Semiconductor Yield Engineering & VLSI
Stapper Negative-Binomial & Bose-Einstein Yield Formula
Calculate Charles Stapper’s Negative-Binomial Clustered Wafer Yield Y_NB = (1 + D_0 * A / alpha)^(-alpha), Bose-Einstein Single-Layer Yield (alpha = 1), and Multi-Mask Critical-Layer Bose-Einstein Yield.
Explain why 800 mm^2 AI accelerator dies beat unclustered Poisson yield!
Why don’t reticle-limit 800 mm^2 AI GPU dies suffer the catastrophic yield collapse predicted by unclustered Poisson EXP(-D_0 * A)? Charles H. Stapper (1973, IBM Journal of Research and Development) proved that cleanroom particle scratches and etch defects cluster spatially: when defect density varies across wafer zones according to a Gamma distribution with Clustering Parameter alpha > 0, compound Gamma-Poisson integration yields the Negative-Binomial Yield Law Y_NB = (1 + (D_0 * A_cm2) / alpha)^(-alpha)! Lower alpha means tighter spatial clustering (defects pile onto a few bad dies, sparing neighboring dies!). At alpha = 1, Y_NB reduces to the Bose-Einstein / Seeds formula 1 / (1 + D_0 * A_cm2), and across N_mask independent critical lithography levels with per-mask defect density d_mask = D_0 / N_mask, Multi-Mask Bose-Einstein Yield is 1 / (1 + (D_0 * A_cm2) / N_mask)^N_mask!
Verify on Reticle_GPU_Die_Area A_cm2 = 8.00 cm^2 (800 mm^2, B2), Fab_Defect_Density D_0 = 0.15 defects/cm^2 (lambda = 1.20 faults/die, B3), Spatial_Clustering_Alpha alpha = 2.0 (typical mature 4nm/5nm fab clustering, B4), and Critical_Mask_Layers N_mask = 20 (B5): Unclustered Poisson Yield is EXP(-1.20) = 0.3012 (30.12% in E2); Stapper Negative-Binomial Yield at alpha = 2.0 jumps to (1 + 1.20 / 2.0)^(-2) = 1.60^(-2) = 0.3906 (39.06% in E3—nearly +9 percentage points higher!); Single-Level Bose-Einstein Yield (alpha = 1) is 1 / 2.20 = 0.4545 (E4); and 20-Mask Bose-Einstein Yield is (1 + 1.20 / 20)^(-20) = 0.3118 (E5)!
| VLSI Clustered Yield Model (8.0 cm^2 GPU Die, D_0=0.15/cm^2 -> lambda=1.20) | Formula | Verified Output |
|---|---|---|
| Unclustered Poisson Yield (alpha->inf) | EXP(-D_0 * A_cm2) | 0.3012 (30.12%) |
| Stapper Negative-Binomial (alpha=2) | (1 + D_0 * A_cm2 / alpha)^(-alpha) | 0.3906 (39.06% Exact) |
| Bose-Einstein Yield (alpha=1) | 1 / (1 + D_0 * A_cm2) | 0.4545 (45.45%) |
| N-Mask Bose-Einstein (N_mask=20) | (1 + D_0 * A_cm2 / N_mask)^(-N_mask) | 0.3118 (31.18%) |
Stapper Negative-Binomial Clustered Yield Formula
Validate A_cm2 > 0, D_0 >= 0, alpha > 0, and N_mask >= 1!
=LET(a_cm, B2, d_0, B3, al_c, B4, n_mk, B5, IF(OR(a_cm <= 0, d_0 < 0, al_c <= 0, n_mk < 1), "Invalid Defect Model", ROUND((1 + (d_0 * a_cm) / al_c) ^ (-al_c), 4)))Multi-Mask Bose-Einstein Yield Formula
Notice the calculus limit: as N_mask (or alpha) grows large (set alpha = 1000 in B4), (1 + lambda / 1000)^(-1000) = 0.3014 converges straight to Poisson EXP(-1.20) = 0.3012!
=IF(ISTEXT(E3), E3, ROUND((1 + (B3 * B2) / B5) ^ (-B5), 4))Try the example
Use the original values or change them to test your understanding.