Extraction and signoff correlation

14.1 Stage purpose

RC extraction measures the resistance (R) and capacitance (C) of the routed wires and vias. The result is a set of parasitics that delay calculation turns into net delay and into the load seen by each cell. Innovus stores them in an RC database, written RCDB, and can write them out as SPEF, the Standard Parasitic Exchange Format, one file for each RC corner. An RC corner is a named set of process data, temperature and scale factors for extraction.

Two extractors matter in this chapter. The in-tool extractor produces the numbers that optimisation and in-tool timing use. The signoff extractor, Standalone Quantus, produces the numbers that the signoff timing run uses. Both read the same geometry, but they model the capacitance differently, and the in-tool engines trade some accuracy for speed. The gap between them is the subject of this chapter.

Capacitance between wires deserves a closer look, because it is where the engines differ most. Figure 20 shows one victim wire with two neighbours on the same layer and one wire above it. Each neighbour couples to the victim, and the capacitance to the substrate and to everything else is the ground capacitance. To save run time, the native extractor lumps a small coupling capacitance into the ground capacitance. Three thresholds decide which couplings are small. A filter mode combines the relative and the coupling thresholds, and the total threshold acts on its own. Lumping keeps the total capacitance almost unchanged and removes the aggressor, so the timing tool can no longer see the crosstalk delay that aggressor would cause.

This chapter keeps four things apart. Tool completion means the extraction or timing command returned. Analysis coverage means every net has parasitics in every active view and RC corner. Stage qualification means the gap between in-tool and signoff results is known and inside your project budget. Signoff is the project signoff run itself, which this chapter does not replace.

Three evidence classes are used. Preliminary evidence comes from preRoute extraction, from native extraction whose scale factors have not been checked against signoff, or from a run on an unfinished route. Implementation evidence comes from postRoute extraction and timing on the final routed database. Signoff evidence comes from Standalone Quantus extraction and the signoff timing engine, run with the project signoff setup on the final database. Whether a signoffTimeDesign run inside Innovus counts as your project signoff evidence is a project decision. A difference between two runs of different classes is a correlation result. It is not a pass.

Scale factors are the second tool. A scale factor multiplies the extracted capacitance, coupling capacitance or resistance of an RC corner, so that the native numbers move toward the signoff numbers. The ledger shows why a scale factor is not a cure for everything. The fitted cap and res factors are small and describe an engine offset. The fitted coupling factor of 1.37 describes something else: two couplings that the filter removed. Match the filter first, and fit only what is left. The factors come from one net here only to show the idea: a real fit uses every net of the SPEF pair, as generateRCFactor does.

The flow that the rest of the chapter follows is drawn in Figure 21 in Section 14.7.1.

14.2 Entry prerequisites

14.3 Relevant files and analysis context

The inputs are the routed database, the capacitance table or Quantus technology file for each RC corner, and, when signoff extraction runs from Innovus, a layer map between the LEF names and the technology file. The reference states that the layer map is required for Standalone Quantus, and that fill needs a stream layer map as well when its data is passed to Quantus. The outputs are the RCDB in memory, saved RCDB directories, SPEF files, and the timing report directories of both runs.

Four groups of settings decide what the numbers mean: setExtractRCMode, the RC corner scale factors, setAnalysisMode with setDelayCalMode and setSIMode, and the timing derates. The derates deserve a warning. The reference states that set_timing_derate can only be used on the command line and cannot be imported as part of an SDC file. Innovus therefore takes derates only from the session, and a derate written in an SDC does not reach it. A signoff run may read derates from its own SDC or scripts, so the mismatch can run in either direction. Compare both sides.

The scale factors belong to the RC corner, so saving the design with saveDesign saves them in the view definition file. Restoring the design restores them. Thus, record the factors with every comparison, because a factor that is applied silently moves the numbers it is meant to explain.

Use the same routed database for both extractors. The reference states that for postRoute correlation the native extraction must run on the same routed design that produced the signoff SPEF. A signoff SPEF from a design that has had one more ECO is a comparison of two different designs.

14.4 Checks and command cards

14.4.1 Pre-stage checks

Before extraction, store four baselines so that a later difference can be attributed: the routed database saved with saveDesign after the fill checks of Chapter 13, the settings records of X-01 and X-07, the post-route timing baseline of Chapters 11 and 12, and the constraint and derate record of the signoff setup.

14.4.2 Post-stage checks

report_unit_parasitics -all_layers -rc_corner <rc>
get_rc_corner <rc> -rcCornerAttribute
getExtractRCMode -engine -effortLevel -coupled
getExtractRCMode -capFilterMode -coupling_c_th -relative_c_th -total_c_th

setExtractRCMode -engine postRoute -effortLevel <level>
extractRC
saveRC <name>.rcdb.d
report_rcdb <name>.rcdb.d

report_annotated_parasitics -view <view>
report_annotated_parasitics -view <view> -list_not_annotated -max_missing <n> > <file>

rcOut -spef <name>.spef.gz -rc_corner <rc>
spefIn <signoff>.spef.gz -rc_corner <rc>

report_net_parasitics <net> -rc_corner <rc>
reportDelayCalculation -from <pin> -to <pin> -view <view> -si

generateRCFactor -preroute false -postroute low -reference signoff -outputFile <f>
update_rc_corner -name <rc> -postRoute_cap {<f>} -postRoute_res {<f>}

getAnalysisMode -analysisType -cppr
getDelayCalMode -SIAware
report_timing_derate
get_global timing_cppr_threshold_ps
getDesignMode -earlyPBAMode -pessimisticMode

signoffTimeDesign -outDir <dir>
report_timing -view <view> -path_type end
timeDesign -postRoute -outDir <dir>
signoffTimeDesign -reportOnly -outDir <dir>
report_timing -view <view> -path_type full_clock -from <pin> -to <pin>

14.5 Required reports, artefacts and how to read them

report_net_parasitics v -rc_corner rc_max -file_name v_native.rpt     ## in-tool, illustrative
Net: v
RC Corner: rc_max
Total Cap = 1.455 ff
Total Res = 134.800 ohms
Total XCap = 0.499 ff
Cap Distribution:
Total Ground Cap = 0.956 ff
XCap With Nets:
a1 = 0.499 ff

report_net_parasitics v -rc_corner rc_max -file_name v_signoff.rpt     ## signoff SPEF read in, illustrative
Net: v
RC Corner: rc_max
Total Cap = 1.476 ff
Total Res = 140.400 ohms
Total XCap = 0.684 ff
Cap Distribution:
Total Ground Cap = 0.792 ff
XCap With Nets:
a1 = 0.506 ff
a2 = 0.138 ff
x = 0.040 ff

Let us read the pair against Figure 20. Total capacitance differs by 1.4 percent and resistance by 4.0 percent, both low in the native run. Offsets of this size are what a scale factor is for, and both are small enough to judge against the project budget. The coupling is different. The native report shows 0.499 fF of coupling and one attacker, the signoff report 0.684 fF and three. The ground capacitance in the native report is 0.164 fF higher. The two missing attackers carry 0.178 fF, so most of that capacitance moved to ground.

Thus, the two reports agree where a quick glance looks, on total capacitance, and disagree where timing is sensitive. A fitted coupling factor of 1.37 would make the coupling number look right, but total capacitance would then read 1.640 fF against 1.476 fF at signoff, about 11 percent high, because the lumped capacitance is still in the ground term. The decision is to compare the thresholds of X-01 with the signoff command file, and to fit scale factors only after that. All values here are illustrative and are not from a tool run.

14.6 Healthy, suspicious and hard-stop examples

14.7 Debugging, corrective action and reruns

A timing gap has several independent sources, and they hide each other. Change one source at a time, rerun X-08 after each change, and write down the gap that remains. The table lists the usual sources, the symptom that points to each, and the card that checks it.

14.7.1 How to bound the remaining gap

To bound a difference means to give an upper limit for it and to show where it came from. Three steps do that. Figure 21 shows the flow they belong to.

First, make the two runs comparable, and do it before anything else: the same routed database, the same SDC, the same views, the same derates, OCV, CPPR and SI settings, and the same RC corner data. The strip at the bottom of Figure 21 lists them. A settings difference found later hides inside every extraction number, and a difference found before this step is not yet a measurement.

Second, decompose the extraction gap. Align the filters, then fit the scale factors, and record the gap after each change in a short table. A source that moves the gap is a cause. A source that does not move it is ruled out.

Third, state the residual against the endpoint list and not only against WNS. Take the endpoint lines from X-08 for both runs, subtract the slack of matching endpoints, and report the largest difference and its sign. The sign matters. A residual where the in-tool run is optimistic is a risk. A residual where it is pessimistic costs area and power. Your project budget decides how large either may be. The comparison script is yours to write, since the endpoint lines are plain text. This book prints no database path for it.

Step 1 changes the in-tool number. Innovus cannot import set_timing_derate from an SDC file, so the derates in the signoff SDC never reached the session, and part of the original gap was a settings error and not an extraction difference. Step 2 moves the in-tool number again because the removed attackers return, as in Figure 20. Step 3 removes the engine offset. The residual of 0.007 ns is what extraction and delay calculation still differ by, and it is a number you can compare with the project budget.

The decision depends on that budget. If the residual is inside it, X-08 is a pass for this view, and you rerun X-03 to X-08 for the other views. If it is outside, the next sources in the table are path-based against graph-based analysis and SI mode. Note that both runs now fail setup. Closing timing is the work of Chapter 12, and this chapter has shown that the failure is real and not an artefact of the extractor.