HomePhysical DesignDay 12 — DFM

Design for Manufacturing (DFM)

Yield optimization, CMP metal fill, redundant via insertion, OPC, SRAF, lithography-friendly routing, double patterning, and production DFM strategies for advanced VLSI nodes.

By EcrioniX Engineering Team · Published June 14, 2026 · ~4,700 words · 14 min read

1. Why DFM Matters — The Yield Equation

Yield = fraction of chips per wafer that work correctly. At advanced nodes, yield is never 100% — random defects from particles, lithography imperfections, and process variation kill some percentage of every wafer. DFM techniques push yield higher, directly translating to revenue.

Yield Economics — Why 1% Yield Improvement Matters
65%
Typical yield
first silicon (5nm)
+1%
Yield improvement
= ~$2M/month
$500
Cost per wafer
(5nm TSMC)
Yield Model (Seeds-of-Defect): Y = e^(-D₀ × A) D₀ = defect density (defects/cm²) — foundry process parameter A = die area (cm²) Example: 5nm TSMC, D₀ = 0.09/cm², die = 100mm² = 1cm² Y = e^(-0.09 × 1) = e^(-0.09) = 91.4% DFM improves effective D₀ by reducing sensitivity to defects: Without DFM: D₀_eff = 0.15 → Y = 86% With DFM: D₀_eff = 0.09 → Y = 91.4% Improvement: +5.4 percentage points = ~$10M/month revenue gain

2. CMP Metal Fill

Chemical Mechanical Planarization (CMP) polishes wafer surfaces flat. Without uniform metal density, high-density areas polish slower (hills) and low-density areas polish faster (dishing). Both cause focus problems in subsequent lithography steps, increasing defects.

CMP Planarization — Density Effect
Without Metal Fill (Bad) Uneven surface → focus errors Density varies 10–90% → hills & dishing With Metal Fill (Good) Flat surface → sharp lithography Density 60–80% everywhere ✓ Functional metal Dummy fill (floating)
Metal Fill Requirements (TSMC N5): Target density window: 60–80% per 10×10µm tile Min density: 60% (prevent dishing) Max density: 80% (prevent hills, meet spacing rules) Fill insertion strategy: 1. Measure density per tile after routing 2. Identify tiles below 60% → insert fill shapes 3. Identify tiles above 80% → slot existing metal or skip fill 4. Fill shapes: minimum size = 2× minimum width rule 5. Fill-to-signal spacing: maintain minimum spacing rule Fill power impact: Dummy fill adds capacitance to nearby signal nets Typical capacitance increase: 2–5% per layer Fill must be excluded from timing-critical net neighborhoods EDA tools automatically set exclusion zones around critical wires

3. Redundant Via Insertion

A single via has a small but nonzero failure probability (~0.01% per million vias at 5nm). With billions of vias on a modern chip, that means thousands of potential failure points. Redundant vias (2×1 or 2×2 arrays) drastically reduce this risk.

Single Via vs Redundant Via
Single Via (Risky)
M8 Metal VIA M7 Metal Failure rate: 0.01% per million
Redundant Via 2×1 (Safe)
M8 Metal V V M7 Metal Failure rate: ~0.000001%
Both vias must fail simultaneously for the net to open. Independent failure probability: 0.01% × 0.01% = 0.000001% — 10,000× improvement.
Redundant Via Insertion Strategy: Priority order: 1. Power/ground vias: ALWAYS use 2×2 arrays (4 vias) 2. Clock nets: ALWAYS use 2×1 arrays (2 vias) 3. Timing-critical signals: use 2×1 where space allows 4. Non-critical signals: best-effort (fill where spacing permits) Coverage targets (industry standard): Power/ground: 100% redundant Clock: 95%+ redundant Critical signal: 80%+ redundant Total average: 60–70% redundant Area overhead: ~5–10% routing area increase Worth it: a single via open failure can kill the entire chip!

4. Optical Proximity Correction (OPC)

At sub-wavelength lithography (features smaller than the light wavelength), diffraction causes printed shapes to differ from the mask shapes. OPC pre-distorts the mask so that the wafer-printed result matches the intended design.

OPC — Mask Correction for Diffraction
Intended Shape
Target
Perfect rectangle in design
Without OPC (printed)
Rounded
Corners rounded, line ends pull back
With OPC (mask)
OPC mask
Serifs added → prints like target

5. Sub-Resolution Assist Features (SRAF)

SRAFs are tiny mask features smaller than the resolution limit — they don't print on the wafer, but they improve the aerial image contrast of the main features around them, reducing line-width variation.

TechniquePurposeWhen AppliedNode
OPC (Optical Proximity Correction)Pre-distort mask to compensate diffractionAll features, especially corners180nm+
SRAF / Scattering BarsImprove contrast of isolated linesIsolated lines and spaces90nm+
Source Mask Optimization (SMO)Co-optimize illumination and mask shapeCritical layers28nm+
Inverse Lithography Technology (ILT)Compute optimal mask from target patternFull-chip critical layers7nm+
EUV + OPC13.5nm wavelength enables direct patterningCritical layers at 5nm and below5nm–1nm

6. Double Patterning (SADP/LELE)

At 14nm and below, a single exposure cannot pattern features at minimum pitch. Double patterning splits one layer into two exposures:

LELE Double Patterning (Litho-Etch-Litho-Etch)
Mask 1
Odd lines
+
Mask 2
Even lines
=
Result
Half-pitch!
Key DFM constraint: Lines from different masks (different colors) cannot connect on the same layer. EDA tools perform coloring — assigning each wire to mask 1 or mask 2 — and flag coloring conflicts where topology makes this impossible.

7. Lithography-Friendly Design Rules

8. DFM Tools

ToolVendorDFM Function
Calibre YieldAnalyzerSiemensYield prediction, critical area analysis
IC Validator DFMSynopsysDFM rule checks integrated with ICC2
Voltus-FiCadenceFill insertion with timing awareness
Mentor Calibre OPCSiemensOPC generation and verification
Synopsys ProteusSynopsysILT and SMO mask optimization

DFM Production Checklist

Next — Day 13: EDA Tools deep-dive — Cadence Innovus and Synopsys ICC2 workflow, Tcl scripting, flow automation, and tool comparison.

← Previous
Day 11: Engineering Change Orders
Next →
Day 13: EDA Tools