 |
Web Hosting by Netfirms | Free Domain Names by Netfirms |
|
|
|
Shekar Krishnaswamy
ABSTRACT
A
significant challenge in implementing the Theory of Constraints in the
semiconductor industry is the complex and reentrant nature of the manufacturing
process. Managing a constraint or bottleneck in such processes is resultantly
complex and is an ongoing topic of research. This paper describes a
straightforward method to avoid starvation of possibly reoccurring bottleneck
equipment. This method is not proximity sensitive, meaning that it does not act
on material only if it is within a specified proximity of the constraint. This
method is capable of reacting to changes in product mix and can act to correct
line imbalances since it is based on real time data from the floor control
system. The assertion is also made that
for the case where a bottleneck has several occurrences in the product flow, all occurrences
of the bottleneck must avoid starvation, not just the tool set. This affects
the ability of the primary constraint to feed itself.
The
Problem
In today's semiconductor
industry we have a unique circumstance where the demand for semiconductor
products is seemingly open ended but the cost of a new fabrication facility is
in the billions of dollars. This requires the manufacturer to take advantage of
every opportunity to increase the utilization of each facility. Even a small
percentage increase in usable capacity translates to large additional profits
and more market share. On the other hand, the semiconductor industry must face
the difficult challenge of responding to a highly dynamic market characterized
by rapidly changing demands and product mixes and sometimes very brief product
life cycles. At the same time, failure of the manufacturer to deliver product
on time can result in financial penalties or even loss of a customer. It is
clear then, that semiconductor fabs have great costs associated with either
under committing or over committing fab capacity and that short cycle times are
of great advantage.
Therefore the primary challenge
in semiconductor manufacturing is to maximize the throughput of the facility
while responding quickly to customer demands through low cycle times. How do we
respond to this challenge?
Addressing
The Problem
Much of the industry has
recognized the validity of the Theory of Constraints as explained by Eliyahu
Goldratt in his book The Goal. This theory directly addresses our challenge.
The Theory of Constraints (TOC) seeks to maximize system throughput while
maintaining the minimum level of inventory. This also yields the lowest cycle
time and therefore the best customer delivery performance. The basics of this
theory are to identify the bottleneck, gauge the input into the system by the capacity
of the bottleneck, never to let the bottleneck be idle, and then elevate the
capacity of the bottleneck.
Challenges
Posed By The Semiconductor Industry
The challenge has been in how to
implement the theory in a complex reentrant system such as semiconductor
fabrication. For example, what if the bottleneck tool is encountered several
times in the process? What occurrence of the bottleneck do you run first? Does
it matter? What is a practical approach of driving inventory to the bottleneck
based on the bottleneck's consumption rate? As far back as 1988, Glassey and
Lozinski [1] discussed techniques to detect starvation of the
bottleneck. Through the years Dr. Glassey has described various methods to
accomplish this goal. The solutions have ranged from graphical assistance for
operators to queue predictions based on simulation experiments. Dr. Glassey has
also attacked the problem of regulating the flow of material into the process
flow based on a linear control rule called descending control. These approaches
have been influenced by the availability of real time data.
Feeding The Constraint
The
solution presented here to the problem of how to feed the constraint is based
in an old methodology called critical ratio. Critical ratio is a method to
drive dispatching decisions based strictly on customer due date. The ratio is
simply the time the lot is expected to take to complete divided by the time
until it is needed to be complete. This ratio of expected/needed time gives a
higher priority to product that is farther behind schedule. This concept of
driving to an end point based on certain criteria is a good one, but it is
obvious that there are no manufacturing considerations here. Operations people
have long been aware that none of the mainstream dispatching schemes are based
on or even consider the need to manage the production system in an efficient
manner. Therefore, the current need is to provide a dispatching algorithm that
considers not only customer delivery requirements, but also manufacturing
efficiency as a means to that end.
Efficiency
At The Constraint
The
Theory of Constraints asserts that the manufacturing system as a whole can
reach optimum efficiency by identifying and feeding the constraints of the
system. The results of this optimization are to minimize inventory and reduce
cycle times
and therefore reduce manufacturing costs and yield more aggressive and accurate
customer delivery dates. This process of feeding the constraint becomes significantly
more difficult when the process is highly reentrant, and the constraint tools are encountered many times in the
process at differing process rates. How then, is the constraint to be fed and
still maintain a linear inventory profile? This is done by recognizing that not
only must the constraint tool be fed, but that all occurrences of the constraint tool be fed, and that this must
be done in a consistent and equitable manner. If this very important concept is
not understood, the manufacturing system could be left with the circumstance
where there are mounds of inventory in front of the constraint tool, but it is
all bound for a single occurrence of the tool with all other occurrences left
dry. In this case a bubble of inventory has been created in one location and a
hole in the inventory profile created in other locations. A reentrant
constraint must ensure that it feeds itself. Add to this the concept of
maintaining a minimum buffer in front of each occurrence of the constraint to
protect it from the inevitable disruptions and fluctuations that could impede a
constant flow to the constraint.
By considering each occurrence
of the constraint an endpoint, work in process (WIP) can be driven to the
constraint based on criteria relevant to the constraint tool. If the goal is to
keep each occurrence of the tool from starving, then the ratio concept of
expected/needed time can be used. The expected time is the time for a lot to
get to the constraint and the needed time is the time that the lot is needed at
the constraint in order to keep it from starving (or depleting a minimum
buffer). This ratio will be referred to as the Hunger Ratio, because it
describes the degree of "hunger" that the constraint experiences for
each lot. The more material that is in the constraint’s buffer or is likely to
arrive before the current lot, the lower the Hunger Ratio for the current lot
will be. This Hunger Ratio concept can also be used to regulate the rate of
starts into the manufacturing system based on the consumption rate of the
bottleneck. This is another major component of the drum-buffer-rope model used
in the Theory of Constraints. Other variations can be used to schedule
non-production material that is driven by stock point levels (or in TOC terms a
buffer level), rather than due dates. This approach provides a priority factor
based on manufacturing efficiency needs that can be combined with more
traditional drivers based on customer delivery dates. The combination provides
a balanced and synergistic approach to scheduling.
The
Role Of Simulation Modeling
Simulation modeling has become
an indispensable tool in the effort to accurately assess the true capacity of a
facility, allowing manufacturers to more confidently commit the full capacity
of the fab. Any company who fails to avail themselves of this competitive
advantage in an aggressive market does so at it's own peril. Simulation
modeling has also provided an economical way to evaluate different dispatching
philosophies, and with the marriage between simulation and real time
dispatching systems, that advantage is being translated into reality.
Simulation
modeling has been widely used in the semiconductor industry for a wide range of
strategic objectives such as fab design, equipment selection, capacity and
cycle time planning, etc. Its usage in
the tactical area has been relatively limited.
The pertaining areas include short-interval scheduling, dispatching of
individual lots, scheduling of equipment events, short-term deployment of
operators, etc. Historically, lengthy
development and execution times made simulation models impractical to use in
this area. Over the years however, commercial fab simulators have been
developed that dramatically reduce the time for model development and are
optimized for performance. Moreover,
the cost of computer hardware has also declined steeply making it affordable to
run models on high-end workstations. In
spite of all these improvements, changes analyzed and deemed as beneficial by
simulation models do not necessarily translate into actions in the shop floor
on a day to day basis.
The primary reason for this is the lack of
integration between the Manufacturing Execution Systems (MES) and the
simulation and rule development systems. The majority of the information used
in these two systems is common information since the simulation tool is
essentially trying to mimic the physical manufacturing system. The integration
and synchronization of these two systems and the use of common data is
imperative in the attempt to both maintain accurate and timely simulation
models as well as to translate simulation based improvements to the physical
facility.
Currently data models within the simulation
environment encompassing the basic constructs of the fab and their associated
relationships, but they vary significantly from that found in WorkStream. Examples of these include product, route,
equipment, personnel, dependent resources like masks, load boards, etc. In a tactical environment, it is essential
that the virtual constructs and associated rules mesh with the real world. Specifically, in AMD’s case, the tactical
area of interest is real time dispatching.
Dispatching in our case implies the sequencing of products ready for
processing at a given operation or set of equipment. The local pre-requirements to begin processing a product lot at
an operation include the readiness of the product, equipment, operators and
necessary sub-resources like masks, fixtures, etc. However, the scope is generally expanded to include other aspects
such as information on adjacent operations, status of bottlenecks, capabilities
of equipment, lots with special priorities, etc.
AMD has embarked upon a Real Time Dispatch
(RTD) project with the objective to implement dispatching at individual
operations. The system used is the
AutoSimulations Real Time Dispatch product called ASI-RTD. It provides the capability to implement
custom rules specific to certain operations or a group of operations. An important aspect is the capability to
exercise these rules using real-time information. The system’s functionality will include the ability to develop
and test rules or heuristics, in an off-line mode using simulation
modeling. Rules deemed beneficial via
simulation would then be transferred to WorkStream for controlling the order of
processing. The pilot phase of this
project has been implemented in AMD Fab25 and is being managed by a team
consisting of representatives from the production planning, manufacturing
operations, modeling, and computer integrated manufacturing groups.
Apart
from intelligent dispatching, this system has already provided other
benefits. It has enabled a
user-friendly access to the WorkStream data that enables users to develop
reports, which in the past could be done only by programmers proficient in
COBOL and Workstream. These reports are
currently providing timely and accurate information for decision-making
purposes.
2.
C. R. Glassey and Jeyaveerasingam George Shanthikumar, and Sridhar
Seshadri, “Linear Control Rules for Production Control of Semiconductor Fabs”
IEEE Transaction on Semiconductor Manufacturing. Vol. 9, No. 4. November 1989.
3.
Eliyahu M. Goldratt and Jeff Cox, “The Goal” 1992.