Investigators
John Fowler, Gerald Mackulak (Arizona State Univ.)
Lee Schruben (Univ. California Berkeley)
Industrial Liaisons
Gurshaman S. Baweja, Texas Instruments Incorporated
Ed Cervantes, Advanced Micro Devices, Inc.
Amit Gupta, Texas Instruments Incorporated
Sarah J. Hood, IBM Corporation
Mani Janakiram, Intel Corporation
Matilda O'Connor, Advanced Micro Devices, Inc.
Chii-Liang Wee, National Semiconductor Corporation
David A. Wizelman, Advanced Micro Devices, Inc.
Edward J. Yellig, Intel Corporation
Task 1
Anticipated Primary Result : A resource-driven (R-D) simulation technique to enable faster execution of semiconductor fab models compared to job-driven (J-D) models. Task 1: 2002 Annual Report
Background : A typical semiconductor fab simulation model can be used to predict the impact of system changes such as different queue disciplines and job priorities, etc. A major problem with the use of simulation is the amount of time required to build and execute appropriate models for these predictions. We address the issue of model execution and build time by further developing the R-D technique advocated by Schruben. The "resource-driven" technique is based on modeling the cycles of resource entities used to process jobs within the factory rather than the more common approach of modeling the flow of job entities through the factory. The major advantage of this paradigm is that the model footprint is a function of the numbers of resources, not numbers of jobs - execution does not slow appreciably as the system becomes more congested. As currently developed, model execution speed is sensitive to both job mix and routing complexity, but we have promising ideas for addressing these issues. While some loss of information is anticipated using R-D logic, we expect that decisions based on the simulation study will not change - or that R-D models can be enriched to include job flow information to recover needed information.
Description : This research will explore an alternative fab simulation methodology that focuses on resource cycles. In a resource-driven (R-D) simulation, individual jobs are passive and are "moved" or "processed" by active system resources such as tools, operators, and AMHS systems. Rather than maintaining a record of every job in the system, only integer counts of the numbers of jobs of particular types at different steps are maintained. The systems state is described by the status of resources (also integers) and these job counts. The speed and space complexity of resource-driven simulations is O(1): execution speed and memory footprint does not change significantly as the system becomes more congested. The events in a resource-driven simulation involve simple elementary integer operations, typically incrementing or decrementing job counts and numbers of available resources. Very large and highly congested queuing networks can be modeled this way with a relatively small, finite set of integers. While many of the traditional performance measures are available, it is expected that resource-driven simulations have inherently lower resolution. Our team has created prototype resource-driven simulations of highly congested systems that execute orders of magnitude faster than corresponding job-driven process flow simulations of the same factory. We propose to build upon this prior work. In this research we will further refine define the concept of R-D simulation modeling with particular emphasis on the limitations of this paradigm. Investigations will include the development of additional experimental models, the definition of the R-D approach as implemented in appropriate software, and a report on the relationship of complexity to execution speed. Furthermore, we will investigate the limitations of the R-D approach in terms of its ability to produce specific output measures. Finally, we will develop a method for incorporating "mixed" modeling details when and if necessary.
Task 2
Anticipated Primary Result : A methodology and proof of concept software tool for converting selected types of models (Test-bed, ASAP) from the J-D paradigm to the R-D paradigm and/or to a mixed (combination of R-D and J-D) paradigm allowing the increased execution speed envisioned for the R-D approach. Task 2: 2002 Annual Report
Background : Constraints exist in the simulation model development process. A useful model requires sufficient detail and accuracy in both the physical and statistical domains. Semiconductor fab modeling presents the additional constraint of short product life cycles resulting in short analysis time cycles and a strong reliance on fast model execution if an analysis is to be useful. Woodward (Mackulaks Ph.D. student now at INTEL) developed a methodology for reverse engineering a graphical model from simulation code so as to allow for automated testing of logic errors, determination of dynamic performance, detection of deadlocks, and model behavioral equivalency. The basic idea is that automatically converting models to an intermediate form allows for algorithmic evaluation of the properties related to output generation. The same concept should be applicable to the conversion of a specific class of already existing models from their J-D form to the R-D form. This conversion facilitates the use of the faster R-D models without requiring additional data collection or modeling effort.
Description : The approach developed by Woodward found that for a certain class of models behavioral equivalence could be ascertained by converting model code (and data) to a form suitable for algorithmic manipulation. The algorithm was then able to determine equivalence of both static and dynamic system components between a set of selected models. Extending this approach means that an existing model written in ASAP could conceivably be automatically converted to the R-D approach using similar conversion concepts. The conversion must of course be conducted on the basis of a selected set of performance metrics and details specific to the R-D modeling paradigm. This task is concerned with identifying the factors necessary for behavioral equivalence and then developing software to enable conversion. The first step to successfully accomplishing this task is the investigation of Woodwards efforts in this area as applied to the approach we are calling "resource-driven" (R-D) modeling. Ideally we will be able to extend prior work so as to be useful for generating the R-D modeling paradigm directly from a selected subset of features of the "job-driven" (J-D) approach. These features may include data, resource specifications, queuing logic, process flow, yields, maintenance, etc. and have yet to be definitively determined. The automated generation of the R-D models will allow comparison of output metrics, dynamic execution behavior, and speed of execution between paradigms. This task will develop the methodology for the conversion process as a function of the final requirements of the R-D simulation paradigm. It will illustrate the value of the conversion by converting existing models from the Testbed and from AutoSched AP, and benchmarking the performance of these conversions against the newly constructed R-D models. The benchmark comparisons will include but not necessarily be limited to output metrics and execution speed.
Task 3
Anticipated Primary Result: Goal-driven optimization methodologies for a variety of semiconductor manufacturing problems (such as capacity planning) that use a "best" combination of R-D, J-D, and mixed models. Task 3: 2002 Annual Report
Background : Modeling semiconductor manufacturing systems is a non-trivial task. Issues related to appropriate model detail, sufficient run length for accurate output parameter estimation, and short project life cycles all influence the analysts ability to produce results in a timely fashion. Tasks one and two propose methods to reduce run time and increase model usefulness through development of a new model building paradigm (Task 1) and the automatic generation of R-D models from ASAP models (Task 2). This task is focused on the development of a strategy for determining when and how to apply the appropriate modeling paradigms to optimally solve semiconductor manufacturing problems that require discrete event simulation.
Description : The initial phase of this task will be the determination of the metrics that permit comparison of the R-D and J-D modeling paradigms. The structure that allows for the evaluation of these two approaches is required before a comparison can be made. For example, is it sufficient to look at the desired output metrics and compare on the basis of statistical confidence intervals or is the comparison of physical system components sufficient? Also, do comparisons depend on the purpose of the study (scheduling, AMHS, batching decisions) or can they be made solely on the output generated? This phase of the task will work closely with tasks 1 & 2. Once a comparison approach has been defined analysis will be undertaken to determine whether the R-D approach (fast execution) performs best as a predictor for the J-D approach (comprehensive detail) under certain sets of physical and/or logical model structures. The idea is that we should be able to develop a plan for where and how to use R-D and where to ultimately switch to J-D models so as to get the required solutions with minimal modeling effort. Optimization methods will be applied to various classes of semiconductor manufacturing problems. The idea behind this investigation is that appropriate application of the R-D models will significantly reduce the overall modeling time and effort of large scale analysis problems. If the R-D models can be used to quickly reduce the solution space we will be able to evaluate more complicated objectives. Finally, the approach developed will be compared to other existing approaches in the literature.