You are on page 1of 6

Proceeding of the 2006 IEEE

International Conference on Automation Science and Engineering


Shanghai, China, October 7-10, 2006

Multi-factory capacity planning in semiconductor assembly


and test manufacturing with multiple-chip products
Sean Sweat Sophia Niu Mike Tao Zhang Zhicong Zhang, Li Zheng
Civil Eng Dept, Operations Research & Solutions, AzFSM Industrial Eng, Industrial Eng Dept,
MIT, USA Intel Shanghai, China Intel Corporation, USA Tsinghua Univ, China

Abstract – Capacity planning determines the optimal is difficult due to complicated product mixes, complex
product mix based on the available tool sets and allocates routing, and various resource requirements. Spreadsheets
production capacity according to the forecasted demands (Occhino 2000, etc.), simulation (Chou and Everton 1996,
for the next few months. MaxIt is the previous capacity etc.), and mathematical programming (Karabuk and Wu
planning system for Intel’s Flash Product Group (FPG) 2002, Uribe et al. 2003, etc.) are the most widely used
Assembly & Test Manufacturing (ATM). It only applied to approaches to capacity planning in semiconductor manu-
single product family scenarios with simple process rout- facturing. Linear programming (LP) or mixed-integer lin-
ing. However, new Celluar Handhold Group (CHG) ear programming (MILP) has an advantage over other
products need to go through flexible and reentrant ATM methods in that it is able to find the optimal solution to an
routes. In this paper, we introduce MaxItPlus, which is an elaborate capacity model with thousands, even millions, of
enhanced MaxIt using MILP (Mixed Integer Linear Pro- constraints in a reasonable amount of time.
gramming) to conduct capacity planning of multiple
product families with mixed process routes in a multi-
factory ATM environment. We also present the detailed
mathematical formulation, the system architecture, and
implementation results. The project will help Intel global
Flash ATM to achieve a single and efficient capcity
palnning process for all FPG and CHG products and gain
$10M in marginal profit (as determined by the finance
department).

Index Terms – capacity planning, semiconductor manu- Fig 1: Production routes for FPG and CHG products.
facturing, resource allocation, MILP (Mixed Integer Lin-
ear Programming) We have studied mid-range capacity planning in both
semiconductor assembly and test manufacturing at the VF
INTRODUCTION (virtual factory) level with multiple factories (Zhang et al.
Semiconductor manufacturing consists of Front-End 2005). The resultant system, MaxIt v1.2, employed an
processing (wafer fabrication and wafer sort) and Back- MIP technique to allocate the capacity, i.e. to arrange the
End processing (Assembly and Test), normally performed product mix in the VF environment in such a way to op-
separately in different factories. Flash memory chip was timize the prioritized objectives within resource
first developed by Intel in 1988. Flash is a rewritable limitations. We made the following assumptions in MaxIt
memory chip that holds its content without power, making v1.2: (1) MaxIt v1.2 performed capacity planning with
it a great solution for products like cell phones, digital only a single product family involved, i.e. the process
cameras, etc. CHG is a new product family for Pocket PCs flow is unique. (2) Although the assembly cycle time and
or smart cellphones. It applies SCSP (Stacked Chip Scale the test&FPO (sometimes just referred to as ‘test’) cycle
Package) technology, which assembles several Flash dies, time of each product in product family ‘PA’ is different,
Logic dies, and SRAM (Static Random Access Memory) MaxIt v1.2 assumed they were both fixed to one week and
dies into one package. Fig 1 demonstrates the production that the products assembled in this week could only be
routes of FPG (non-flex rounting) products and CHG (flex tested in the next week or later, as is called ‘Fixed Cycle
rounting) products. Time Assumption’. (3) All of the workload allocated to
The fundamental issue pertaining to capacity plan- assembly (or test&FPO) during a week can be completed
ning is the determination of an optimal product mix based within that week – that is, all of the assembly (or
on the available tool sets or allocation of production ca- test&FPO) operations of each dispatched product can be
pacity. Capacity planning in semiconductor manufacturing completely processed during any week. This we call
‘Uniform Workload Assumption’. (4) The cross-shipping


Given the sensitive and proprietary nature of the semiconductor industry, only normalized performance data and concep-
tual modeling are discussed in this paper.

1-4244-0311-1/06/$20.00 ©2006 IEEE 247


load Assumption’. (4) The cross-shipping time between the uniformity assumption and can generate compact LP
two production factories was ignored. formulations. Hung and Cheng (2002) relaxed the uni-
In this paper, we extend MaxIt v1.2 to a more ad- form assumption and proposed a partition approach which
vanced capacity planning system: MaxItPlus. MaxItPlus is more realistic for industrial applications. The number
is capable of planning with multiple product families, each of periods is large in these formulations, and results in a
with different process flows. We also propose Accumu- great number of decision variables and constraints, hence
lated Workload Formulation to break the Fixed Cycle creating very complex LPs.
Time Assumption, i.e. the cycle time is not predetermined In order to eliminate the assumption that the cycle time
to be fixed. Furthermore, it considers the cross-shipping of each operation (for each product) is an integer multiple
time between two production factories and the lead time of the planning period length, Leachman (1993) formu-
of subcontract work. lated the discrete lot-based production with a continuous-
time rate-based model (Hackman and Leachman 1989).
LITERATURE REVIEW The rate-based model assumes that the dispatching of each
product in each period is uniformly distributed over the
Problem formulation, i.e. transformation of a capacity
period. Then, given the estimated cycle time of each op-
planning problem into an LP/MIP model, is essential for
eration, the workload of each operation in the process
applying LP/MIP techniques. Some methods interpret the
flow can be calculated in terms of the dispatching rate of
planning horizon as a single period. IBM (Bermon and
each product type in the previous periods. This model
Hood 1999) developed an LP-based strategic capacity
allows the cycle time of each operation to be a non-
optimization planning system (CAPS). It is able to deter-
integer. However, it also assumes a fixed or average cycle
mine the optimal product mix that maximizes profits with
time for every operation, as does most other currently
tool capacity constraints and works as a decision support
published research.
system for capacity planning at IBM’s largest semicon-
Most of the previous studies on capacity planning in
ductor manufacturing line. Uribe et al. (2003) proposed a
semiconductor manufacturing focus on the front-end proc-
modeling technique for minimizing the investment in ca-
ess, especially wafer fabrication, while few concentrate on
pacity planning for discrete manufacturing sites under an
back-end processing. Wang and Hou (2003) solved the
uncertain demand stream. The mathematical model is a
problem of capacity expansion and allocation in the semi-
two-stage stochastic integer program, where the first stage
conductor testing industry with a limited budget. They
makes the optimal response of the system under uncer-
took into account the simultaneous usage of tester and
tainty and the second stage selects a tool set based on the
handler resources, assuming that an order is fixed to a
first stage with budget constraints.
specific tester and handler during the entire planning hori-
Other methods divide the planning horizon into multi-
zon.
ple periods. In multi-period planning, formulating the
‘operation balance’ constraint, i.e. consistency of work-
PROBLEM STATEMENT
load on different operations across multiple periods, in
accordance with the reality is difficult but critical. Con- MaxItPlus performs capacity planning of assembly and
ventional capacity models assume that all operations of a test processes in a VF environment which includes two
product can be completed within a period, or perform op- production factories, SH (Shanghai, China) and CV
timization at the step level not the operation level. Swa- (Cavite, Philippines), and subcontractors as well. The VF
minathan (2000) developed several models to find tool produces a variety of product types, classified into three
procurement policies capable of dealing with the uncer- product families: PA (Flash), PB (Digital Device) and PC
tainty in future demands for various products concerning (CHG). Products in PA and PB are assembled from raw
single-period or multi-period planning, and utilization materials. Products in PC are made from raw materials
measure or cost measure. Christie and Wu (2002) pre- and products belonging to PA and PB. That is, PA and
sented a multi-stage stochastic LP model for strategic ca- PB products are divided into two groups: one that is de-
pacity planning, aiming to minimize the gap between ca- livered to customers and another that is used to create PC
pacity demands and capacity allocation. Hood et al. products and thus reenters the assembly line. The PA and
(2003) employed stochastic integer programming to de- PB products that only go through the assembly line once
termine a tool set configuration that adapts to demand (thus creating finalized PA and PB products) go through
uncertainty. Çatay et al. (2003) examined the problem of what is called ‘straight routing’. The PA and PB products
planning wafer production over multiple time periods by a that go back into the assembly line to create PC products
multi-period MIP model. go through ‘reentrant routing’. Hence, the overall process
Leachman and Carmon (1992) divided the planning flow is mixed or flexible routing. In the past, as shown in
horizon into many periods and assumed that the average Fig. 2, MaxIt v1.2 could only perform capacity planning
flow time of each operation for each product was an inte- for one product family, PA, while PB and PC had to be
ger-multiple of period lengths. They proposed a Work- planned manually. MaxItPlus can cover all of the product
load Allocation Formulation and Direct Product Mix families.
Formulation to model alternative machine capacity prob- SH, CV, and subcontractors cooperate to fulfill cus-
lems in semiconductor manufacturing, which are based on tomers’ requirements. Both SH and CV are qualified to

248
assemble and test all product types, but subcontractors can We focus on the capacity allocation in the VF environ-
only perform assembly. The raw materials are dispatched ment and determine the optimal product mix which mini-
in SH, CV, or subcontractors. If the two factories’ assem- mizes the total cost. The capacity model of MaxItPlus is
bly capacities cannot meet requirements, the subcontrac- based on mixed-integer programming.
tors may assume some assembly tasks and ship the semi-
finished products to SH or CV for testing. There is an MODELING
inventory between assembly and test&FPO in each fac-
How to ensure the consistency of workload between
tory. These two inventories are shared virtually, meaning
consecutive operations within a period or between two
that semi-finished products can be shipped from one fac-
adjacent periods, and how to depict the workload allo-
tory to the other for test after assembly if needed. How-
cated to an operation, is stressed in capacity planning
ever, if assembly and test can be performed in the same
models. These models are concerned with the cycle time
factory, cross-factory shipping is not preferred because of
of assembly or test&FPO processes. The Fixed Cycle
increased production costs and the resultant lengthening
Time Assumption and Uniform Workload Assumption
of the cycle time. Fig 3 illustrates the overall process
used in MaxIt v1.2 are not in accordance with real produc-
flow.
tion for the following reasons: (1) The cycle times of as-
sembly and test&FPO, especially for product families PB
and PC, are not fixed to one week. Some are much longer
or shorter than one week. Moreover, in practice, the cycle
time varies under different production environments. (2)
Not all the assembly (or test&FPO) operations of each
dispatched product can be completely processed during
the week. For example, the work released to assembly
line on Thursday or Friday may not be finished assem-
bling within this week. In addition, dispatching is not
uniform during the whole planning horizon in most cases,
so the production volume of each product on each assem-
bly (or test&FPO) operation is different in some weeks.
(3) Products may be assembled and tested within the same
week.
Fig 2: MaxIt v1.2 model for PA, PB and PC In order to make the planning closer to practical pro-
duction, we propose the Accumulated Workload Formula-
We have the following assumptions, verified by real tion to eliminate the Fixed Cycle Time Assumption and
practice: (1) The model uses weekly time buckets and Uniform Workload Assumption. We also formulate the
weekly demand volumes for each product. (2) The reentrant processing routing, cross-shipping time, and
throughput of a machine is indicated by the volume this subcontract lead time in this model.
machine can produce in one week. The value has already Each of the following has a variable associated with it
considered the average of exceptions, maintenances, con- which MaxItPlus’s goal is to minimize: (1) material proc-
version time and rework consumption. (3) The yield is the essed by subcon; (2) “excess” inventory (as defined by
average ouput percentage of input, estimated during real user); (3) unsatisfied customer demand; (4) material
production. (4) In industrial practice, the key bottleneck in shipped between sites; (5) resources (machines and tools)
production is tool capacity rather than materials supply. used; (6) inconsistency of Intel site loading; (7) inconsis-
Therefore, in the capacity planning stage, we concentrate tency of subcon site loading; (8) unsatisfied experimental
on production capacity and assume materials are always engineering time requests; and (9) the total of all changes
available. (5) The finished products are delivered to cus- in weekly Tester Utilization.
tomers as soon as they are completely processed. In other Finally, we will present the mathematical formulation
words, there is no final product inventory. of MaxItPlus. The actual MIP was programmed in OPL
Suite’s ILOG program. In order to easily understand the
formulation, however, we must first introduce the vari-
ables that we used. The subscripts are defined first:

1. Subscript Parameters
pen penalty type
p product (which also changes for each stage)
pIN product (as input)
pOUT product (as output)
s site
w week
Fig 3 CHG production process in Intel global ATM m month
o operation
rs resource set (possible resource combinations)

249
r resource The following equations are components of the objec-
e enablers
tive function – for all of which the goal is to minimize.
Each equation is also accompanied by a specific, user-
2. User Defined Variables defined, penalty variable. This has not been shown here
Weight r tester weight (ratio respective to a particular
“reference tester”) but should be noted.
PenV pen,m penalty value (no units – wholly arbitrary) 4.1 Minimize Total Subcon Loading
PP p,m product priority (no units – wholly arbitrary) where s  subcon, and m is the month containing week w
surplusWeight an arbitrary value assigned to the permissibility
of surplus inventory over the CTO
obj_subconLoadLevel >
¦ ¦ ¦ input p,s,w * PPp,m @
p s w
BPRequest p,m customer demand (in units)
Commit p,m committed volume: units which have already The goal here is to delegate as little as possible to subcontractors.
been committed to production (in units) 4.2 Minimize Total Excess Inventory
LoadPerc p s,m load percentage: target percentage of a product’s produc- where w is the last week of month m
tion that is carried out at site s (ratio) obj_CTOstatus
CTOGoal p,m CTO goal (in units)
ShipTime s1,s2 ship time (in weeks) ¦ ¦ PPp,m ª¬ surplusWeight * surplusInv p , m  slackInv p ,m º¼
p  m
ResRatio rs,r resource ratio (number of resource r in resource
The first priority is bringing inventory up to the CTO goal, but once that
set rs)
is accomplished, we want to accumulate as little inventory over the CTO
EnablerRatio r,e enabler ratio (number of enabler e in resource r)
(excess) as possible. This prioritization is accomplished by the user
POR p,s plan of record (a reference to a particular week)
setting surplusWeight to something between 0 and 1. The lower the
BOH p,s beginning on hand: inventory (in units)
value of surplusWeight, the more likely we are to allow surplus inven-
Inv s,r,w inventory of resources (in units)
tory.
NomWeek p,s,w nominal week, or portion of week which is avail-
4.3 Minimize Total Unmet Demand
able (ratio)
LC p,s,w portion of full capacity available due to unfamili- obj_unmetDemand >
¦ ¦ slackDemand p,m * PPp,m @
p m
arity (ratio)
Yield p,s,w portion of input that successfully becomes output 4.4 Minimize Total Cross-shipped Materials
(ratio) where s1 z s 2
RampMIN p,s,w minimum production level allowed (in units)
RampMAX p,s,w maximum production level allowed (in units) obj_amountXshipped ¦ ¦ ¦ ¦ xShip p,s1,s2 ,w
p s1 s 2 w
InputRatio p,pIN number of pIN required to make one p (in units)
OutputRatio p0,s,w,p portion of p0 that becomes p (ratio) 4.5 Minimize Total Assembly Equipment Used
TPT p,s,m throughput time (in weeks) where r  APOeq
EETimeReq p,s,r,w request for experimental engineering time with p
obj_APOeqUsed ¦ ¦ ¦ resUsed s,r,w
and r (in hours) s r w
StartWeek p,s,o,rs first week that this scenario is “qualified” (a 4.6 Minimize Total Testers Used
reference to a particular week)
EndWeek p,s,o,rs last week that this scenario is “qualified” (a ref-
where r  testers
erence to a particular week) obj_testersUsed ¦ ¦ ¦ resUsed s,r,w
Runrate p,s,o,rs,w production speed (in units per machine per week) s r w
TU hueristic value of Tester Utilization 4.7 Minimize Total FPO Equipment Used
where r  FPOeq
3. Decision Variables obj_FPOeqUsed ¦ ¦ ¦ resUsed s,r,w
input p,s,w input to a PPM (in units) s r w
inputSplit p,s,o,rs,w allocation to different resource sets within an 4.8 Minimize Total Enablers Used
operation (in units) obj_enablersUsed ¦ ¦ ¦ enablersUsed s,e,w
output p,s,w output from a PPM pre-binsplit (in units) s e w
ship p,s,w weekly shipments to customers (in units) 4.9 Minimize Total Intel Loading Inconsistency
ship p,m monthly shipments to customers (in units) where s  Intel
slackDemand p,m unmet demand (in units) obj_IntelLoadSmooth ¦ ¦ ¦ diffLoad p,s,w
toInv p,s,w post-binsplit output – all output goes to inventory p s w
before any other destination (in units)
4.10 Minimize Total Subcon Loading Inconsistency
fromInv p,s,w volume that comes from an inventory (in units)
slackInv p,s,w slack of inventory in relation to the CTO goal (in where s  subcon
units) obj_subconLoadSmooth ¦ ¦ ¦ diffLoad p,s,w
surplusInv p,s,w surplus of inventory to the CTO goal (in units) p s w
inventory p,s,w inventory level (in units) 4.11 Minimize Total Unsupported EE Time
resUsed s,r,w number of resources used (in units) obj_unmetEETime ¦ ¦ ¦ unmetEETime p,s,w
slackEETime p,s,r,w unmet EETime (in hours) p s w
xShip p,s1,s2,w cross-shipped product – product shipped between A certain amount of experimental engineering time is requested each
sites (in units week. We minimize the requests not met.
enablersUsed s,e,w number of enabler components used (in units) 4.12 Minimize Total Tester Utilization Inconsistency
eeTime p,s,r,w amount of experimental engineering time af-
obj_TUSmooth ¦ ¦ diffTU s,w
forded (in hours) s w
diffLoad p,s,w the change in allocation from the previous week
(in units)
diffTU s,w the change in Tester Utilization from the previ- 5. Constraints
ous week (unitless) 5.1 Material Flow Constraints
5.1.1 Throughput
4. Objective Function

250
p, s, w, where w f w  TPTp,s,m , resUsed s,r,w
and where m is the month containing w ª inputSplit p,s,o,rs,w * ResRatiors,r T
output p,s,w f input p,s,w * Yield p,s,w ¦¦ ¦ ¦ «
¬ p,s,w * Runrate p,s,o,rs,w * LC p,s,w
p o rs w « NomWeek
This defines the product output for a site and week. (Note that the prod-
uct differs between stages, so to specify the product also specifies the
stage of the process which produced said output.) 
eeTime º» p,s,r,w

5.1.2 The Bin Split 168 * NomWeek »¼



p, s, w , toInv p,s,w output p0 ,s,w * OutputRatio p0 ,s,w,p p,s,w

his complicated constraint determines what resources are used when. It


Some products are split into what are basically sub-products, labeled as takes into account the split input, the number of a particular resource
“bins”. These binned products are the final output that then goes to an used in the applicable resource set, the nominal week, runrate, and learn-
inventory. ing curve. It also considers the resources used for satisfying EE time.
5.1.3 Product Bundling 5.2.3 Resource Limitation
fromInv pIN,s,w
p, s, w, pIN , input p,s,w s, r, w, where r  kits , resUsed s,r,w d Invs,r,w
InputRatio p,pIN
5.2.4 Enablers Used
The creation of a certain product may require more than one of a particu- s, e, w, and where r  kits
lar previous product. This constraint accounts for that.
5.1.4 Ramping enablersUsed s,e,w >
¦ resUsed s,r,w * EnablerRatior,e @
r
p, s, w , RampMIN p,s,w d output p,s,w d RampMAX p,s,w Here we calculate the enablers we are using as determined by the re-
When there are production (ramping) guidelines as decided by manage- sources we are using.
ment, this constraint considers them. 5.2.5 Enabler Limitation
5.1.5 Inventory Movement s, e, w , enablersUs ed s,e,w d Invs,e,w
p, s, and w t 2, 5.2.6 EETime on Resources
where w 0 ShipTimes ,s p, s, r, w , eeTime
0 p,s,r,w
rs ¬
¦ ª eeTime p,s,rs,w * ResRatiors,r º ¼
­ w 2, BOHp,s 5.2.7 EETime Status
°
and where PrevWeekInv ®
°¯w t 3, inventoryp,s,w -1 p, s, r, w , eeTimep,s,r,w EETimeReq p,s,r,w - slackEETimep,s,r,w
5.2.8 Loading Difference
inventoryp,s,w Pr evWeekInv  toInvp,s,w-1 - fromInvp,s,w-1 p, s, w such that w -1 t POR p,s ,
§ · § ·
¨ s ¸ ¨ s ¸

 ¨ ¦ xShipp,so ,s,w0 ¸ - ¨ ¦ xShipp,s,sd ,w ¸ - ship p,s,w diffLoad p,s,w input - input
p,s,w p,s,w 1
© o ¹ © d ¹
This constraint calculates the loading difference between two successive
This complex constraint considers the addition to amd subtraction from weeks per site/stage and product. We would like the input each week to
inventories in accordance with stage outputs, stage loading, and cross- be as similar as possible, thereby creating diffLoads near zero.
shipping. The cross-shipping is not considered instantaneous and atten- 5.2.9 Tester Utilization Constraint
tion should be paid to the week specifications and the boundary condi-
tions. s1, s2 , w, such that s1 z s2 and where r  testers
5.1.6 Demand Status TU d
p, m and where w is a week of month m ª § · º

« resUsed s ,r,w  ¨ ¦ eeTime p,s ,r,w ¸ »
BPRequest p,m ¦ ¦ ship  slackDemand
s w
p,s,w p,m «
«
1

0.001 
¨ p
©
Invs1,r,w
1


¸
¹ »
»
This is simply how much of the monthly demand was and was not met.
¦ Weightr «« »
» d TU
5.1.7 CTO Status § ·»
p, m and where w is a week of month m
r «

« resUsed s 2 ,r,w ¨ ¸
 ¦ eeTime p,s 2 ,r,w »
¨ p ¸
«- © ¹»
CTOGoal p,m «
¬
0.001
 Invs 2 ,r,w »
¼

§ · This constraint ensures that the actual Tester Utilizations are between

¨¨ ¦ ¦ inventory p,s,w ¸¸  slackInv p,m - surplusInv p,m the user-defined TU limits.
© s w ¹
5.2.10 Tester Utilization Weekly Difference
We keep track of how much a particular inventory is above or below the
user-defined CTO goal. s; w t 2, and where r  testers
5.2 Resource Allocation Constraints diffTU s,w
5.2.1 Product Splits (or Operation Splits)
ª § · º
p, s, o, w , input p,s,w ¦ inputSplit p,s,o,rs,w
« resUsed s,r,w  ¨ ¦ eeTime p,s,r,w ¸
¨ ¸
»
rs « © p ¹ »
« »
For some operation within a stage, there may be multiple resource sets 0 .001  Inv
that are capable of doing that operation. This constraint figures out how ¦ Weightr «« s,r,w »
»
the input to each operation will be split amongst the various possible « § ·»
resource sets. The way that the operation subscript is used here is unusual
r
«
resUsed s,r,w-1  ¨ ¦
¨ p

eeTime ¸
p,s,r,w-1 ¸ »
«- © ¹»
because we calculate this constraint for each instance of o but do not see « »
¬ 0 .001  Inv s,r,w-1 ¼
it in the non-summation variable. This is because the input to each opera-
tion is the same as the input to the entire stage (at least here at the weekly The absolute differences in TUs from week-to-week are calculated
level – this would not hold true for a greater level of detail in terms of here. LP techniques for dealing with absolute values are used to
time). convey this constraint in the actual program.
5.2.2 Resources Used
s, r, w

251
IMPLEMENTATION REFERENCE
By implementing the MaxItPlus system, $10M+ mar- [1] Mike Tao Zhang, Sophia Niu, Marky Mai, Qi Li, Zhicong Zhang,
and Li Zheng. "Multi-factory optimization enables kit reconfigura-
ginal profit will be gained by the material flow and re-
tion in semiconductor manufacturing," IEEE Int’l Conference on
source/kits optimization. The assumption to achieve bene- Automation Science and Engineering, Edmonton (Canada), 2005.
fit is that the optimization will provie certain % additional [2] Hackman, S. T., Leachman, R. C., 1989, A general framework for
capacity vs. manual capacity allocation, if those capacity modeling production. Management Science, 35, (4), 478-495.
could support same % more volume, then the marginal [3] Leachman, R.C., Modeling techniques for automated production
planning in the semiconductor industry. In: Ciriani, T.A., Leach-
profit for the more volumes would be huge. man, R.C. (editors), Optimization in industry, Wiley, Chichester,
Besides the marginal profit, four major benefits can be pp.1-30. 1993.
achieved as well from the implementation in our expecta- [4] Bard, J. F., Srinivasan, K., Tirupati, D., 1999, An optimization
tion: approach to capacity expansion in semiconductor manufacturing
facilities. Intl. Journal of Production Research, 37, (15), 3359-3382.
1. Avoid potential business breakpoint of current man- [5] Bermon, S. and Hood, S. J., 1999, Capacity optimization planning
system (CAPS). Interfaces, 29(5), 31-50.
ual process if demand volume or complexity in- [6] Bhatnagar, S., Fernandez-Gaucherand, E., Fu, M. C., He, Y., Mar-
creases cus, S. I., 1999, A markov decision process model for capacity ex-
2. Data Quality – Reduce human error in BP response pansion and allocation. Proceedings of the 38th Conference on De-
by reducing manual works. cision & Control, Phoenix, Arizona, USA, pp. 1380-1385.
[7] Cantamessa, M. and Valentini, C., 2000, Planning and managing
3. Business Process consistency – Achieve consistent manufacturing capacity when demand is subject to diffusion ef-
BP response biz process for all the Flash products fects. International Journal Production Economics, 66, 227-240.
4. Added additional functionality on assembly kits ca- [8] Çatay, B., Erengüç, S. S., Vakharia, A. J., 2003, Tool capacity
pacity analysis planning in semiconductor manufacturing. Computers & Opera-
tions Research, 30, 1349-1366.
We started the MaxItPlus project in May 2005 and we [9] Chou, W. and Everton, J., 1996, Capacity planning for development
successfully developed a prototype in Auguest 2005. The wafer fab expansion. Proceedings of the 1996 IEEE/SEMI Ad-
vanced Semiconductor Manufacturing Conference, Cambridge,
prototype employed the MILP model and demonstrated MA, pp. 17-22.
the feasibility and effectiveness of the model. MaxItPlus [10] Christie, R. M.E. and Wu, S. D., 2002, Semiconductor capacity
is on the right track be released as scheduled in March planning: stochastic modeling and computational studies. IIE
2007. To avoid technical risk and shorten development Transactions, 34, 131-143.
[11] Hood, S. J., Bermon, S., Barahona F., 2003, Capacity planning
time, we apply a new methodology that separates near under demand uncertainty for semiconductor manufacturing. IEEE
200 features to 6 releases, and conduct the functional tests Transactions on Semiconductor Manufacturing, 16(2), 273-280.
and integration tests right after each release. [12] Hung, Y.-F. and Cheng, G.-J., 2002, Hybrid capacity modeling for
alternative machine types in linear programming production plan-
ning. IIE Transactions, 34, 157-165.
CONCLUSIONS [13] Iwata, Y., Taji, K., Tamura, H., 2003, Multi-objective capacity for
In this paper, we describe MaxItPlus, which is an agile semiconductor manufacturing. Production Planning & Con-
trol, 14(3), 244-254.
MILP solver incorporating the changes and additions nec- [14] Karabuk, S. and Wu, S. D., 2002, Decentralizing semiconductor
essary for optimizing both the FPG and CHG product capacity planning via internal market coordination. IIE Transac-
families. Beyond that, a flexible data structure has been tions, 34, 743-759.
introduced to eliminate the hardcoding of the stage-to- [15] Leachman, R. C. and Carmon, T. F., 1992, On capacity modeling
for production planning with alternative machine types. IIE Trans-
stage routing and to enable complicated mapping among actions, 24(4), 62-72.
materials, products, and bins. Other objectives were [16] Occhino, T. J., 2000, Capacity planning model: the important in-
added as well, to better reflect Intel’s business rules. puts, formulas, and benefits. IEEE/SEMI Advanced Semiconductor
These included a minimum inventory goal, and splitting Manufacturing Conference, pp. 455-458.
[17] Robinson, J. K. and Giglio, R., 1999, Capacity planning for semi-
the tester utilization goal to be specific to memory testers conductor wafer fabrication with time constraints between opera-
and logic testers. The hope is that this flexibility might tions. Proceedings of the 1999 Winter Simulation Conference, Ari-
enable future product additions to simply require adjust- zona, USA, pp. 880-887.
ments in data preperation. The prototype system verified [18] Swaminathan, J. M., 2000, Tool capacity planning for semiconduc-
tor fabrication facilities under demand uncertainty. European Jour-
the MILP model, and we are confident that MaxItPlus to nal of Operational Research, 120, 545-558.
be delivered by September 2006 will bring Intel the return [19] Uribe, A. M., Cochran, J. K., Shunk, D. L., 2003, Two-stage simu-
on investment for around US $11 million and will de- lation optimization for agile manufacturing capacity planning. In-
crease the time spent creating the monthly mulit-factory ternational Journal of Production Research, 41(6), 1181-1197.
[20] Wang, K.-J. and Hou, T. C., 2003, Modelling and resolving the
production plans by two or three days. joint problem of capacity expansion and allocation with multiple
resources and a limited budget in the semiconductor testing indus-
ACKNOWLEDGMENTS try. International Journal of Production Research, 41(14), 3217-
3235.
The project is funded in part by the Intel Higher Edu-
cation Program, Intel Products (Shanghai) Ltd., and the
National Science Foundation of China (Grant
No.50375082). Their support is very much appreciated.

252

You might also like