August_EDFA_Digital
edfas.org 5 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 20 NO. 3 plexities, they do not account for the phenomena of defect clustering. One way to illustrate how defect clustering aects yield is to consider two cases in- volving a wafer with 100 die that is sub- ject to exactly 100 chip killing defects. In Case 1, the 100 defects are distributed evenly across the wafer such that one defect falls on each die. The yieldwill be 0% (all 100 die are killed by one of the 100 defects). In Case 2, the 100 defects are highly clustered and hit only one of the die on the wafer. Although the fault density is exactly the same in both scenarios, the Case 2 yield will be 99%. If clustering is not accounted for, it can result in an underestimation of the yield of large die. Amodel that takes into account both circuit complexity and clustering is the negative binomial model: [3,4] Y = 1/(1 + A C D/ α ) α where A C = critical area, D = defect density and α = cluster factor A C is the portion of the die area that is susceptible to defects of a certain size that cause electrical failures. The A C can be algorithmically determined using various commercial DFM tools. [5] The lower the cluster factor α , themore clus- tered the defects. The technology α for a fab can be extracted by plotting yield vs. area for various groupings of die yield and typically ranges between 1 and 10. [2] Note that as α approaches infinity, the negative binomial model approaches the Poisson model, as shown in Fig. 2. SYSTEMATIC VS. RANDOM DEFECT YIELD LOSS In the yield model equations discussed so far, it is assumed that the yield loss is driven entirely by randomly distributed defects. In reality, a portion of the yield loss can o¨en be attributed to nonrandomsystematicmecha- nisms. Examples includenon-centereddeviceparameters, misaligned lithography levels, under- or over-polished insulator layers, etc. [4] These types of mechanisms tend to kill die independent of their layout complexity. To account for this nonrandomdefect loss, a systematic yield component (Y 0 ) must be added to the model: Yield = Y 0 * Y D = Y 0 * 1/(1 + A C D/ α ) α This negative binomial model takes into account design complexity, clustering, and systematic yield loss mechanisms, as shown in Fig. 3. Being able to categorize the type of yield loss as randomor systematic is important for estimating the yield impact on other products. The random yield (Y D ) impact scales with circuit complexity, while the systematic yield (Y 0 ) impact is independent of circuit complexity. The random defect yield can be scaled as follows. A Poisson model is used for simplicity: Product 1: Y D1 = e - A 1 D ; D = -ln(Y D1 )/A 1 Product 2: Y D2 = e - A 2 D = e - ( - ln(Y D1 )/A1)A 2 = Y D1 A2/A1 Fig. 2 Yield vs. design complexity for dierent cluster factors α (negative binomial model). Lower α corresponds to a higher degree of clustering, which results in better yield for larger, more complex designs. Fig. 3 Yield vs. design complexity. Negative binomial-based model that accounts for both systematic (Y O ) and random defect (Y D ) eects.
Made with FlippingBook
RkJQdWJsaXNoZXIy MjA4MTAy