Sign-off: IR Drop, EM, Crosstalk & Antenna
Before tape-out a chip must pass every sign-off check. Learn static and dynamic IR drop, electromigration (EM) and Black's equation, crosstalk delay and glitch, antenna effect, multi-corner multi-mode (MCMM) STA, ECOs, the final checklist (STA, LEC, DRC, LVS, IR/EM, SI) and GDSII tape-out — with an IR-drop calculator for a power rail.
A routed design can still fail in silicon: supply voltage may sag where many cells switch (IR drop), wires may wear out from current (electromigration), neighbouring wires may disturb each other (crosstalk), and plasma during fabrication can damage gates (antenna).
Sign-off analysis checks all of these across every process, voltage and temperature corner and every operating mode. Only when all checks are clean does the design go to the foundry as GDSII/OASIS — the tape-out.
An analogy
Sign-off is the final safety inspection of a building: the electrical wiring can carry the load (EM), the voltage holds up when everything is switched on (IR), neighbours don't interfere (crosstalk), and construction didn't damage anything (antenna).
IR drop
Current through the resistance of the power grid lowers the voltage reaching cells, making them slower (and ground bounce raises VSS). Static IR uses average currents; dynamic IR simulates simultaneous switching and is usually much worse locally. Fixes: more straps and vias, decoupling capacitors, spreading high-activity cells, and staggering clock-gating enables.
Electromigration
High current density pushes metal atoms along a wire, forming voids (opens) and hillocks (shorts) over years. Black's equation shows lifetime drops steeply with current density and temperature. Power nets see DC EM; signal and clock nets see AC (RMS/peak) limits. Fixes: wider wires, more vias, lower drive strength, NDRs.
Crosstalk and antenna
Two classic signal-integrity and manufacturing problems:
- Crosstalk delay: an aggressor switching opposite to the victim increases its effective capacitance (slower); same direction speeds it up
- Crosstalk glitch: a quiet victim gets a noise bump that can flip a latch or dynamic node
- Fixes: spacing, shielding, buffering the victim, downsizing aggressors
- Antenna: long metal connected only to a gate collects charge during plasma etch and can damage the thin oxide; fixes are layer jumping (bridge up to a higher metal) or antenna diodes
MCMM sign-off and tape-out
Timing is signed off in many scenarios: corners (SS/FF/TT × voltages × temperatures × RC extraction corners) and modes (functional, test, low-power). ECOs fix late violations with minimal changes (often metal-only, reusing spare cells). The final checklist: STA all scenarios, LEC RTL↔netlist, DRC/LVS/antenna/density clean, IR/EM clean, ESD/latch-up checks, then GDSII/OASIS stream-out and the foundry's tape-out review.
Formulas
- IR drop:
ΔV = I · R_grid (budget typically ≤ 5–10% of V_DD) - Black's equation (EM):
MTTF = A · J⁻ⁿ · e^(E_a / kT) (n ≈ 1–2) - Crosstalk glitch (charge sharing):
ΔV ≈ V_DD · C_c / (C_c + C_g) - Miller effect on delay:
C_eff = C_g + k·C_c, k = 0 (same dir) … 2 (opposite) - Antenna ratio:
AR = A_metal connected to gate / A_gate ≤ rule limit
Worked example: Static IR
Given: Grid resistance 0.05 Ω to a block drawing 2 A
Find: IR drop.
- 0.05 × 2
Answer: 100 mV (10% of 1.0 V — at the limit)
Worked example: Crosstalk glitch
Given: C_c = 20 fF, C_g = 30 fF
Find: Glitch on a floating victim.
- 20/(20 + 30)
Answer: 0.4·V_DD — dangerous; a driven victim sees much less
Worked example: EM lifetime
Given: Black's n = 2, current density doubled
Find: Lifetime change.
- MTTF ∝ J⁻²
- 2⁻²
Answer: Lifetime drops to 1/4
IR drop along a power rail
# IR drop along a power rail fed from one end: N cells, each drawing I,
# rail segment resistance R between cells.
VDD, N, I, R = 1.0, 10, 5e-3, 0.5 # V, cells, A per cell, ohm per segment
v = VDD
worst = VDD
for k in range(N):
current_in_segment = (N - k) * I # current still flowing downstream
v -= current_in_segment * R
worst = min(worst, v)
drop = VDD - worst
print(f"worst-case rail voltage {worst*1000:.1f} mV, IR drop {drop*1000:.1f} mV ({drop/VDD*100:.1f}% of VDD)")
# Feeding the rail from both ends roughly quarters the worst drop
half = N // 2
v2 = VDD - sum((half - k) * I * R for k in range(half))
print(f"fed from both ends: IR drop {(VDD - v2)*1000:.1f} mV")
# worst-case rail voltage 862.5 mV, IR drop 137.5 mV (13.7% of VDD)
# fed from both ends: IR drop 37.5 mVThe first segment carries all the current, so the drop accumulates toward the far end. Feeding from both ends (more straps) cuts it from 137.5 to 37.5 mV — exactly why power grids are meshes with frequent straps.
Common mistakes
- Wrong: Checking only static IR. — Right: Run vector-based or vectorless dynamic IR as well.
- Wrong: Signing off at a single corner. — Right: Cover all PVT + RC corners and modes (MCMM).
- Wrong: Ignoring timing impact of IR drop. — Right: Use IR-aware STA or margin for voltage droop.
Applications
- Automotive chips: 15-year EM lifetime at 150 °C junction temperature.
- AI accelerators: Hundreds of amps: dynamic IR is the top risk.
- Mobile SoCs: Dozens of MCMM scenarios for DVFS modes.
- Open-source tape-outs: OpenLane runs DRC/LVS/antenna before shuttle submission.
Key takeaways
- IR drop = I·R; mesh the grid and add decaps.
- EM lifetime ∝ J⁻ⁿ e^(Ea/kT).
- Crosstalk changes delay and causes glitches.
- Tape-out needs clean STA, LEC, DRC/LVS, IR/EM, SI.
Going further
- Self-heating: FinFET/GAA self-heating accelerates EM on signal wires.
- Aging (BTI/HCI): Transistors slow over life; sign-off includes aged libraries.
- Chip-package co-analysis: Package inductance adds L·di/dt droop.