Test and Yield Leaning

In addition to the normal sorting function, test provides the essential feedback loop for yield-learning. Specifically, product-based diagnostics are increasingly vital. The need to base diagnostics on actual product hardware is driven by systematic defect mechanisms now being increasingly complex functions of neighboring shapes, local pattern densities, etc. As a result many failure mechanisms may be visible only on product. In addition, product-based diagnostics automatically put focus on key yield-limiting failure mechanisms. Volume-based diagnostics are critical since individual occurrence of any given systematic defect mechanism may be rare. The pooling of data across many failing die are needed to identify true systematic mechanisms.

== 1.1     Electrical Test-Based Diagnosis ==

Parametric-related feedback is needed for (1) device and interconnects parameters and (2) design-process interactions. Measurement of device and interconnect parameters have traditionally relied upon test structures, especially scribe line FETs and interconnect resistance and capacitance monitors. Increasing across-chip variation (intra-die variability) enhances the negative impact of scribe-line-to-chip offsets. Moreover, test structures are limited by the number of physical and electrical configurations they are able to cover. As circuit parametrics are increasingly affected by such configurations, including within-standard-cell and transistor-layout configurations, it becomes necessary to base learning on product test or test of product-like layout configurations. Embedded, distributed monitor circuits such as thermal and VDD sensors, process-monitoring ring oscillators and critical path proxies are now standard on microprocessor-class ICs and can be used to help diagnose parametric fails and understand variability. Understanding variability includes unraveling the structure of variations into spatial and cross-parameter components (variation in transistor length, Vt, source-drain resistance, etc.)  The spatial component includes both die-to-die and within-die components.

Cross-parameter variations, potentially including a spatial component, are important to analog/RF circuits, as well as digital. Methods for understanding/characterizing the manufacturing process and operating environment that are sensitive enough for analog/RF are needed. Moreover, product test is uniquely well-suited to provide feedback on design-process interactions, including those leading to noise-related fails, such as power-grid-droop and crosstalk fails.

Top challenges for test-based yield-learning include:

·       Better resolution for cell-internal defects. Latest advances in structural testing and scan-based logic/layout-aware diagnosis methods are adequately addressing interconnect and via defects. Statistical approaches built into volume-based diagnostics are able to predict interconnect-related defect modes without an over-dependence on Physical Failure Analysis (PFA). Innovation is required, however, for cell-internal-defect-targeted diagnostics to be able to identify systematic fail modes inside standard cells. Observations of hardware suggest a shift toward a larger percentage of the defect distribution being cell-internal defects, as opposed to interconnect-related. Current best methods for cell-internal defect diagnostics are cell truth-table and gate-exhaustive model-based, with the truth tables established via static SPICE simulations of modeled cell-internal parasitics. These methods suffer from aliasing issues and over-reliance on potentially inaccurate modeling of cell-internal defects used in SPICE simulations. In addition, diagnosis resolution needs to be better due to limitations in the PFA process.

·       Managing design data for yield learning. A tremendous amount of design data can be brought to bear for yield learning purposes, but it is often not organized effectively for this purpose. In addition, with hierarchical design and DFT flows, the overall management of this data at most companies today is ad-hoc and limiting its effective use.

·       Inadequacy of LEF/DEF as the basis of layout-aware diagnosis. LEF/DEF suffices for the purposes of layout-aware ATPG but is too early in the design cycle to be used effectively for layout-aware diagnosis. LEF/DEF is getting farther and farther away from final mask shapes due to complex OPC, boolean and retargeting steps.

·       Yield-Learning in a OSAT/Fabless/Foundry environment. Yield-learning capabilities must be cognizant of the environment that has become the dominant model for our industry. If the technology cannot deal with the security and logistical concerns of this environment, it cannot be effective. Factory integration issues must be

addressed. Data capture and management capabilities must support increasing reliance on statistical analysis and data mining of volume production data for yield-learning. Secure mechanisms for yield-data flow for distributed design, manufacture and test, including fabless/foundry and 3rd party IP, are needed. Standard test data formats, such as STDF-V4-2007 scan fail data, and infrastructure to support their transmittal, are needed to support automation and sharing of data. Distributed design, manufacture and test also creates an emerging role for methodologies and tools to help determine in which area problems lie, e.g., design house, foundry or test house.

·       Test and data-collection time increases due to longer scan chains. These increases drive a need for focus on LBIST methodologies and scan compression for both test and diagnosis.