DETERMINING OPTIMAL LOT SIZE

FOR A SEMICONDUCTOR BACK-END FACTORY

Juergen Potoradi and Gerald Winz, Ph.D.
Infineon Technologies
Munich, Germany
ABSTRACT
Modeling analysts at Infineon Technologies are
using a methodology that applies queuing theory
logistics laws and simulation to factory
performance analysis. These methods are being
applied at the Singapore and Malacca back-end
factories, where a major focus is on achieving
capacity increases with minimal equipment
additions.
This paper describes this technical methodology
and investigates an optimum lot size for the
Malacca factory based upon given throughput and
cycle time targets. The analysis provides a
recommended lot size of 6800 for the overall
production area, allowing the factory to maximize
throughput while still meeting overall factory cycle
time goals. The model indicates a potential 14%
increase in throughput by selecting the optimal lot
size.

KEY WORDS
Lot size, semiconductor manufacturing, queuing,
simulation, production logistics.

INTRODUCTION

Productivity
improvement
efforts
in
the
semiconductor industry have historically focused
on wafer fab operations. However, it has recently
been recognized that the so-called back-end
factories have great potential for improvement,
particularly in the area of production logistics.
Figure 1 shows the general material flow of
semiconductor manufacturing from wafer fab
through the elements of the back-end factory (preassembly through mark/scan/pack).
Wafer
Start

Wafer

The challenge for back-end productivity

improvement is to maximize equipment utilization
and achieve high capital efficiency while
maintaining cycle time targets. Cycle time is
important to meet product delivery dates and to get
a short feedback time about quality and yield to the
wafer fabs. Of course, high utilization and short
cycle time are two conflicting goals. Because high
utilization leads in most cases to high WIP, Littles
law tells us this will increase cycle times. Queuing
theory applications [Hopp and Spearman, 1996]
teach us that the cycle time depends not only on the
utilization, but also on variability and raw process
time (RPT). The latter is largely determined by the
lot size in back-end operations. Other studies have
also indicated that for production systems with
reentrant flow, one of the three most influential
factors on system performance is lot size [Adachi
et. al. 1989].
Determining an optimum lot size for the back-end
factory is not a straightforward exercise since the
optimal lot size in assembly is generally different
from the one in test. The material flow in preassembly and assembly is linear, much like the
linked lines in the automotive industry. Therefore,
these tend to have less variability. This is not true in
the test area where re-entrant flow significantly
increases non-uniformity of the arrival rates.
Additionally, the back-end contains areas with
stand-alone equipment and areas with dedicated
auto-lines (similar to the automobile manufacturing
industry). These inconsistencies in the material
flow, the high degree of product variety of most
logic-product back-ends, and the varying lot sizes
make production logistics difficult. The effective
coordination of tools, process, WIP and manpower
is critical to reducing variability.

Fab

Pre-Assembly

ANALYTIC APPROACH: LOT

SIZE INFLUENCE

2.1 CYCLE TIME AND LOT SIZE

Assembly

When looking at the influence of the lot size on a

given type of back-end equipment, you might
encounter a curve of the shape shown in Figure 2.

End Of Line
Burn In / Test
Mark Scan Pack

Product
Ship

Figure 1: A Simplified Semiconductor Production Flow

The utilization is determined by the mean of the

service time B and the inter-arrival time A of the
lots. Consequently we just need the following
formula for the mean inter-arrival time to get the lot
size dependency of . The mean inter-arrival time is
determined by the mean lot size and the arrival rate
based on units per unit of time, unit .

E[CT] (3.6K time units)

4
3.5
3
2.5
2
1.5
1
0.5

E[ A] =

0
100

1100

2100

3100

4100

5100

6100

7100

So we get for the utilization :

E[L] (units)

Figure 2: Cycle Time CT

In order to explain this curve, we consider the cycle

time formula for M/GI/1 systems [Tran-Gia, 1996].
There the mean cycle time CT is given as a function
of utilization , service time B and the coefficient
of variation of the service time cB 2 :

1 + cB 2
E[CT ] =
E[ B] + E[ B]
2 1
(For an English derivation of this formula, see
Hopp and Spearman [1996, pg. 295].) As a next
step, we have to look at the lot size dependency of
the three factors B, and cB2 .
The service time increases linearly for increasing
lot size L, according to the following formula. The
constant overhead part is denoted by C.

E[ B] = E[ L ] + C

For simplification, we assume that the time needed

to process one unit equals one time unit. Typical
values for the overhead C are 800 time units for
testers and 1500 time units for trim/form tools.
Figure 3 shows the service time curve:
3
2.5
2
1.5

E[ B ] unit ( E[ L ] + C )
=
E[ A]
E[ L ]

Based on this formula we get a condition for unit in

order to have a stable system:

unit <

E [ L]
E [ L] + C

The utilization/lot-size curve for different values of

C has a distinctive shape, as shown in Figure 4.
0.9
0.8
0.7
utilization

This curve shows the relationship of cycle time and

lot size for a trim/form tool, but is indicative of
many factory equipment types. What these types of
equipment all have in common is a certain time that
is needed for loading and unloading of a lot. That
means there is a fixed amount of time that has to be
spent for each lot, independent of the lot size. For
every type of equipment where the processing time
of a lot consists of a fixed time for overhead and a
time slice for each unit, the cycle time curve has
this askew U shape.

E[B] (3.6K time units)

E[ L]
unit

0.6
0.5
0.4
0.3
0.2
0.1
0
100

1100

2100

3100

4100

5100

6100

7100

E[L] (units)

Figure 4: Utilization versus Lot Size L

The curve in Figure 4 (C=1500, unit=0.211) shows
a overloaded system for lot sizes below 500. In this
case the condition for unit does not hold for lots
smaller than 500.
For this analysis we assume that a change of the
mean lot size has no influence on the distribution of
the arrival and service processes. Therefore, the
coefficient of variation cB 2 , as well as the
Markovian arrival process, is independent of the lot
size change. This might not be true in a real
manufacturing system, since large-size lots might
be treated differently from smaller-sized lots. For
example if an automated transportation trolley is
used always at full capacity for lot transport to a
tool, then the number of lots arriving at the same
time depends on the lot size. These kinds of
dependencies are considered in the simulation
models.

1
0.5
0
100

1100

2100

3100

4100

5100

6100

7100

E[L] (units)

Figure 3: Service Time B versus Lot Size L

To summarize, it can be said that the lot size has an

influence on the utilization and on the service time.
Accordingly the cycle time dependency on the lot
size is a function of both utilization and service
time dependency. This can easily be seen from the
curves in Figures 2, 3, and 4. The left part of the
curve in Figure 2 is influenced by the utilization.
2

The steep drop of the utilization is even amplified

in the cycle time formula. The right part of this
curve reflects the linear increase of the service time.
2.2 CHANGING UTILIZATION AND
VARIABILITY
The curve in Figure 2 shows clearly that 1300 is the
optimum lot size for the described tool. But what
will happen, if we start more material on the tool or
the variability of the service time increases?

2.3 THROUGHPUT, OPERATING CURVES

Since people on the production floor are very
throughput-oriented, the question might arise of
what lot size will give maximum throughput for a
given cycle time. In this case we use the same cycle
time equation to calculate, for different lot sizes, the
number of units that can be released in order to
achieve a certain cycle time target.
0.5
0.45

For increasing unit (from 0.1 to 0.6) the cycle time

curve shifts to the right upper corner (see Figure 5).
This happens under the influence of the utilization,
since the service time is not affected by a change in
unit .

(units per sec)

0.4

0.2
0.15
0.1
0

0.6

100

1100

2100

3100

4100

5100

6100

7100

E[L] (units)

8
7

0.5

6
5

0.4

0.3
0.2
0.1

3
2
1
0
0

1000

2000

3000

4000

5000

6000

7000

8000

E[L] (units)

Figure 5: Cycle Time versus lot size for different

arrival rates
In Figure 5 we can see the minimum lot size
increasing faster than the linearly-increasing arrival
rate unit .
Compared to the significant influence of equipment
utilization on the optimum lot size, a higher
variability of the service process has a rather
negligible effect on optimum lot size.
20
18

Figure 7: Arrival Rate that gives a cycle time of 3.5

hours for each lot size
Figure 7 shows the arrival rate unit for lot sizes
between 100 and 8000 and for a target cycle time of
3.5 hours (one hour equals 3600 time units). This
curve has an inverted U shape with an optimum
lot size at the maximum arrival rate unit . This is
also the maximum throughput since we consider
only stable systems. The explanation of the U
shape is again found in the steep utilization curve
for smaller lots and the increasing service time for
larger lots.
A summary of the various curves discussed thus far
is given in Figure 8, showing the relationship
between cycle time and arrival rate (throughput)
[Hopp and Spearman, 1996]. At Infineon
Technologies, this is referred to as the operating
curve. In Figure 8 the operating curves for five
selected lot sizes are shown.

16
14

lot size 500

12
3

10

lot size 1000

lot size 3000

lot size 5000

lot size 7000

14
12

1
0

4
2
0
0

1000

2000

3000

4000

5000

6000

7000

8000

E[L] (units)

Figure 6: Cycle Time for different service time

distributions
Figure 6 shows the cycle time curves for the
coefficient of variation cB 2 increasing from 0 to 4.
Based on this we can say that the optimum lot size
is much more sensitive on a utilization increase
than to a variability increase.

cycle time (hours)

E[CT] (3.6K time units)

0.3
0.25

0.05

10

E[CT] (3.6K time units)

0.35

10
8
6
4
2
0
0.1

0.2

0.3

0.4

0.5

0.6

0.7

units per sec

Figure 8: Operating Curves for different lot sizes

The curve shown in Figure 7 can be viewed as a
horizontal section at cycle time 3.5. Note that the
curve for lot size 3000 crosses the line furthest to
the right, showing again that for a cycle time of 3.5,
3000 is the lot size that gives maximum throughput.
The curve in Figure 2 can be viewed as a vertical
3

section at the arrival rate 0.21. To get the curves in

Figure 5, vertical sections for the correspondent
arrival rates have to be considered.

distributions were estimated [Law and Kelton,

1991].
3.3 VALIDATION

The application of queuing formulas is more

complex for systems with more than one piece of
equipment, several consecutive process steps,
variable down times, setups, and other additional
parameters. Nevertheless, these simple curves and
formulas are very useful to explain simulated cycletime curves of these complex systems. The
simulated curves for a multi-step system consist
more or less of superimposed curves of the single
steps.

SIMULATION APPLICATION:
LOT SIZE INFLUENCE

3.1 PROJECT APPROACH

This analysis is part of the long-term goal of
introducing simulation as a tool for management
decision support in the back-ends in Malacca
(Malaysia) and Singapore. The specific question
concerning the factory is: What is the optimum lot
size?
In order to answer this question for the Malacca
back-end, two different models where built: one for
the combined assembly/end-of-line areas and one
for the burn-in/test area. For the pre-assembly area,
lots arrive from the previous production department
in a predetermined size. The overhead time is
negligible in the mark/scan/pack area, so lot size is
not so important. Therefore these areas are not
considered in this analysis.
The models were built, validated, and used for
analysis by a team of process engineers from the
specific production areas, in partnership with a
simulation expert. This team structure, as well as
regular meetings with production people and
management, ensured that the simulation results
met expectations and requirements. Details of this
project management approach can be found in
Chance, Robinson, and Fowler [1996].
3.2 DATA
Data from the planning department was used as the
core input data for the simulation models. More
detail was added concerning scheduled and
unscheduled down times, process flow, and
variability. Since the static planning data uses just
average values for calculations, any data about
distributions (e.g., lot size, lot arrivals, cycle time)
had to be gathered from the manufacturing
execution system (MES) and other standard factory
reporting systems. Data about equipment states and
process times was gathered manually and was
restricted to the average values (more sophisticated
CIM online data collection systems are not installed
in this factory). For any missing data, triangular

For a successful simulation project, two conditions

are very important:

models must be valid.

the model and the output must be trusted.

The model validation was done by comparing

model output to actual historical data for a given
period. We used cycle time and OEE (Overall
Equipment Efficiency) output charts for the
comparison. By comparing these parameters we
ensured that the variability in our model was
comparable to the actual system. The cycle time is a
function of utilization and variability [Hopp and
Spearman, 1996]. Therefore, if we compare the
utilization (OEE chart) and the cycle time, we can
infer a match of the variability.
Proof that results from a simulation study are
trusted is evidenced only by the factory managers
using simulation output in their decision-making
process. Simulation for simulations sake is of no
value in a production environment. The team
structure and the regular meetings contributed
significantly to meeting this goal.
3.4 SOFTWARE
This project used the performance analysis software
Factory Explorer, from Wright Williams and
Kelly [Chance, 1996], which proved to be a very
effective tool for modeling back-end operations.
Building from previous modeling experiences
within Infineon Technologies [Domaschke, et. al.,
1998], the total time to train the factory analysts,
build valid models, and conduct the analysis was
less than three months.
3.5 RESULTS
The optimum lot size is dependent upon the
objective function of the simulation exercise.
Optimizing for the minimum cycle time, for
example, will give a different answer than
optimizing for maximum throughput. In order to
answer both questions (i.e., determining the
optimum lot size for a given cycle-time-constrained
capacity) the simulation output is shown as the
operating curves described in section 2.3. Cycletime-constrained capacity [Fowler and Robinson,
1995] is defined as the maximum throughput rate
sustainable for the factory for a given product mix,
line yield, and equipment set, and a constraint on
the average cycle time.

lot size 5.5K

lot size 7.5K

cycle time

lot size 3.5K

lot size 9.5K

+ 14%

Throughput

Figure 9: Operating Curves for different lot sizes;

assembly/end-of-line

The throughput could be increased 8.5% by

going from 3000 lot-size to 9500 lot-size and
will still meet the cycle time target. The
reasons for the high throughput jump from
3000 lot-size to 9500 lot-size, and the
relatively small changes for lots larger than
9500, are the same as in the assembly/end-ofline analysis. One reason is the increasing
flatness of the utilization curve for increasing
lot-sizes (see Figure 4 section 2.1). A second
reason is the change of the bottleneck at lotsizes around 5500 from a lot-size sensitive tool
to a tool that is not lot-size sensitive.

The lot size of 3500 is the minimum lot size for

the planned throughput. Static calculations
from the Factory Explorer analysis indicate
that a lot size of 3000 will lead to an unstable
model.
The minimum cycle time for the planned
throughput can be achieved with lot sizes
around 5500. Lots with 3500, 7500, and 9500
show a higher cycle time. However, this
minimum cycle time is only slightly lower than
the cycle time for a 7500 unit lot size, and the
latter has significantly higher throughput.
There is a potential for a 14% increase in
throughput by going from 3500 lot-size to 7500
lot-size and still meeting the cycle time target.
The 14% throughput gain is due to the
decreasing utilization for increasing lot sizes,
as discussed in section 2.1. The smaller
increase from 5500 lot-size to 7500 lot-size is
based on the flat utilization curve for the larger
lots. Additionally, when the lot size is
increased above 5500, the bottleneck tool
changes from a tool that is lot-size sensitive to
a tool that is not lot-size sensitive.
The throughput can not be further increased for
the given cycle time target, by increasing the
lot size to 9500 and higher. See also Figure 7 in
section 2.3, where the maximum throughput is
decreasing for lots bigger than the optimum.

Figure 10 shows operating curves for different lot

sizes for the burn-in/test area. Again we have a
horizontal line for the cycle time target and a
vertical line for the planned throughput. For the
burn-in/test area, we can say the following:

3000 is the minimum lot-size for the planned

throughput. The figure shows clearly, that any
smaller lot-size will lead to an unstable system.

The minimum cycle time along the vertical line

can be achieved with lots around 5500 and less.

lot size 5.5K

lot size 6.8K

cycle time

lot size 3k

In Figure 9 the operating curves for different lot

sizes for the assembly/end-of-line area are shown.
The horizontal line stands for the cycle time target,
the vertical line for the planned throughput. From
this analysis we can say the following:

lot size 9.5K

lot size 11.5K

+ 8.5%

Throughput

Figure 10: Operating Curves for different lot sizes;

burn-in/test

A further increase of the lot-size to 11500 and

above will not increase the throughput for the
given cycle time target. The maximum
throughput lies around 9500 lot-size. See also
Figure 7 in section 2.3, where the maximum
throughput is decreasing for lots bigger than
the optimum.

CONCLUSION

For this modeled factory, the optimum lot size is

7500 for the assembly area and 9500 for the burnin/test area. This allows the factory to maximize its
throughput in the certain area while still meeting
the established cycle time goal.
However, an additional operational consideration is
that the factory operates the MTX ovens on a fullbatch loading policy. The maximum batch-size for
these ovens is 6800, so for the burn-in area this
would be considered a locally-optimum size for
effective production operations.
Comparing the 6800 lot-size to the optimum lotsizes for assembly/end-of-line (7500) and burnin/test (9500), the simulation results show relatively
minor differences in the throughput. Therefore, for
the overall production area the recommended lotsize is 6800. This finding reinforces the fact that
simulation should be used to analyze not just local
improvements. It is also very important to consider
the overall impact on the factory as one entity.

This study reinforces the importance of applying

queuing theory and simulation together as one
unified approach. Used separately, the Operating
Curve Management (OCM) approach (using
queuing theory) simply indicated the benefit of
increasing the lot size at specific tools. Only after
applying simulation in partnership with OCM was
the team able to fix an upper bound on the optimum
lot size.
A general observation can be made concerning lot
sizes. For areas of the factory that are highly
utilized, a larger lot size is required to meet
throughput. For areas less utilized, a smaller lot size
can be implemented to minimize cycle time.
This analysis demonstrates that simulation models
can be rapidly built and effectively maintained to
provide quick updates and recommendations as the
factory changes. With part-time simulation
analysts, the Malacca factory is able to easily
provide continuous simulation results for
managerial decision-making.

ACKNOWLEDGMENTS
The authors gratefully acknowledge the technical
contribution of Dr. Jennifer Robinson, Chance &
Robinson, Inc., and the editorial assistance of Mr.
Steven Brown, Infineon Technologies. We thank
our many partners from the Malacca and Singapore
factories for their active participation.

REFERENCES

Hopp, W. J., and M.L. Spearman. 1996. Factory

physics:
foundation
of
manufacturing
management. Chicago: Irwin.
Law, A. M., and W.D. Kelton. 1991. Simulation
modeling & analysis. McGraw Hill.
Tran-Gia, P. 1996. Analytische Leistungsbewertung
verteilter Systeme. Berlin: Springer.

AUTHOR BIOGRAPHIES
JUERGEN POTORADI is a Factory Modeling
and Simulation analyst and project leader with
Infineon
Technologies
(formerly
Siemens
Semiconductor Division). He is currently
responsible for implementing simulation techniques
and methodologies in the Asian back-end factories:
Singapore and Malaysia. Mr. Potoradi received his
undergraduate and graduate degrees in Computer
Science from the University of Wuerzburg in
Germany. He has extensive experience in
simulation analysis of semiconductor wafer fab and
back-end production operations. His email address
is [email protected].
GERALD WINZ received his doctorate in
engineering from Fraunhofer Institute, Dortmund.
After logistics consulting for the German
production and trade industry, he joined Infineon
Technologies (then Siemens Semiconductor
Division) in 1997. Dr. Winz is in charge of OCM
introduction (Operating Curve Management, a
queuing theory application) in the Asian back-end
factories: Singapore and Malaysia. His email
address is [email protected].

Adachi, T., J.J. Talavage, and C.L. Moodie. 1989.

A rule-based control method for a multi-loop
production system. Artificial Intelligence in
Engineering, Vol. 4, No. 3, 115-125.
Chance, F. 1996. Factory Explorer users guide.
Chance, F., J.K. Robinson, and J.W. Fowler. 1996.
Supporting manufacturing with simulation:
model design, development, and deployment.
In Proceedings of the 1996 Winter Simulation
Conference, ed. J.M. Charnes, D.J. Morrice,
D.T. Brunner, and J.J. Swain, 114-121.
Institute of Electrical and Electronics
Engineers, Piscataway, New Jersey.
Domaschke, J., S. Brown, J.K. Robinson, and F.
Leibl. 1996. Effective implementation of cycle
time reduction strategies for semiconductor
back-end manufacturing. In Proceedings of
the 1998 Winter Simulation Conference, ed.
D.J. Medeiros, E.F. Watson, J.S. Carson, and
M.S. Manivannan, 985-992. Institute of
Electrical
and
Electronics
Engineers,
Piscataway, New Jersey.
Fowler, J.W. and J.K. Robinson. 1995.
Measurement
and
Improvement
of
Manufacturing Capacity (MIMAC) Designed
Experiment
Report,
SEMATECH
Technology Transfer #95062861A-TR.
6