disappear, and its keep attributes must
remain intact. A consumed spare must be named in the intended ECO. Restore any
unintentionally removed spare before rerunning the native tools. See
design-for-eco for the design intent.
Captured rule
Treat a document-authorized FP_CORE_UTIL percentage range as an implementation
choice, not an unparseable scalar: choose its lower (roomier) endpoint while
retaining the original range and all electrical/physical acceptance limits.
Do not generalize this selection to timing, power, or DRC limits. Keep scalar,
PDK-scoped, and explicit implementation overrides authoritative.
Measured on Subservient SKY130A, LibreLane Classic, image 0.4.76: ignoring the
35–45% document range selected the native default; post-CTS legalization rose
from 94.3% to 106.1% after 459 hold buffers and failed DPL-0038. Selecting the
authorized 35% made post-CTS repair PASS, with native instance utilization
0.741263, setup slack +5.46794 ns and hold slack +0.281741 ns. Routed repair
also completed; candidate SI still FAILed its unchanged zero-slack limit
(hold −0.01205236 ns, setup +0.26383874 ns). This is not full-chain closure.
The bounded Chip control remained 28 PASS / 1 FAIL with identical routed
setup −5.38684845 ns and hold −0.04831780 ns. Native config tests: baseline
102 passed; changed 108 passed, no failures. Evidence packets:
subservient-{ic,ip}-core-util-range-i0476-oai-20261109 and
subservient-core-util-range-unit-i0476-oai.
Captured rule
For a macro floorplan power-connection failure, compare the LEF supply-pin types, current ODB pin bindings, parent netlist and supplied connection hooks. Zero new connections can mean a correctly typed pin is already bound to a different rail. Resolve the binding from the admitted design intent in an explicit authored native configuration and disclose any override of supplied settings. Preserve the fixed interface and numeric physical constraints; do not remove hooks or force a conflicting connection to silence the error.