Supporting Material is available for this work. For more information, follow the link from
the Table of Contents to "Accessing Supporting Material".
Table of Contents
The Impact of Endogenous Demand on Push-Pull
Production Systems
Paulo Gongalves'
Jim Hines?
John Sterman*
Charlie Lertpattarapong*
University of Miami
School of Business Administration
Coral Gables, FL USA 33124
Work reported here was supported by the Intel Corporation. The authors would like to thank the many
people at Intel that have collaborated to this project. In particular, we would like to thank Mary Murphy-
Hoye, Jim Kellso, David Fanger, Gordon McMillan, Karl Kempf, Daniel Mckeon, George Brown, Brian
Kelly, Robert Kelly, Gene Meieran, Dean Phillips, Roberta Bailey Roberts, Nick Rose, Michael Waithe,
Ann Johnson, Ray Lucchesi and Jay Hopman.
Complete documentation on the model as well as more information on the research program that generated this
paper can be found at http://web.mit.edu/palomar/www/.
! UM, School of Business Administration, KE 404, Coral Gables, FL, 33124. Phone: 305-284-8613; Fax: 305-284-
2321; paulog@miami.edu
2 MIT, Sloan School of Management, E53-309, Cambridge, MA, 02142. Phone: 617-253-9413; Fax: 617-258-7579;
jhines@mit.edu
4 MIT, Sloan School of Management, E53-351, Cambridge, MA, 02142. Phone: 617-253-1951; Fax: 617-258-7579;
jsterman@mit.edu
4 MIT, Technology Management and Policy Program, Cambridge, MA, 02142. Phone: 617-452-5458; Fax: 617-
258-7579; clertpat@ mit.edu
The Impact of Endogenous Demand on Push-Pull
Production Systems
Abstract:
Though often analyzed separately, supply chain instability and customer demand interact
through product availability. We investigate the feedback between supply chain performance and
demand variability in a model grounded in first-hand study of the hybrid push-pull production
system used by a major semiconductor manufacturer. While customers’ response to variable
service levels represents an important concern in industry, with sizable impacts on company
profitability, previous models exploring supply chain instability do not account for it. This
research incorporates two effects of customer responses to availability. The sales effect captures
the negative feedback whereby product shortages cause customers to seek alternate sources of
supply, reducing demand easing the shortage. The production effect captures the delayed impact
of changes in demand on the manufacturer’s production decisions: lower demand leads to
reduced production, prolonging shortages that depress demand, a destabilizing positive feedback.
We show how the sales and production effects interact to destabilize the supply chain and lower
average performance. Supply chain models that assume exogenous demand may therefore
underestimate the amplification in the chain, and the value of inventory buffers. In addition,
incorporating the customer response leads to different inventory and utilization policies than
those in use by the company. The model yields insights into the costs of lean inventory
strategies and responsive utilization policies in the context of hybrid production systems and
endogenous demand.
Keywords:
Supply chain management, push-pull systems, demand amplification, lost sales,
system dynamics, and simulation.
1. Introduction
Despite the supply chain revolution of the past decade, companies in diverse industries
such as computers, electronics, autos, toys, seeds, and pharmaceuticals still struggle with
production and shipment delays. An important effect of such delays, generated by supply chain
glitches and instability, is reduced shareholder value. Hendricks and Singhal (2003) show that an
abnormal decrease of over 10% in shareholder value is caused by part shortages, order changes
by customers, and production ramp-up and roll-out problems, among others. In addition, it has
been long recognized that the bullwhip effect tends to amplify the instability in orders as one
moves upstream a supply chain, potentially making upstream companies, such as semiconductor
manufacturers, more prone to supply chain glitches. For instance, Intel Corporation, the number
one U.S. semiconductor manufacturer, has consistently struggled with part shortages, high
variability in demand, and order changes and cancellations by customers. In November 1999,
facing shortages of Pentium III processors, Intel planned to introduce a new fabrication facility
in the following year. In late 2000, blaming order cancellations by large customers and economic
slowdown, Intel warned that its revenues would fall short of projections and that sales would be
flat for the quarter (Gaither 2001).
The challenge of demand variability, instability and order amplification is complicated by
long production delays. In semiconductors, long throughput times (approximately 13 weeks)
affect the ability of manufacturers to maintain adequate inventory levels in the face of demand
variability. When customer demand varies, factory managers must adjust capacity utilization to
maintain adequate service levels while avoiding excess inventory. However, the combination of
variability in demand and long fabrication delays often leads to alternating periods of scarce and
excess supply. The resulting supply variability can feed back to customer demand and
profitability as a company’s inability to meet demand leads to lost sales, eroded reputation, and
decreased goodwill. For instance, Gateway Inc. increased the number of microprocessors it
bought from Advanced Micro Devices Inc. “to offset Intel's inability to match rising demand”
(Hachman 2000). In addition, Intel’s variability in supply can reduce its profitability. In
December 1998, Intel struggled with shortages of its low-end Celeron microprocessors, allowing
AMD, Intel’s main competitor in the U.S market, to increase its market segment share by more
than two percentage points, even after Intel cut prices on its Celeron chips (Hachman 1999).
The interactions of supply chain instability and customer response raise several
interesting questions: What is the impact of endogenous demand on supply chain variability?
What is the impact of supply chain variability on customer response? What policies can Intel
and other companies implement to stabilize their supply chains?
To address these questions, this research builds and analyzes a model of a semiconductor
supply chain in which customer demand responds to product availability. Based on a year-long,
in-depth field study of Intel’s supply chain, the model captures the material flows of production
and the customers’ response to the manufacturer’s service level. In particular, it incorporates two
effects of customer response. First, the sales effect captures the negative feedback whereby
product shortages cause customers to seek alternate sources of supply, reducing demand easing
the shortage. That is, a change in demand feeds back to mitigate the impact of the initial
disturbance. Second, the production effect captures the delayed impact of changes in demand on
the manufacturer’s production decisions through a positive feedback loop. If demand falls,
manufacturers reduce demand forecasts and capacity utilization to avoid excess inventory. After
the production delay, lower production leads to lower inventory and service levels, causing a
further drop in customer demand. The delayed production effect generates a reaction that
reinforces the impact of the original disturbance.
We show that endogenous customer response to availability leads to greater supply chain
instability compared to models in which customer demand is treated as exogenous. Therefore,
supply chain instability models that assume exogenous demand may underestimate the
amplification in demand and the value of inventory buffers. Moreover, treating demand
endogenously leads to different inventory and utilization policies than those currently in use by
the firm. In particular, the supplier should maintain higher safety levels in assembly work-in-
process (WIP) and finished goods inventory (FGI); and reduce the responsiveness of utilization
to changes in customer demand caused by inadequate service levels. Based on the costs
associated with lost sales and holding inventory at assembly and finished goods, we derive a
recommendation for the optimal location and quantity of safety stocks. The policy heuristic
provides a sharp reduction in supply chain instability and minimizes the impact of lost sales. The
model analyzed in this paper gives insights into the costs of lean inventory strategies and
responsive utilization policies in the context of hybrid production systems and endogenous
demand.
The next section reviews the relevant literature. Section 3 discusses the research site and
the following section presents the model. Section 5 introduces the simulation results, analyzes
the model, and derives policies for supply chain stabilization. We conclude with discussion of
our main results, managerial and theoretical implications, and directions for future research.
2. Literature Review
Supply chain instability and the influence of inventory level on demand have attracted the
attention of researchers and practitioners in different fields, such as economics, system dynamics
and operations management. In economics, research on supply chain instability dates back to
Thomas Mitchell’s (1924) descriptions of the mechanisms through which retailers caught short
of supply increased their orders to suppliers. In system dynamics, the study of supply chain
instability helped lay out the foundations necessary to create the field (Forrester 1958, 1961).
Subsequent research explored applications in diverse areas including interactions between the
supply chain and labor force (Mass 1975); the performance of Material Requirements Planning
(MRP) systems (Morecroft 1980); laboratory study of people’s ability to manage complex
systems such as the supply chain in the Beer Game (Sterman 1989a, 1989b); decision-making
under varying levels of feedback complexity (Diehl and Sterman 1995); and the impact of
business cycles on capital equipment supply chains (Anderson and Fine 1999). In addition,
models in the system dynamics tradition often incorporate the feedback of inventory availability
on customer demand, like Forrester’s (1968) “market growth” model and Graham’s (1977)
model investigating the impact of adding a minor loop to oscillatory systems. While this paper
emphasizes the combination of endogenous customer response with supply chain instability, the
major contribution to the system dynamics literature is our investigation of the impact of the two
in hybrid push-pull production systems.
In operations management, research investigating supply chain instability typically
assumes exogenous customer demand while studies exploring the influence of inventory level on
customer demand do not consider multiple-stage supply chains. Examples of the former include
Lee et al.’s (1997a, 1997b) models of demand signal processing, rationing, order processing, and
price variations; Baganha and Cohen’s (1998) hierarchical model; Graves’ (1999) single item
inventory system with non-stationary demand; and Chen et al.’s (2000) model with a demand
forecasting technique and order lead time. Examples of the latter include Dana and Petruzzi’s
(2001) extended newsvendor model where customers choose between the company and an
outside supplier; Gans’ (1999a, 199b) dynamic model of individual consumer behavior, where
consumers update their prior beliefs about the company after each contact; and Hall and Porteus’
(2000) model where the expected service level is a function of firm capacity and firms compete
by investing in capacity to service customers. Our research fills a gap in the operations
management literature by exploring both the effect of endogenous customer responses and
supply chain instability. Our results extend Danna and Petruzzi’s (2001) result to the case of a
multistage supply chain with production delays, showing that when a company accounts for the
effect of inventory availability on demand it is optimal to hold more safety stock.
3. Research Site
The results draw on a year-long, in-depth analysis of Intel’s supply chains. Intel is the
technology leader in microprocessor manufacturing. Among many firsts, Intel was the first to
produce 0.13-micron technology, allowing it to double the size of the processor's cache memory
while reducing die size by over 30%. Such improvements resulted in faster microprocessors and
increased number of chips manufactured per wafer. The company was also the first to transition
from 200 mm to 300 mm wafers, leading to higher chip production efficiency. To manage the
variability in product line, production, and demand, Intel employs about 1,500 planners who
address short- and long-term production decisions, using sophisticated systems and detailed
guidelines directing decisions. Model development entailed interviewing planners with diverse
decision scopes and responsibilities to understand the decision making processes in Intel’s
production system. In addition, the research team interviewed managers in diverse areas of the
corporation, such as operations, supply chain management, information technology, demand
forecasting, marketing and sales. In total, we conducted almost one hundred semi-structured
interviews both through site visits and weekly conference calls. The research also involved
reviewing Intel’s logs detailing guidelines for decision-making, and collecting related
quantitative and qualitative data. The former included time-series data on quarterly capacity,
utilization, wafer starts, shipments, forecasts, service level, and market share. The latter included
managers’ decision heuristics, company’s guidelines and incentives, and information
dependencies among business areas. These data helped us establish the assumptions used in the
model that captures Intel’s semiconductor manufacturing.
4. Model Assumptions
Semiconductor manufacturing is commonly divided into the “front-end,” including the
initial steps of fabrication and sorting, and the “back-end,” including assembly/testing,
packaging, and distribution. The front-end process takes place in a wafer fabrication facility, or
Fab. The process takes 200 mm/300 mm polished disk-shaped silicon substrates, the wafers, as
inputs and transforms them, through multiple stages of photolithography and etching, into
fabricated wafers, composed of hundreds of -inch square integrated circuits, known as dies.'
Wafers are cut into dies and stored in Assembly Die Inventory (ADI) warehouses, collocated
with Assembly/Test plants. In the back-end, the dies are first tested; upon passing the tests, they
receive a protective package and metal connections, completing the microprocessors, or chips,
which can then be stored in finished goods warehouses. The model proposed represents the
manufacturing process by a three stage supply chain consisting of fabrication, assembly, and
distribution (Figure 1).
' The actual number of dies per wafer range from 100 to 1000, depending on the chip architecture — whether the chip
is “logic” or “memory” — and its specific design. Each die is composed of individual devices such as transistors and
memory cells.
Replacing
Shipments .
Finished
Goods
Inventory
Assembly
wIP Net
Assembly
+ Completion
Shipments
Production
Rate
Desired vine Customer
Wafer 4 Forecasted , Demand
Stats “~~... Customer
Figure 1 —-Hybrid push/pull production system for semiconductors.
Thick lines and patterned background refer to a push system, indicating that the upstream production
process operates as a push. Thin lines and clear background refer to a pull system, indicating that the
downstream stages operate as a pull.
In addition, microprocessor production takes place in a hybrid push-pull production
system, combining a push system at upstream stages and a pull system at the downstream stages.
Therefore, the front-end is characterized by a push production system: long-term demand
forecasts, updated weekly, and adjustments from fabrication and assembly WIP serve as the
basis for the desired wafer production rate, or wafer starts. In contrast, the back-end operates as a
pull system, with assembly/testing, packaging, and distribution based on current customer orders.
Four main assumptions based on the fieldwork drive the behavior of the model.” The first
three assumptions address managers’ decisions regarding (a) push and pull production, (b)
capacity utilization, and (c) demand forecasting. These assumptions reflect Intel managers’
locally rational heuristics to control their systems. While they are not optimal, they reflect
heuristics managers use to make their everyday decisions and evolved because they are locally
? A full description of the model, formulations, and assumptions can be found in Gongalves (2003).
10
adapted to conditions in the company and its fabs. The fourth assumption captures customer
response to inventory availability.
4.1. Push and Pull Production Decisions
All production decisions ultimately depend on customer demand. Current demand drives
shipments and assembly completions; long-term demand forecasts influence production starts.
All incoming orders are logged by Intel’s information system and tracked until they are shipped
to customers or cancelled. If the microprocessors are available in FGI, orders can be filled
immediately. Therefore, incoming customer orders “pull” the available microprocessors from
FGI. In turn, replenishment of FGI shipped to customers “pulls” microprocessors from assembly.
If the microprocessors are not available in FGI, backlogged orders “pull” parts directly from
assembly die inventory (ADI). Since the parts have to be assembled, filling orders from ADI
increases the delivery delay experienced by customers and reduces the flow of shipments below
customer orders as ADI inventory and assembly capacity limit shipments.
4.1.1. FGl and Assembly Pull
To model the pull characteristic of assembly and finished goods, we must capture their
dependency on current demand. Actual shipments (S) from FGI are given by the minimum of the
desired (pull) and feasible (push) shipment rate. By design, shipments operate in a pull mode,
with shipments being determined by the desired rate, however, if not enough FGI is available the
system will ship out only what it is possible. Desired shipments depend on the ratio of backlog
(B) and the desired delivery delay (DD’). Feasible shipments depend on the stock of FGI and the
minimum order processing time (gp); a first-order process is assumed for simplicity.
S(t) = MIN(S (1), Sax O) qd)
11
S(t) = B(t)/DD® (2)
Suax (1) = FGI(t)/Top G3)
While shipments (S) deplete FGI, the net assembly completion rate (Ay) replenishes it.
The product of gross assembly completion rate (Ac) and the unit yield (Yv), i.e. the fraction of
good chips per assembled die, define the net assembly completion rate (Ay). The gross assembly
completion rate (Ag) is given by the minimum of the desired (a pull signal) or the feasible (a
push signal) gross assembly completion rate (equation 4). By design, assembly operates in a pull
mode, with assembly outs being determined by the desired rate (equation 5). However, if not
enough WIP is available the system can complete only what it is feasible (equation 6). The
feasible assembly rate is determined by the availability of assembly WIP (A WIP) and the
assembly time (7); for simplicity a first-order delay is used. The desired gross assembly rate
(A’e) is determined by the desired net assembly rate (Ay) adjusted by the unit yield (Yy).
Ag (t) = MIN (Push, (t), PullAg (t)) (4)
PullAg (t) = AyO/Y, (5)
PushAg(t) = AWIP(t)/T., (6)
The determination of desired net assembly rate (A"y) by division planners begins with
recent shipments (ES), a proxy for current demand that is more stable and reliable than orders.
Desired net chips out is then adjusted above or below recent shipments to close any gap between
target and actual FGI and to eliminate excess backlog (equation 7).
A= aro ES'
(+ FGI (t)- FGI) , a-F') fa
Tren Tp,
12
where FGI’ and FGI are target and actual finished goods inventory, B” and B are
target and actual backlog, and tq and 7 are the adjustment times for the
elimination of gaps between them.
4.1.2. Fabrication Push
The wafers produced in the fabrication process are pushed into the Assembly Die
Inventory (ADI) where they are stored until orders for specific products pull them from ADI into
assembly and distribution. While the gross assembly completion rate (Ag) depletes AWIP, the
die completion rate (D;) replenishes it. The die completion rate (D;), measured in dies/month, is
given by the net fabrication rate (Fy), measured in wafers/month, adjusted by the number of dies
per wafer (DPW); the die yield (Yp), i.e. the fraction of good die per wafer; and the line yield
(Y;), i.e. the fraction of good fabricated wafers (equation 8). The gross fabrication rate (FG) is
determined by the availability of fabrication WIP (FWIP) and the fabrication time (7); for
simplicity a first-order delay is used (equation 9).
D, (t) = Fg (t)» DPW -Y, -Y, (8)
F(t) = FWIP(t )/T, (9)
While the gross fabrication rate (Fg) depletes fabrication WIP (FWIP), wafer starts (WS)
replenish it. The decision on actual production rate, wafer starts (WS), is based directly on the
desired wafer starts (WS"). Fab planners determine the desired wafer starts considering the
desired die inflow (D,’) requested by Assembly/Test plants and an adjustment for fabrication
work-in-process. The latter is based on managers’ heuristic to maintain fabrication WIP (FWIP)
at a desired level (FWIP’). Equation 10 shows fabrication planners’ heuristic for managing wafer
starts.
13:
(10)
ws" ()=ma{o D;(t) n furry niet)
“DPW -Y,-Y, Cin
where Zywip is the fabrication WIP correction time; and the non-negativity
constraint prevents negative production targets.
The desired die inflow rate (D/) depends on long-term demand forecasts (ED) and an
adjustment from assembly WIP. The assembly WIP adjustment component reflects assembly
planners’ goal to replenish assembly WIP when the current level is below the target to correct the
discrepancy over time (typ). Equation 11 shows division planners’ heuristic for managing the
desired die inflow (Dr), incorporating information on long-term demand forecast (ED)?
Division planners provide information on the desired die inflow to Fab planners, allowing them
to set production starts.
(11)
Dj (t)= MAX [ ED(t)/¥, -Auro- ari)
Tawp
where Zwip is the assembly WIP correction time; and the non-negativity
constraint prevents negative die inflow rates.
4.2. Capacity Utilization
To set the capacity utilization (CU) of their Fabs, managers consider the desired
production rate and the available capacity. Capacity utilization is a nonlinear function of the ratio
of desired wafer starts (WS’) and available capacity (K) operating at the normal capacity
utilization level (CUy).*
Ws’ (t)
Ci =fu
OO= tale ou,
) (12)
3 While the heuristic for the desired die inflow rate does not take into consideration the adjustment from FGI
explicitly, information from FGI is used to set the desired assembly WIP. (AWIP’).
* We assume that the normal capacity utilization level at Intel to be equal to 90% of maximum capacity.
14
When desired production (WS) equals the normal capacity utilized, capacity utilization is
set at the normal operating point (90%), allowing all desired production to be met. The
remaining 10% slack capacity is often used for engineering purposes (process improvement and
development runs) as well as to accommodate manufacturing instability. When desired
production is large relative to normal capacity utilized, Fab managers increase utilization,
therefore reducing the capacity that is available to engineering. The opposite takes place when
desired production falls. If the function lay on the 45° reference line, utilization would vary
exactly enough to ensure that wafer starts always equaled desired starts exactly (subject to the
capacity constraint). Field study showed, however, that the utilization function characterizing
actual wafer start decisions lies above the 45° reference line and has a flatter slope at the normal
operating point. Fab managers seek to avoid shutdown and prefer to keep their fab running even
when desired starts fall below normal, preferring instead to build inventory; similarly, they
increase output less than enough to meet desired starts fully when desired starts exceed normal
output so as to maintain some room for engineering purposes and to avoid yield problems. A
concave function where f, 20, fy >0, fy <0, fy,(0)=0, fy) =CU nom Fy (2) = CU ax»
captures the response of Fab managers to variations in desired wafer starts relative to capacity.
While the general shape of the function (fy) is plausible, the slope of the function around
the normal operating point and the normal utilization fraction play an important role in model
behavior. Data for estimating such parameters are both proprietary and Fab specific. Therefore,
we provide sensitivity analysis (section 5.3) over a broad range of plausible parameters for
capacity utilization functions and investigate the impact of these parameters on model behavior.
15
4.3. Demand forecasting
The marketing organization is responsible for demand forecasting at Intel. As in many
firms, marketing generates an initial demand forecast for microprocessors based on estimates of
customer demand from different geographic regions and customer types (for another example in
the semiconductor industry, see Sterman (2000), pp. 449-462). A process known as “Judged
Demand” is then used to go from the initial to a final forecast. The “Judged Demand” process
receives its name due to the judgment and subjective adjustments involved in elaborating the
forecast. First, favorable (unfavorable) macroeconomic indicators are incorporated to increase
(decrease) the initial estimates based on the total available market for personal and business
computers. Second, an “executive adder” process often adjusts the aggregate forecast upwards to
reflect the optimistic goals and aspirations of company executives. Finally, the marketing group
“filters” (i.e., smooths) the demand estimates from different regions to account for local
incentives. With respect to regional information, according to a platform manager in marketing:
“Customer numbers get rolled up, aggregated, and judged with a set of assumptions that may or
may not be correct; and customer-level insight, when provided, gets watered down.” In
particular, when demand for certain products is high, regional warehouse managers tend to
increase their orders to ensure that they will be able to meet demand, the familiar “phantom
ordering” generated as different customers compete for larger slices of what they perceive to be a
shrinking pie (Forrester 1961, Sterman 2000, pp. 743-755, Goncalves 2003). In contrast, when
demand is low, regional managers have the tendency to decrease orders to make sure they are not
stuck with undesired inventory. Therefore, marketing “filters” the forecast to come up with its
final forecast. Analysis of the forecast data confirmed this: when compared, regional forecasts
were more variable than marketing ones.
16
We model the demand forecast (ED) as a first-order exponential smooth of actual orders
(D) - in practice obtained from the aggregation of regional orders — updated over a period of one
month (7p,qj), the frequency with which marketing updates their forecasts.
D(t)- ED(t)
ED(t)= (13)
Tsai
For simplicity, we do not take into consideration the random macroeconomic factors that
may influence the demand forecast and the executive adder process, making the a fortiori
assumption that marketing is able to filter out the noise and bias caused by these processes.”
4.4. Customer response
In this model we capture customers’ response to supply availability, measured by the
fraction of orders filled (FoF). Customers respond to a low fulfillment fraction by seeking
alternative sources of supply; as they succeed, their orders drop. Intel’s attractiveness to
suppliers (A;) is a nonlinear function of customers’ perception of supplier delivery reliability
(PFoF). In turn, customers’ perception of delivery reliability (PFoF) adjusts from the actual
delivery reliability — Fractional orders Filled (FoF) — with a third-order Erlang lag (A), with an
average time constant of six months. The third-order Erlang distribution captures the plausible
distribution of responses by OEMs. At the instant of a decrease in the service level, all OEMs
will still perceive the supplier as reliable, and there will be no shifts to alternative sources of
supply. Therefore, the immediate response of the distributed lag should be zero. If service level
remains low, however, some customers will change their perceptions about supplier reliability
and seek other suppliers. The distribution of OEMs’ reactions eventually peak, and then
decrease, reaching zero after a sufficient time has elapsed. The delay captures the time required
5 It is straightforward to add random noise to the forecast to capture the impact of these sources of error and
adjustment.
AT.
for OEMs to perceive changes in availability, to determine that the changes are not temporary
and warrant a search for alternative sources, and to close deals with those alternative sources. For
simplicity, we assume that competitors maintain constant delivery performance (i.e. a constant
attractiveness (Ac) over time). This assumption allows us to measure changes in system behavior
due to customers’ reactions only due to changes in supplier conditions; relaxing it is a promising
direction for future work.
The nonlinear function (f4), characterizing Intel’s attractiveness (Aj), is a logistic curve.
Attractiveness varies on a scale going from zero to one (0 $ Ajyyi, < Azjyay £1). A logistic curve
captures customers’ mild response to small changes in supply availability, and more significant
responses to large changes in supply availability.
A,()= f,(PFoF (0) (14)
While the logistic shape of the function is plausible the model behavior depends heavily
on the slope of the function and the minimum value. At the same time, the data for estimating
such parameters are not reliable or easily available. Here too we provide sensitivity analysis
(section 5.3) over a broad range of plausible parameters for the function governing customer
responses and investigate the impact of these parameters on model behavior.
4.5. Feedback structure
The Intel managers’ heuristics (push-pull production, capacity utilization, demand
forecast, and customer response decisions) close the feedback loops shown in Figure 2.
Balancing loop (B/) — FGI Pull — describes the company’s pull system operating at the finished
goods inventory (FGI) level. An increase in backlog (due to additional orders) increases desired
shipments, boosting shipments and reducing the backlog—if there is sufficient FGI. When the
availability of FGI is limited, the negative loop FG/ Availability (B2) — limits shipments and
18
forces the backlog up.
Available 2, 2 (ad
Capacit
oe > —— [Fat avaiiabinty
J Finished yi
& Fabrication p>) Assembly | Goods
wip Net F =
Production Assembly Shipments
ate Completion —
A e + na Fai Pull
Desired tes)
Capacity gust + vv; Adust Ani | Shipments
Utilization Fabwip Lost
Mion FabWiP “Adjust Be AWIP Adjust _/ Replenishment J + + sales Fraction of
Adjust Adjust Fal + duigohig Orders Filled
es AWIP é Expected Orders ——?"
eee 4 Desired Net Shipments (ae
Starts + Desired Net i) Assembly + iia) R2 DELAY
_ Production oy Completion hes) Gath
Rate Production Incoming Through
+ Push “sa Orders, Service +,
Bkig 5 ike Market
Forecasted ‘Adjust \ othe
pa Customer *~
——[petay | Demand
Figure 2 - Supply-Demand feedback process for Intel’s hybrid production system.
With FGI constraining shipments, the pull system cannot operate at the FGI level.
However, the system can still pull from assembly WIP. The balancing loop Assembly Pull (B3) is
analogous to the FGI Pull loop, but pulls chips out of the assembly WIP, which takes longer.
Therefore, when FGI is constrained, the system still operates as pull but with longer fulfillment
delay. The actual assembly completion rate is also adjusted by two other loops: a balancing loop
that corrects the levels of finished goods inventory (Adjust FGI — B4) and a reinforcing loop
(Replenishment — R1) that replenishes all shipments from FGI. The system can pull from
assembly as long as the assembly WIP inventory is sufficiently high. When the availability of
assembly WIP decreases, the first order control for assembly — balancing loop AWIP Availability
(BS) — prevents AWIP from going negative. If assembly WIP constrains net assembly, the
system can’t even pull from AWIP, and the entire system reverts to push.
Production at the upstream stage is based on long term demand forecasts, information on
19
the desired assembly completion rate, and adjustments due to inventory corrections in assembly
(Adjust AWIP — B6) and fabrication (Adjust FWIP — B7). The push part of the production system
is determined by the first order control for fabrication — balancing loop Wafer Availability (B8).
In terms of customers’ response, the reinforcing loop Growth Through Service (R2) describes the
ability of the company to grow its market share as it is able to meet customer demand. In
contrast, the balancing loop — Lost Sales (B9) — describes the inverse dynamics. As customer
demand grows, the company’s ability to maintain its service level (fraction of orders filled)
decreases, reducing its ability to retain customers. If the company cannot adequately fill
customer orders, it will lose market share to competitors. Finally, the feedback from the
company’s supply chain to customer demand is described in the reinforcing loop — Production
Push (R3), which captures the long delays associated with production and customer reactions. If
demand falls, manufacturers reduce demand forecasts and capacity utilization to avoid excess
inventory. After the production delay, lower production leads to lower inventory and service
levels, causing a further drop in customer demand. These feedback processes are capable of
generating the dynamic behavior observed in the company and replicated in the model.
5. Model Analysis and Results
The model constitutes a ninth-order nonlinear differential equation system. Since the
equations are highly nonlinear it is not possible to obtain closed-form solutions. Hence, we
simulate the model to gain intuition on its behavior. While the parameters chosen for the base
case (Table 1) reflect Intel’s manufacturing system, the values are disguised to maintain
confidentiality.
20
Table 1. Base Case Parameters.
Parameter Definition Value Units
D Customer demand 5.0 Million units /month
MS Initial market segment share 80 %
DPW Number of die per wafer 200 Die/wafer
CUn Normal capacity utilization 90 %
Y, Number of good wafers per total produced 90 %
Yp Number of good die per wafer 90 %
Yu Number of good chips per good die 95 %
K Available capacity 28.9 ‘000 wafers/month
For a given customer demand (D), the equilibrium capacity (K) required to meet that demand can be
computed from the normal capacity utilization and yields. The formula for equilibrium capacity (K) is
D-MS
given by: k=")
CU, -DPW-Y, -Y, -Y,
Figure 3 shows the behavior of backlogs and finished goods inventory coverage for two
simulation runs. The model is initialized in equilibrium with constant industry demand. Of
course semiconductor demand has been growing exponentially for decades, but since we are
interested only in the stability of the system around the demand, we focus on a detrended
demand signal. Interactions of growth with supply chain stability are left for future research. In
equilibrium the hybrid push pull system functions as intended: the supplier meets its target
delivery delay, fills 100% of incoming orders, and maintains the desired quantities of FGI,
AWIP, and wafers. From equilibrium we introduce a demand pulse by increasing customer
demand by 5% and then 20% respectively for a single month at the end of the first simulated
year. The demand shocks increase the backlog (Figure 3a). As planners observe the increase in
demand and backlogs, they quickly realize the need to raise production, and desired wafer starts
rise (Figure 3b). Since capacity is fixed in the short run, managers’ must raise capacity utilization
(Figure 3c) to increase wafer starts. Fab managers quickly adjust utilization to its maximum.
Higher utilization increases the level of fabrication (Figure 3d), assembly WIP (Figure 4a), and
21
FGI coverage (Figure 4b). After fabrication and assembly delays, finished goods eventually
become available to meet the demand.
Backlog Coverage (months)
Desired Normalized Wafer Starts (months)
0.4 300
Pulse 20% Pulse 20%
200 Pulse 5%
0.3 Pulse 5%
. 100 soblte
02 x Equilibrium 0 Equilibrium
0 12 24 36 48 12 24 36 48
Time (Month) Time (Month)
Capacity Utilization Fabrication WIP Coverage (months)
2 Pulse 5% 35 2
7 Pulse 5%
0.8
og | Eauiliorium Pulse 20% a ae
Equilibrium \ pulse 20%
0 25
0 12 24 36 48 12 24 36 48
Time (Month) Time (Month)
Figure 3 — (a) Backlog coverage, (b) desired wafer starts, (c) capacity utilization and (d)
fabrication coverage for the two simulated scenarios.
In both cases, the system immediately responds to the surge in backlog by increasing
shipments (not shown) and pulling more chips from FGI. In the case of the 5% increase, the
depletion in FGI is insufficient to constrain shipments. Here, while the demand shock creates
some supply chain instability, safety stocks in FGI and AWIP allow the system to operate as
desired, i.e. as a hybrid push-pull system. Despite the shock, the company is capable of meeting
its target delivery delay and filling 100% of its incoming orders.
The 20% shock, however, produces a different outcome. For a large enough shock, the
safety stocks in FGI and AWIP are not capable of maintaining the system in desired operation
mode. Here, the system behaves as a pure push system, reacting to demand changes with the
22
much longer production delay. The depletion in FGI constrains shipments, limiting the fraction
of orders filled down (Fig 3c). With FGI constraining shipments, the pull system cannot operate
at the FGI level. The system compensates for the lack of FGI, however, and pulls chips from
assembly WIP, increasing assembly rate. As the availability of assembly WIP decreases, it
eventually constrains assembly. Now, the system can’t pull from AWIP, so it reverts to push.
Assembly WIP Coverage (months) Finished Inventory Coverage (months)
12 0.300
Pulse 209 2
Equilibrium fae Pulse 20% Pulse 5%) Pe 20%
\ Pulse 5% i
1 ie 0.275 ! sae
: ¥ ar . t | J
Equilibrium
1.0 0.250
0 12 24 36 48 OO 12 24 36 48
Time (Month) Time (Month)
Perceived Fraction of Orders Filled Market Share (%)
1.00 p 81
Equilibrium’ * f Equilibrium Pulse 5%
<quilibrium” Pulse 5% :quilibriut \ fi ulse 5%
0.95 80 seen =
ro
\puise 20% oe \
0.90 79 Pulse 20%
0 12 24 36 48 0 12 24 36 48
Time (Month) Time (Month)
Figure 4 - (a) Assembly WIP and (b) finished inventory coverage, (c) perceived fraction of
orders filled and (d) market share for two simulated scenarios.
In push mode, the supplier is unable to meet all customer orders. Customers (i.e. OEMs)
perceive the drop in service level (Figure 4c) after a delay (accounting for decision-making and
reporting delays in information systems) and seek alternative sources of supply. Market share
decreases (Figure 4d) and the drop in orders contributes to the reduction in backlog coverage.
Ultimately, the company’s inability to meet customer demand results in reduced market share,
offsetting the impact of the original increase in demand. As customer demand continues to
23
decrease, it eventually equals the volume of shipments, allowing the backlog coverage (Figure
3a) to stop increasing and the fraction of orders filled to stop declining. Even after additional FGI
becomes available, market share continues to decrease due to the delay in customers’ perception.
Customers’ response to inventory availability feeds back to the supplier decision on production.
Capacity utilization (Figure 3c) drops as the supplier reacts to declining demand. The decrease in
utilization lowers the level of fabrication, assembly WIP, and FGI coverage. When customers
finally perceive improved company performance, orders increase and market share rises. With
time, orders increase past shipments, leading to an increase in backlog. Once again shipments are
not sufficient to meet customer demand and the fraction of orders filled decreases. The 20%
shock in demand shock generates an oscillation response that decays as some of the excess
demand is lost and the supplier closes any remaining demand gap with capacity utilization above
normal. The interaction of the firm’s locally rational decision rules for shipments and capacity
utilization with the market’s response to product availability lead to a lightly damped oscillation
when the system is subjected to a single shock, and the firm’s market share remains depressed
for a long time after the temporary increase in demand.
5.1. Impact of Pull System
To obtain additional insight into the causes of oscillation, we first consider how the
balancing FGI Pull (B1) and Assembly Pull (B3) loops influence system behavior. The pull from
FGI allows the company to close the gap between desired shipments and actual shipments by
running down backlog. Naturally, this loop can only operate while there is sufficient finished
goods inventory to allow shipments to take place. The ability of the FGI pull loop to operate so
effectively is due to the short time constant (1 week) associated with the desired shipment rate.
When that loop is off — shipments are a function of the level of FGI — the system operates as a
24
push system, and the oscillatory behavior of the system increases (Figure 5). Similarly, the
balancing Assembly pull loop pulls inventory from assembly to allow the FG/ Pull loop to
operate as desired. This loop can only operate while there is sufficient assembly inventory to
allow assembly completions to take place. It also has a short time constant (1 month) determined
by the time to adjust backlog. Turning off the pull from assembly reduces system stability as it
restricts the ability of the FGI pull loop to operate effectively. Figure 5 shows the effect of
turning off the FGI and assembly pull loops compared to the base case.
Assembly WIP Coverage (months) Finished Inventory Coverage (months)
13 0.50
A Pulse 20% (AWIP+FGI Pull)
No AWIP Pull _
44 bese: “i iv va eee by
Ak e— No FGI Pull
09 No AWIP Pull 0.20 Aaa No FGI Pull
0 12 24 36 48 0 12 24 36 48
Time (Month) Time (Month)
Figure 5 —- FGI and AWIP Pull loops turned off compared to the 20% pulse in demand.
5.2. Impact of Endogenous Demand
The impact of endogenous demand on system behavior offers further insight into the
causes of oscillation. To make demand exogenous, we knock out the customer response loop by
setting the time to perceive the fraction of orders filled to a very large number (pr =). This
change breaks the feedbacks from the fraction of orders filled (FoF) to customer demand (D),
making demand exogenous while reducing the system to a production and backlog response to a
change in demand. The production loop establishes “wafer starts” based on expected demand and
inventory adjustments throughout the chain. Notice first that if inventory levels in fabrication,
assembly, and finished goods are fully visible to managers and they use the same time constant
for inventory adjustments, the system could be reduced to an effectively-first-order system
25
(Graham 1977). The behavior would then be a smooth increase in production to meet the
additional pulse in demand, followed by a smooth decline.
The structure of the production loop, however, is different. First, while the system has
full visibility, FGI is only used to set the desired level of assembly inventory, instead of also
being used to set the rate of assembly outflow. Second, expected shipments, smoothed with a
short time constant, are used to inform the adjustment needs for desired inventory, but expected
demand, smoothed with a long time constant, is used to set production. Hence drops in shipments
confound changes in demand with changes in availability: inventory shortages that may constrain
shipments send a spurious signal that additional output is not needed. Finally, the balancing FGI
pull and assembly pull loops introduce additional complexity to the production process. The
resulting behavior is damped oscillation of production in the manufacturing supply chain (Figure
6). By increasing the time adjustment for the inventory corrections or smoothing the expected
demand over a longer time constant, we can dampen the oscillations.
Fabrication WIP Coverage (months) Finished Inventory Coverage (months)
3.5 0.300
A No Customer Response
A NoCust.Resp.
ee Sy
\ biee 20% \bae 20%
2.5 0.250
t) 12 24 36 48 0 12 24 36 48
Time (Month) Time (Month)
Figure 6 — Production response to demand change compared to the base case.
The interaction of customer response with the rest of the system amplifies the oscillatory
behavior of production. As demand increases, the company’s ability to meet demand decreases.
With a delay, customers perceive that service levels are decreasing. In parallel, the company
increases production. After the manufacturing delay, the additional finished goods allow the
26
company to meet a greater fraction of demand than it would otherwise. However, just as more
finished goods become available, the delayed responses from customers reduce orders. Hence,
the manufacturer finds itself with more inventory just as demand drops. The combined effect of
excess FGI and the drop in demand decreases production, limiting the company’s ability to meet
future demand. While the system converges, the interaction of customer response and production
amplifies supply chain instability.
While endogenous demand amplifies supply chain instability, more importantly it affects
the company’s policies regarding capacity utilization and inventory control. Consider first the
inventory policy. Since the system reaction to a change in demand is more stable when demand
is assumed exogenous, a tight inventory policy, with reduced levels of safety stock, is still
capable of providing a high service level. However, more unstable systems, as in the case of
endogenous demand, require larger inventory buffers to provide the same level of service.
Therefore, when demand is assumed exogenous, a lean inventory policy is recommended; when
demand is assumed endogenous, inventory buffers are recommended. Figure 7a and 7b show net
present value of the costs associated with endogenous and exogenous demand under the two
inventory policies for the cost structure defined in the following section.®
Now consider the capacity utilization policy. When demand is endogenous, inventory
availability affects demand. Inventory shortages that may constrain shipments decrease service
level and customer demand, sending a spurious signal that additional output is not needed. Since
the decrease in demand is caused by an inventory shortage, additional output is highly desirable.
An unresponsive capacity utilization policy that does not lower production level due to a
decrease in demand provides a higher service level when demand is endogenous. In contrast,
when demand is exogenous, inventory availability does not affect demand. A responsive capacity
° The lean and safety inventory policies are tested with 0% and 5% respectively of safety stock in AWIP and FGI.
27
utilization policy allows the company to prevent the accumulation of excess inventory during
periods of low demand. Therefore, when demand is assumed exogenous, a responsive capacity
utilization policy is recommended. In contrast, an unresponsive utilization policy is
recommended when demand is assumed endogenous. Figure 7c and 7d show net present value of
the costs associated with endogenous and exogenous demand under the two capacity utilization
policies for the cost structure defined in the following section.”
NPV Costs ($) — Exogenous Demand NPV Costs ($) — Endogenous Demand
80M Safety Stock NN Lean Inventory ~~,
= -
_-—4 = 7
40M gor
Lean Inventory Safety Stock
OM OM
0 12 24 36 48 } 12 24 36 48
Time (Month) Time (Month)
NPV Costs ($) — Exogenous Demand NPV Costs ($) — Endogenous Demand
80M Unresponsive 80M Responsive
Utilization \ Utilization
40M 40M Unresponsive
Utilization
Responsive
Utilization
OM OM
} 12 24 36 48 0 12 24 36 48
Time (Month) Time (Month)
Figure 7 - Impact of endogenous and exogenous demand of inventory and utilization
policies.
Figure 7 shows that typical policy prescriptions of adopting lean inventory and
responsive utilization policies are reversed when demand is made endogenous. With endogenous
demand, inventory buffers and an unresponsive capacity utilization policy yield higher service
levels.
7 The comparison for capacity utilization policies are performed with 0% safety stock in AWIP and FGI to make the
impact of the utilization policy more salient.
28
5.3. Sensitivity Analysis
Model behavior is highly sensitive to the assumptions embedded in the capacity
utilization (fy) and customer response (f,) functions. In particular, the model is sensitive to (1)
the slopes of the nonlinear functions fy and fy, (2) the maximum capacity utilization, and (3) the
minimum of the customer demand response. The sensitivity analysis follows a common
procedure to obtain its results. We represent each nonlinear function — capacity utilization and
customer response — as a linear combination of two polar cases, capturing extreme assumptions.
By varying the weight in the linear combination it is possible to obtain a range of behaviors in
the model.
5.3.1. Sensitivity to Capacity Utilization
Consider two extreme cases of Fab managers’ reactions to desired production: responsive
and unresponsive. Both managers respond to increases in desired production the same way. They
respond differently, however, to decreases in the desired production volume. A manager with
unresponsive reactions, characterized by function (fy), does not respond much to a reduction in
desired production. When desired production is low, the unresponsive manager prefers to keep
the Fab running and build up inventory levels down the chain rather than slowing the line or
shutting down. In the limit, of course, utilization must fall to zero as desired production falls to
zero. The unresponsive policy is shown in Figure 8; utilization has a flat slope near the normal
operating region. In contrast, a responsive manager, characterized by function (fy), responds
aggressively to decreases in desired production by cutting utilization in proportion to the decline
in desired production, avoiding the buildup of inventory and making the unneeded capacity
available for process improvement, test runs or preventive maintenance (Figure 8). Intermediate
cases are obtained from the linear combination of the two extremes; the base case sets w; = 0.5.
29
CU =wfo + (l-w) feos wie [0,0] (15)
Specification Sensitivity: Market Share (%)
80
5
O15 .
5 fy: Unresponsive
Soa dence To 100% Res
5 9° A
205/14
oS t .
Sf | 7 fyg: Responsive
So 6 78
0 0.50 1 1.50 2 0 12 24 36 48
Desired Production/Normal Capacity Time (Months)
(WS*/(K*CU,))
Figure 8 —- Market share sensitivity to capacity utilization specification.
Figure 8 shows market share for different specifications of capacity utilization function.
System variability increases with managers’ responsiveness to changes in desired production. It
appears intuitive that more responsiveness should enhance stability by preventing the
accumulation of excess inventory during periods of low demand. The opposite is observed,
however, because demand is endogenous: by cutting production aggressively when demand is
perceived to be low, the firm ensures that inventory will be even less available, driving market
share still lower. The unresponsive policy pushes product into the supply chain, improving
availability and bringing customers back to the firm more quickly.
5.3.2. Sensitivity to Customer Response
Now consider the two extreme cases of customer responses. An insensitive customer
base, characterized by function (f,7), does not respond to changes in the perceived service level.
The insensitive customer response function around the operating point (1,1) is flat. This extreme
case cuts the feedback from the supplier’s service level to customer demand. In contrast, a
sensitive customer base, characterized by function ( f42), responds aggressively to changes in
perceived service. The slope of the sensitive customer response function around the operating
point is steep. Sufficiently low perceived service levels can reduce product attractiveness to the
minimum possible level. A general customer response function is obtained from the linear
combination of the two polar cases (f47 and f42); in the base case w2 = 0.5.
CR = Wy fg, + (l- Wy) fags wre [0,1] (16)
Specification Sensitivity: Utilization (%)
109% Exogenous
f,¢: Insensitive asenal F 100
f,: Sensitive Customer 75
Product Attractiveness (AL)
°
a
50 .
“100% Endogenous
0.25 25
0 S (1)
0 0.25 0.50 0.75 1 0 8 16 24 32 40 48
Fraction of Orders Filled (PFoF) Time (Months)
Figure 9 — Utilization sensitivity to customer response specification.
Figure 9 shows the results for different degrees of customer response. System variability
increases as customers become more responsive to product availability. This result is expected.
When demand is exogenous the production loop operates independently. By adding the customer
response balancing loop, we have noted that oscillations in production increase. It is also sensible
to expect that a more sensitive customer base, reacting with a perception delay to inventory
availability, will introduce more variability in demand and consequently to production. Customer
response sensitivity sows that supply chain instability with exogenous demand is much smaller
than the instability with endogenous demand. That is, capturing the feedback of product
availability with customer demand amplifies supply chain instability. This result suggests that
models that adopt exogenous demand underestimate the instability in supply chains, which
undermines associated policy prescriptions.
In section 5.1, we learned that the assembly and FGI pull loops are capable of stabilizing
the system, but that only takes place when sufficient inventory is available. Therefore, an
important policy to improve system stability is to maintain safety stocks in both assembly and
FGI. The next section investigates the optimal levels of assembly and FGI safety stocks.
5.4. Optimal Safety Stock Location Analysis
The desired level of safety stock in assembly and finished goods is based on a control
heuristic that optimizes the trade off between the cost of holding inventory and the cost of lost
sales. The criterion to evaluate the optimal level of safety stocks is minimization of net present
value of cumulative discounted costs (CDC), with a discount rate ws
coc = [eFC (dt (17)
Total costs (TC) are the sum of total holding cost (HC7), given by the sum of the
assembly holding cost (HC,wip) and the finished goods holding cost (HCrg;), and lost sales cost
(LSc). For simplicity, we do not account for holding costs in fabrication.
TC = HC gyp + HC pg, + LS (18)
The holding cost at each stage is given by the product of the inventory volume (AWIP,
FGI) and the unit cost of maintaining the inventory. Unit inventory holding costs in each stage
are assumed to be a fraction (f, 6) of the unit finished goods cost (9).
HC yp = 8 -@- AWIP (19)
HC pq, = 5-0+ FGI (20)
* In practice, the integration limits (from 0 to 2) go from 0 to a finite (but large) end simulated time. End time is
selected ensuring that discounting has reduced the remaining contribution to NPV to a negligible amount.
Lost sales cost is the product of a factor (@) of unit finished goods cost (g) and the
amount of lost sales, given by the difference between the initial market share (MSo) and the
actual (MS,).
LS, =a: -(MS, -MS,) (21)
The fraction of inventory volume destined as safety stock in each stage is given by a
percentage (paw, prci) of the total volume in that stage (AWIP, FGI). The optimal safety stocks
are the values of the safety stock percentage at each stage that minimizes the net present value of
CDC over the simulation period. Demand is specified as the sum of the current level and an
autocorrelated (pink) noise term with a standard deviation (0) of 5% (representative of demand
variations for Intel).
We investigate the optimal volume of safety stocks for four different levels of lost sales
cost (@) and four ratios of unit inventory holding costs in assembly and finished goods (p."°
Figure 10 the results of the optimization runs for (a) the percentage of the total volume in
assembly (pawip) and (b) the percentage of the total volume in finished goods (prc). As
expected, optimal safety stocks in both stages increase with the lost sales cost (a). Likewise, the
allocation of safety stock between assembly and finished goods is highly dependent on the ratio
of holding costs in the two stages. Lower holding cost ratios benefit the allocation of safety stock
in assembly. In addition, for large enough holding cost ratios, the manufacturer holds more
safety stock in FGI as the cost of lost sales increases. This result makes intuitive sense as higher
safety stocks in FGI allows the company to meet demand with a shorter response time.
° The same random number seed is used for all simulations, allowing each optimization run to use exactly the same
realization of the random process.
'° We assume also that the parameters for unit costs of finished goods (@) = 50 $/unit, and the holding cost fraction
of @ associated with finished goods (6) = 0.01/month.
Assembly WIP FGI
24% 24%
+ Alfa =2
3 hs AAlfa = 15 =
£ i X Alfa = 05 8
3 20% S 20%
8 o cb >
FS Sx
5 zg
Py 5
= v z
= 16% weeeeeee Moe ae a 16% }---------------p---- g---B___.
z x g 2 a x
= x 8
8 & A x
é x
12% 12%
0 01 02 0.3 04 05 0.6 0 01 0.2 0.3 0.4 0.5 06
Holding Cost Ratio (8/5) Holding Cost Ratio (p76)
Figure 10 - Optimal percentage safety stocks in (a) Assembly WIP and (b) FGI.
Furthermore, as the throughput time in assembly (7) is much longer than in finished
goods (Zp), the same percentage of safety stock volume in assembly and finished goods will
translate into a higher safety stock coverage in assembly. For instance, as the throughput time in
assembly (%) equal four weeks compared to one week in finished goods (Zop), a 15% safety
stock in assembly and finished goods translates into a safety stock coverage of 0.6 weeks in
assembly and 0.15 weeks in finished goods. Hence, the same dollar investment in safety stocks
yield higher inventory coverage in assembly. This result, while intuitive, has a direct impact on
the market share that the company can retain and the resulting instability of the supply chain.
Comparing the impact of the same dollar amount in safety stocks, the higher inventory coverage
in assembly has a more stabilizing effect in the supply chain variability, which results in shorter
delays, fewer orders expedited, and more satisfied customers. While counter to the intuition that
safety stocks should be placed in FGI, due to the increased responsiveness to demand,
performance is improved by keeping the safety stock in assembly. Therefore, the heuristic of
maintaining additional safety stock in assembly can help stabilize the system operation and
improve service level. Figure 11 compares the performance of a 5% additional safety stock
policy in assembly WIP and FGI from the base case run.!!
Perceived Fraction of Orders Filled
1.00
5% Additional AWIP SS
0.95
5% Additional FGI SS
0.90
0 12 24 36 48
Time (Month)
Figure 11 - Performance comparison with equal safety stock investments in
AWIP and FGI.
The holding costs of inventory are easily measured, highly salient, and unambiguous. In
contrast, the costs of lost sales are hard to assess, and demand variability is easily explained
away as resulting from factors outside the firm. Thus it is likely that managers will underestimate
the costs of poor delivery performance (i.e. lost sales, reputation losses, customer inconvenience,
etc.). If the manufacturer underestimates the cost of lost sales, it will also underestimate the
optimal safety stocks to be maintained in assembly and finished goods. In addition, not only will
holding too little inventory lead to higher costs than necessary, but doing so alters the dynamics
of the system by reducing the stability of the supply chain. This instability has probably many
other costs that aren’t accounted for. Other operational costs include shorter run lengths and
smaller batch sizes, more frequent set-ups and changeovers, higher error and rework rates, and
more idle time between lots and between set-ups.
"To obtain a comparable dollar amount for the same amount of safety stocks in FGI and AWIP, we set the ratio of
holding costs (£/6) to 0.25, which compensates exactly for the inventory coverage ratio between the two stages.
6. Discussion and Directions for Future Research
This paper addressed the causes of oscillatory behavior in capacity utilization at a
semiconductor manufacturer and the role of endogenous customer demand in influencing the
company’s production and service level. The modeling effort drew on extensive field work,
including direct observation, collection of archival materials and data, and structured and semi-
structured interviews with managers at Intel. The paper contributes to our understanding of the
role that customer response has on increasing demand amplification across supply chains by
exploring the mechanisms through which endogenous customer demand interacts with managers
production heuristics. The interaction of the sales and production effects greatly amplifies the
oscillatory behavior of production. This result suggests that models that adopt exogenous
demand underestimate the instability in supply chains, which undermines associated policy
prescriptions. Our research shows that capturing the feedback of product availability with
customer demand amplifies supply chain instability, ultimately reversing traditional policy
prescriptions about inventory and capacity utilization policies. Simulation runs suggest that
typical policy prescriptions of lean inventory and responsive utilization policies hold only when
demand is assumed exogenous. With endogenous demand, inventory buffers and an
unresponsive capacity utilization policy yield higher service levels.
Loop knockout analysis suggests that the balancing FG/ Pull (B/) and Assembly Pull
loops (B3) stabilize the system. However, since they can only operate effectively when sufficient
inventory is available, the manufacturer should benefit from maintaining larger inventory buffers
at assembly and finished goods. While the heuristics of keeping inventory buffers at assembly
and finished goods for improving system robustness is not new, this research underscores their
importance in the operation of hybrid push-pull systems. In addition, the policy analysis supports
an amount of safety stocks that increases with lost sales cost and an allocation dependent on the
ratio of holding costs between assembly and finished goods. Therefore, the manufacturer can
effectively reduce supply chain instability and reduce the impact on lost sales, as long as it
internalizes the lost sales cost.
In general, semiconductor manufacturers, as well as firms in other industries, tend to keep
low inventory levels and run lean supply chains, allowing them to reduce inventory costs. This
practice presents manufacturers with a strategy to avoid costs associated with inventory
obsolescence in industries with short product life cycles. The mental model motivating lean
inventories assumes demand variability is exogenous: in a world with unpredictable demand
changes, costly FGI, and rapid technological obsolescence, keeping inventories lean minimizes
the risk that the firm will be caught with excess stock if demand unexpectedly declines.
However, demand is not exogenous—product availability affects demand, which then feeds back
to availability. Low finished and work-in-process inventories increase the chance of stockouts in
different stages in the supply chain, boosting the likelihood that the system will operate in an
undesirable mode (e.g., as a push system). Considering the typical inventory management
heuristics adopted by companies, like the constant adjustment of desired inventory levels to
reflect current demand signals, and the potential increase in demand variability introduced by
customer responses, we note that companies may underestimate the true costs associated with
stockouts. Moreover, managers’ heuristics of adjusting capacity utilization to respond to
variability in demand — caused by the supplier’s inability to satisfy customer — can amplify the
demand variability. Because demand is endogenous, by cutting production aggressively when
demand is perceived to be low, the firm ensures that inventory will be even less available,
driving market share still lower. Hence, the supplier’s effort to meet customer demand in the
short run may actually hurt customer service in the long run. In contrast, the unresponsive policy
pushes product into the supply chain, improving availability and bringing customers back to the
firm more quickly.
There are a number of opportunities for future research motivated by this study. First,
there are many other industries (e.g. automobiles, electronics), where the effects reported here
may also play an important role. Second, our study currently abstracts away from the
introduction of new products over time and the characteristic demand patterns over the product
lifecycle, where there are often initial shortages during production ramp up, followed by a
secular demand decline at the end of the product life. Optimal safety stock levels and production
heuristics may change over the course of the lifecycle. Moreover, our model incorporates only
the response of customers due to current service level (e.g. supply reliability). However, if
customers consistently experience poor delivery reliability they may choose to make other firms
their primary suppliers, reducing sales for other products and for future products, and also
increasing the variability of demand (Risch, Troyano-Bermudez, and Sterman 1995). Order
cancellations can also be added to the model. If order cancellations occur as a result of a decrease
in service level, they are likely to amplify the effects caused by lost sales, which would
strengthen the results presented here. In addition, the current model does not incorporate the
possible inflation of orders by customers, creating phantom demand or bubbles, when multiple
OEMs hedge against supply shortages (Goncalves 2003, Sterman 2000 ch. 18.3). While not
incorporated in the model, phantom demand is important since it is likely to balance the effects
of lost sales and counter the effects observed in this research. While not reported here, we
incorporated the assumption and conducted a number of simulations to investigate the impact on
the results. Our analysis concludes that for plausible values of inflationary ordering the main
results of this paper still hold. While this study does not incorporate these important assumptions
— order cancellation and order inflation — they have been addressed thoroughly by one of the
authors in another study (Gongalves 2003). The hope is that by separating the effects we can help
clarify their distinction and impacts to supply chains.
7. References
Anderson, E. and C. Fine. 1999. “Business Cycles and Productivity in Capital Equipment Supply
Chains.” In Tayur et al. Quantitative Models for Supply Chain Management. Kluwer
Academic Publishers, Norwell, MA.
Baganha, M. and M. Cohen. 1998. “The Stabilizing Effect of Inventory in Supply Chains.”
Operations Research. 46: S72-S83.
Chen, F., Z.Drezner, J.Ryan, and D.Simchi-Levi. 2000. “Quantifying the Bullwhip Effect in a
Simple Supply Chain: The Impact of Forecasting, Lead Times, and Information.”
Management Science, 46(3): 436-443.
Croson, R. and K. Donohue. 2000. “Behavioral Causes of the Bullwhip Effect and the Observed
Value of Inventory Information.” Wharton School of Business working paper, University
of Pennsylvania.
Dana, J. and N. Petruzzi. 2001. “Note: The Newsvendor Model with Endogenous Demand.”
Management Science. 47(11): 1488-1497.
Diehl, E. and J.D. Sterman. 1995. “Effects of Feedback Complexity on Dynamic Decision
Making.” Organizational Behavior and Human Decision Processes. 62(2): 198-215.
Forrester, J.W. 1958. “Industrial Dynamics — A Major Breakthrough for Decision Makers.”
Harvard Business Review. 36(4): 37-66.
Forrester, JW. 1961. Industrial Dynamics. Cambridge, MA: Productivity Press.
Gaither, C. 2001. “Intel Beats Forecast; Warns of Revenue Shortfall,” The New York Times,
January 17, Cl.
Gans, N. 1999a. “Customer learningand loyalty when quality is uncertain.” WorkingPaper,
OPIM Department, The Wharton School, University of Pennsylvania, Philadelphia, PA.
Gans, N. 1999b. “Customer loyalty and supply strategies for quality competition.”
WorkingPaper, OPIM Department, The Wharton School, University of Pennsylvania,
Philadelphia, PA.
Gongalves, P. 2003. “Demand Bubbles and Phantom Orders in Supply Chains.” Unpublished
Ph.D. Dissertation, Massachusetts Institute of Technology, Cambridge, MA.
Graham, A. 1977. “Principles on the Relationship Between Structure and Behavior of Dynamical
Systems.” Unpublished Ph.D. Dissertation, Massachusetts Institute of Technology,
Cambridge, MA.
Graves, S. 1999. “A Single-Item Inventory Model for a Non-Stationary Demand Process.”
Manufacturing & Service Operations Management. 1: 50-61.
Hachman, M. 1999. “MPU Supply Tightening.” EBN Online. Jan. 15.
http://www.ebnonline.com/showArticle.jhtml ?articleID=2902238
Hachman, M. 2000. “Despite inventory concerns, Intel sets 4Q fiscal records.” EBN Online. Jan.
13. http://www.ebnonline.com/business/opinion/showArticle.jhtml?articleID=2906438
40
Hall, J. and E. Porteus. 2000. “Customer Service Competition in Capacitated Systems.”
Manufacturing & Service Operations Management. 2 (2): 144-165
Hendricks, K.and V. Singhal. 2003. “The Effect of Supply Chain Glitches on Shareholder
Wealth.” Journal of Operations Management. 21:501-522.
Lee, H., Padmanabhan, V, and Seungjin Whang. 1997a. “Information Distortion in a Supply
Chain: The Bullwhip Effect.” Management Science. 43(4): 546-558.
Lee, H., Padmanabhan, V, and Seungjin Whang. 1997b. “The Bullwhip Effect in Supply
Chains.” Sloan Management review, Spring: 93-102.
Mass, N. 1975. “Economic Cycles: An Analysis of Underlying Causes.” Cambridge, Mass.
Wright-Allen Press.
Mitchell, TW. 1924. “Competitive Illusion as a Cause of Business Cycles.” Quarterly Journal of
Economics, 38(4):p. 631-652.
Morecroft, J.D.W. 1980. “A Systems Perspective on Material Requirements Planning.” Decision
Sciences. 14: 1-18.
Risch, J., L. Troyano-Bermiidez, and J. Sterman (1995). "Designing Corporate Strategy with
System Dynamics: A Case Study in the Pulp and Paper Industry." System Dynamics
Review 11(4): 249-274.
Singhal, V. and K. Hendricks. 2002. “How Supply Chain Glitches Torpedo Shareholder Value.”
Supply Chain Management Review. January/February. 18-33.
Sterman, J.D. 1989a. “Modeling Managerial Behavior: Misperceptions of Feedback in a
Dynamic Decision Making Experiment.” Management Science. 35(3): 321-339.
Sterman, J.D. 1989b. “Misperceptions of Feedback in Dynamic Decision making.”
Organizational Behavior and Human Decision Sciences. 43(3): 301-335.
Sterman, J.D. 2000. “Business Dynamics: Systems Thinking and Modeling for a Complex
World.” Chicago, IL, Irwin-McGraw Hill.
4l
Back to the
CC BY-NC-SA 4.0