August_EDFA_Digital
edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 20 NO. 3 4 EDFAAO (2018) 3:4-7 1537-0755/$19.00 ©ASM International ® SEMICONDUCTOR YIELD MODELING: A PRIMER John Hopkins GlobalFoundries, Malta, N.Y. john.hopkins@globalfoundries.com INTRODUCTION A semiconductor fab typically produces a variety of semiconductor designs that exhibit awide range of electri- cal wafer final test yields depending on the process tech- nology maturity and product design complexity. At any given time, the average percentage of electrically gooddie on each finished production wafer can range from <10% to well over 90%. It is important to accurately predict the expected yield of each product to determine whether the yield is meeting expectations, and if not, to identify the characteristics of the shortfall. This article provides an overview of models used to predict electrical wafer final test yield and how those projections may be used when deciding how to direct valuable failure analysis resources. ELECTRICAL TEST YIELD MODELS Several dierent models are used in the industry to predict electrical wafer final test yield. [1,2] Although the math diers, all of the models account for process fault density and product design complexity. For the purposes of modelling electrical test yield, a fault is defined as a physical defect that causes the device not to meet the electrical test specification. Design complexity can be described in terms of either a physical or normalized die area. For example, one simplemodel is the Poissonmodel that can be written as: Y = e - ADo where A = die area and D 0 = fault density A shortcoming of this model is that for two chips of the same area, there is no distinction between a densely populated logic design (very susceptible to defects) and amore forgiving design (for example, onewith redundant SRAM). For a given fabricator fault density, this model would predict the same yield for each die. The model can be improved by breaking out the total die area into sub-areas of dierent circuit categories and applying a factor to each area to account for the diering circuit complexity. For example, A1 might be the area of the die populated with dense logic, A2 the area of the die populated by SRAM with redundancy, and A3 the area of the die populated by SRAM without redundancy, etc. The total A would be A1F1 + A2F2 + A3F3 +…where F(n) is between 0 and 1 and accounts for the diering susceptibility to failure of the dierent circuit categories. Another commonly used yieldmodel is the Bose-Einsteinmodel, which takes the form: Y = 1/(1 + AD) N where A = die area, D = fault density/layer, N = complexity factor The N factor in the Bose-Einstein model accounts for process complexity and is based on the number of critical mask levels. Figure 1 illustrates the eect of the N factor on the predicted yield for a particular fault density. Although these models can account for diering design and process com- Fig. 1 Yield vs. design complexity for dierent process complexity factors N (Bose- Einsteinmodel). Higher N corresponds to higher complexity and lower yield.
Made with FlippingBook
RkJQdWJsaXNoZXIy MjA4MTAy