What Is This Tool?
This calculator estimates how many good, working chips (dies) come off a silicon wafer after manufacturing. It models the real-world relationship between wafer size, die dimensions, defect density, and yield — the key economics of semiconductor manufacturing.
💡 Tip: Start with the Calculator tab. Enter your parameters and hit Calculate. Use this Guide tab to understand what each value means.
How to Use
- Choose your Wafer Diameter — larger wafers hold more dies and lower per-die cost.
- Set your Die Width & Height in millimeters — your chip's physical footprint.
- Set Scribe Lane Width — the gap cut between dies (typically 60–100 µm).
- Enter D₀ (Defect Density) — how many random defects per cm² your fab produces.
- Enter α (Clustering Factor) — how clustered defects are (2–5 is typical).
- Pick a Yield Model — Negative Binomial is the industry standard.
- Click Calculate — results show instantly.
Understanding the Inputs
Wafer Size
Silicon wafers come in standard diameters. Larger wafers have more usable area and lower per-die manufacturing cost because fixed costs (lithography, cleaning, inspection) are spread over more dies. The semiconductor industry transitioned from 8" (200 mm) to 12" (300 mm) wafers in the 2000s. 18" (450 mm) wafers exist but are not yet in mass production.
Die Size (Width × Height)
This is the physical size of one chip on the wafer. Bigger dies cost more because:
- Fewer fit on a wafer → fewer units per wafer
- Larger area = higher chance of hitting a defect → lower yield
GPU dies are often 400–800 mm². Mobile SoCs are typically 60–120 mm².
Scribe Lane Width
The narrow street between dies where a diamond saw cuts them apart. This space is "wasted" area — not a die, not useful. Reducing it reclaims area for more dies, but requires precision cutting. Typical values: 60–100 µm.
Process Node (nm)
The minimum feature size of the transistors. Smaller nodes pack more transistors per mm² but have higher D₀ initially (process immaturity), higher mask costs, and require advanced lithography (EUV below 7nm). This field is informational — used in the formula display for context.
D₀ — Defect Density
The number of random, fatal defects per cm² the fab produces. This is the single most important fab quality metric. Mature processes have D₀ around 0.01–0.05 def/cm². New processes can start at 0.5–2.0+ def/cm² until they stabilize.
| D₀ Range | Interpretation |
|---|
< 0.05 | Excellent — mature, high-volume process |
0.05 – 0.2 | Good — established process |
0.2 – 0.5 | Fair — ramping or aging process |
> 0.5 | Poor — early development or process issues |
α — Clustering Factor
Real-world defects don't spread randomly — they cluster around contamination events, equipment faults, or edge effects. The α parameter in the Negative Binomial model captures this:
- Low α (0.5–1): defects are highly clustered — one bad spot kills many dies together but leaves others untouched → better yield than Poisson predicts
- High α (10–20): defects are nearly random → approaches Poisson model
- Typical: 2–5 for most production processes
Yield Models Explained
Three models exist for estimating yield. Each makes a different assumption about how defects are distributed across a wafer.
Industry standard
Negative Binomial
D₀defect density (def/cm²)
Adie area (cm²)
αclustering factor (> 0)
Accounts for defect clustering. α → ∞ converges to Poisson.
Empirical classic
Murphy's Model
D₀defect density (def/cm²)
Adie area (cm²)
Triangular D₀ distribution. No clustering param — useful when α is unknown.
Worst-case bound
Poisson Model
Y
=
e−D₀·A
D₀defect density (def/cm²)
Adie area (cm²)
Fully random defects assumed. Pessimistic — gives a lower-bound on yield.
📐 Deep Dive: The Negative Binomial Model
The NB model is the industry standard because it matches what engineers actually observe on real wafers. Its core insight: defects don't scatter randomly — they cluster.
Why Poisson falls short
Poisson assumes each wafer point has an equal, independent defect probability. That's statistically clean but physically wrong. Defects come from discrete events — a contaminated bath, a worn machine part, an ESD spike — each damaging a localized region. One event can kill 5–20 adjacent dies while the rest of the wafer stays clean. Poisson can't model this, so it underestimates yield.
The statistical foundation
NB is best understood as a Poisson-Gamma mixture: D₀ isn't a fixed constant — it varies region-to-region following a Gamma distribution. Some zones have very high local D₀ (contaminated), most have near-zero. The α parameter controls how wide that Gamma spread is. Integrating over all possible local densities yields the Negative Binomial distribution naturally.
Low α (0.5–1) — highly clustered
Defects bunch together in a few hot zones. Most dies are untouched → yield higher than Poisson predicts
High α (10+) — nearly random
Defects spread evenly. Every die faces roughly equal risk → converges to Poisson
Understanding α — the clustering factor
| α range | Pattern | Typical root cause | vs. Poisson |
|---|
0.5–1 | Highly clustered | Particle bursts, CMP slurry events, ESD | Significantly higher |
1–3 | Moderately clustered | Mixed contamination, edge effects | Noticeably higher |
3–6 | Mildly clustered | Stable equipment, mature process | Moderately higher |
10–20 | Nearly random | Diffusion-limited defects, cosmic rays | ≈ Same |
∞ | Fully random | Theoretical limit | Identical to Poisson |
As α → ∞, the NB formula converges to Poisson. Mathematically: lim(α→∞) [1/(1+x/α)]^α = e^(−x) where x = D₀ × A.
Live model comparison
Adjust D₀ and α to see how the three models diverge as die size grows.
Negative Binomial
Murphy's
Poisson
Worked example
D₀ = 0.1 def/cm², die = 15 × 15 mm = 2.25 cm², α = 3:
Neg. Binomial
1 / (1 + 0.1 × 2.25 / 3)³
→
1 / (1.075)³
→
75.9%
Murphy's
[(1−e^−0.225) / 0.225]²
→
78.1%
Poisson
e^(−0.1 × 2.25)
→
e^(−0.225)
→
79.9%
⚠️ NB isn't always the highest. At α = 3 above, NB actually gives the lowest yield. The clustering advantage is strongest when α < 1 and D₀ × A is large — meaning many defects hitting a small fraction of the wafer area. Try α = 0.5 in the chart above to see this flip.
How to extract α in practice
- Collect yield data (Y) across multiple lots, with known D₀ and die area A per lot.
- Rearrange the NB formula or use nonlinear least-squares fitting to solve for α per lot.
- Average across lots to characterize the process.
New nodes often show high α (random-looking defects) during development. As the process matures and contamination sources are isolated, α drops — paradoxically meaning defects become more clustered but fewer, which improves yield.
NB vs. Murphy's: which to use?
- NB has a physically motivated basis (Poisson-Gamma mixture); Murphy's is purely empirical
- α is tunable and extractable; Murphy's has no equivalent knob
- NB fits modern fabs better, especially at D₀ < 0.05 def/cm²
- NB handles extreme clustering (α < 1); Murphy's cannot
Murphy's is still a fast sanity check when α is unknown, sitting between Poisson (pessimistic) and NB (realistic).
Reading the Results
- Die Area — total area of one die including scribe lanes
- DPW (Dies Per Wafer) — gross count of dies that physically fit, before yield
- Yield % — probability a random die is functional
- GPPW — Good Parts Per Wafer = DPW × Yield — actual sellable chips
📈 Cost insight: Die cost ≈ (Wafer cost) / GPPW + Test cost. Doubling yield halves die cost — yield improvement is often more valuable than shrinking to a smaller node.
Typical Values Reference
| Product | Die Size | Node | Wafer | Est. Yield |
|---|
| Mobile SoC | ~80 mm² | 4–5 nm | 300 mm | 70–85% |
| High-end GPU | 500–800 mm² | 4–5 nm | 300 mm | 40–65% |
| MCU | 5–20 mm² | 28–40 nm | 200 mm | 90–98% |
| Server CPU | 200–400 mm² | 4–7 nm | 300 mm | 60–80% |