How to read a RedHawk dynamic IR drop report: files, metrics and glossary

A dynamic IR drop run is answered by four files: the log’s Worst Dynamic Voltage Drop table, Dynamic/{design}.dvd, layer_drop.rpt and power_summary.rpt — corner and window first.

0. Reading order (follow it, or you will draw false conclusions)

redhawk -b always exits with status 0, so the exit code proves nothing. Validity is judged only by the absence of PWR-330 / ANL-013 / GEN-010 and by what adsRpt/Error/redhawk.err.<timestamp> says.

1. Pin the corner first: a number without its corner is meaningless

A dynamic IR drop result is decided by five things, and a report must be able to answer all five:

Key points:

• RedHawk extracts the PG network’s R from the tech file itself (SPEF only covers signal nets), so the tech corner decides the drop almost linearly. Measured: same waveform, same window, only rcmax swapped for rcmin — worst drop differed by 6-7x. Tech is the number-one variable, bigger than window or VCD choice.
• VCD_FILE { SELECT_RANGE -1 -1 SELECT_TYPE WORST_POWER_CYCLE } makes the tool slide a window over the whole waveform and take the worst one; the selected window is written to the log (Time span selected (A ~ B)), and the grid step defaults to the period of the highest frequency in the STA timing file.
• Pin the window with START_TIME {ps} before comparing variants, otherwise each variant is assigned a different window and 30-40% of the difference is simply the window.
• setup package/wirebond/pad -r 0 -c 0 -l 0 = ideal off-chip, i.e. no Ldi/dt, so the drop you get is only the on-die resistive part and is a lower bound.

2. The verdict line: the Worst Dynamic Voltage Drop table

This block in the log is what you quote. Line by line:

Two checks you must do yourself:

1. Self-consistency: WIRE(VDD) + WIRE(VSS) ≈ minTW. Measured once: 171.30 + 164.50 = 335.80 mV against minTW 335.68 mV — the worst point is the same location where both sides add up.
2. Do not treat maxTW as the worst value: in the same report avgTW 297.02 mV and minTW 335.68 mV, while maxTW 107.03 mV — different definitions, and maxTW is clearly smaller. Quote only minTW/minWC.

Definition of the drop percentage (state it when reporting to a project or a customer, otherwise the number differs by a factor of two):

IR_drop% = (Videal − min_pg_tw) / Videal
         = (1.800 − 1.4643) / 1.800 = 18.65%

• VDD side only: 171.30 / 1800 = 9.52%
• VDD + VSS (the voltage the device actually loses): 335.68 / 1800 = 18.65%
• The eff_Vdd column is not the worst value; it is a window-average definition

A common sign-off habit is dynamic IR drop ≤ 10% (digital core, depending on process / voltage / product), so 18.65% has to be judged together with min_pg_tw and the standard cells’ minimum operating voltage, not as a percentage alone. Also state that this is the lower bound of an off-chip-ideal run.

3. Dynamic/{design}.dvd: per-instance data, down to the point

Column order (line 1 is a comment):

# loc_x loc_y eff_Vdd max_pg_tw min_pg_tw min_pg_sim max_pg_sim min_vdd_tw max_vss_tw domain inst_name
xxx.xx yyyy.yy 1.5466 1.7991 1.4643 1.4643 1.7994 1.6289 0.1645 VDD C_{n}_IBUF

Practical commands (rank the worst, look at the distribution):

# the 10 worst instances
grep -v '^#' {design}.dvd | awk '$5!="n/a"' | sort -k5,5g | head -10

# distribution: how many instances exceed 10%/15%/20%
# (says more about "one point" vs "the whole block" than a single maximum)
grep -v '^#' {design}.dvd | awk '$5!="n/a"{d=(1.8-$5)/1.8*100; if(d>10)a++; if(d>15)b++; if(d>20)c++} END{print a,b,c}'

Rules of thumb:

• Worst top-10 all on the same coordinate row / the same instance type (e.g. a row of IBUF) ⇒ local supply bottleneck; local strap/via work is enough.
• Areal distribution, top-10 scattered ⇒ a genuine global PG mesh problem.
• n/a rows = the instance has no timing window (not covered by STA); count them separately; in this example 14125 had data and 689 were n/a.
• Same instance, same coordinate across runs is the most reliable “same-point” comparison (especially after a layout swap, when coordinates drift: pair by instance name plus nearest coordinate).

4. Dynamic/layer_drop.rpt: local bottleneck or global mesh

Format: #Layer Min_drop Max_drop, in three blocks: #TOP / Net VDD / Net VSS. One measured example (maximum drop per layer):

How to read it:

• Worst drop concentrated on M1 (bottom layer) ⇒ segments/vias too close to the devices are insufficient; local. The layers above add only ~70 mV, so the backbone mesh is fine.
• Near-uniform per layer with large values on the upper layers ⇒ global mesh / power-planning problem (number and width of straps, pad placement).
• layer_drop and the .dvd columns min_vdd_tw / max_vss_tw cross-check each other: at the same location the numbers should agree.

5. power_summary.rpt: first prove the run itself is sane

• Recommended dynamic simulation time, {T} psec, to include {P}% of total power: follows the STA clock (1/highest frequency), independent of the VCD. If the window is much shorter, power statistics are incomplete.
• Power split by frequency domain / clock root / Vdd domain / cell type: use it to check (a) whether the clock-network share is implausible (in this example clock network 11.85 mW / 39.75 mW = 29.8%), (b) whether a large block of instances lands in NA* (not connected to power, or not assigned to a clock root), (c) whether #inst_count matches expectations.
• Order-of-magnitude self-check: Iavg = P_total / V. In this example 39.75 mW / 1.8 V ≈ 22 mA, consistent with the voltage distribution seen in the dynamic run — that is how you know the power really got in.

6. When to look at the other files

7. Three things that must ship with any number

1. Definition: VDD side only, or VDD+VSS; minTW (worst) or avgTW (average); _tw or _sim.
2. Corner and window: temperature + tech + spef + VCD + window (start-end in ns) + time step.
3. Whether off-chip is ideal: r/c/l = 0 means no Ldi/dt, so the conclusion must read “on-die resistive lower bound”.

8. Common traps (each one paid for with real data)

• DEF/SPEF/GSR was swapped but the mtime was missed, and the number is compared with last week’s ⇒ the difference actually comes from the layout. A post-ECO comparison must re-run the same corner set and check whether the worst-point instance coordinate still holds.
• Using eff_Vdd as the worst value ⇒ reported too low; using avgTW as the worst value ⇒ also too low.
• Looking only at the .dvd maximum and not at the distribution ⇒ cannot tell “one point is bad” from “the whole block is short”.
• Reporting an off-chip-ideal run with an “including package” definition ⇒ too optimistic; using it as a measured value is equally wrong.
• The frequency in the VCD disagrees with what the STA deck declares (common with behaviour-model clocks): do not bend the SDC to match the VCD — under VCD-driven analysis, changing the SDC frequency has almost no effect on IR drop. State which mode/rate the waveform came from, and add a worst/full-speed scenario.

9. Glossary

9.1 Physics and concepts

9.2 Report fields and jargon

Sources

• RedHawk 19.0.3 User Manual (Ansys/Apache), dynamic voltage drop chapter and .dvd report format
• Measured: one RedHawk 19.0.3 dynamic IR drop run on a 1.8 V, 1P3M digital core design (about 19k instances), 125 °C, rcmax tech, max SPEF, min VCD, 30 ns window, 20 ps step
• Sensitivity conclusions come from single-variable comparison experiments at the same corner (tech corner / spef corner / VCD / window / SDC frequency)