Implementation of a Fluctuation Smoothing Production Control Policy - - PowerPoint PPT Presentation

▶
implementation of a fluctuation smoothing
SMART_READER_LITE
LIVE PREVIEW

Implementation of a Fluctuation Smoothing Production Control Policy - - PowerPoint PPT Presentation

May 01, 2023 265 likes •455 views

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere Implementation of a Fluctuation Smoothing Production Control Policy in IBMs 200mm Wafer Fab James R. Morrison Central Michigan University Department of

slide-1
SLIDE 1

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 1 –

Implementation of a Fluctuation Smoothing Production Control Policy in IBM’s 200mm Wafer Fab

James R. Morrison Central Michigan University Department of Engineering and Technology Brian Campbell Elizabeth Dews John LaFreniere IBM, Systems and Technology Group

slide-2
SLIDE 2

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 2 –

Presentation Overview

  • IBM’s 200mm wafer fabricator
  • Competing control methodologies
  • The basic FSVCT policy and extensions
  • Estimation of cycle times
  • Implementation challenges
  • Performance evaluation
  • Concluding remarks
slide-3
SLIDE 3

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 3 –

IBM’s 200mm Semiconductor Wafer Fabricator

  • Reentrant process flow –30 or more visits to

certain tool groups

  • Products from over 30 different semiconductor

technologies

  • Dozens of products within each technology
  • Up to 400 stages of processing for a single product
  • Over 200 distinct tool groups
  • Production capacity on the order of 1000

wafers/day

  • Cycle times generally ranging from 40 to 60 days

with actual production time around 20 days

  • Historical focus on technology excellence rather

than manufacturing efficiency

Tool Group 1 Tool Group 2 Tool Group 3 Tool Group 4 Tool Group 5 Tool Group 6 Tool Group 7 Tool Group 8 Tool Group 9 Tool Group 11 Tool Group 12 Tool Group 10

Lots enter

slide-4
SLIDE 4

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 4 –

Competing Control Methodologies

  • Objective: Choose next lot to enter into production (lot sequencing or

dispatching) to minimize overall cycle time and meet due date targets

  • Variant of continuous flow manufacturing (CFM) – Existing policy

– Produce a fixed minimum number of lots of each kind per day – Not sensitive to changes in loading – Prioritization scheme employed to implement preference

  • Critical ratio (CR)

– Due date based policy – Considered easy to implement

  • KANBAN

– Enforce maximum queue lengths at each stage of production

  • Fluctuation smoothing for the variation of cycle time (FSVCT)

– Reduce variation of cycle time by driving all lots to the same average CT – Successful in many simulation studies

  • Fluid limit inspired policies

– Theory suggests potential for improved cycle time performance – Requires optimization and fab model

  • Deterministic finite–horizon mathematical and constraint programming

– ILOG software product (own CPLEX optimization engine) – Requires detailed data and setup

slide-5
SLIDE 5

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 5 –

Toward the Basic FSVCT Policy

  • First “law” of queueing

– D/D/1 queue at 95% loading  Normalized CT = 1 – M/D/1 queue at 95% loading  Normalized CT = 10.5 – Variation leads to cycle time

  • Reducing the variation of lateness

– Let d(l) = due date (time) of lot l – Let h(l) = expected remaining cycle time of lot l – Slack(l) = – expected lateness = – [ h(l) – ( d(l) – Now ) ] – Select lot with least slack (greatest expected lateness)  drive all lots to the same lateness

  • Example:

– Now = – Let d(l) = due date (time) of lot l = 10.8 days – Let h(l) = expected remaining cycle time of lot l = 12.3 days – Slack(l) = – expected lateness = – [h(l) – ( d(l) – Now ) ] = – [ 12.3 – ( 10.8 – 0 ) ] = – 1.5 days – Give priority to lots with greater expected lateness

slide-6
SLIDE 6

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 6 –

Basic FSVCT Policy

  • Reducing the variation of lateness

– Let e(l) = exit date of lot l – Above policy should reduce the variation of lateness(l) = e(l) – d(l) = exit date – due date

  • Reducing the variation of cycle time

– If set d(l) = arrival date (time) = a(l) – Expect a reduction in the variation of e(l) – a(l) = cycle time – Slack(l) = – expected cycle time = – [ h(l) – ( a(l) – Now ) ]

  • Example:

– Now = – Let a(l) = arrival date (time) of lot l = -15.5 days – Let h(l) = expected remaining cycle time of lot l = 14.5 days – Slack(l) = – expected cycle time = – [h(l) – d(l) ] = – [ 14.5 – ( – 15.5 – 0 ) ] = – 30 days – Give priority to lots with greater expected total cycle time

  • Expect a reduction in overall cycle time if we reduce the variation of the

cycle times

slide-7
SLIDE 7

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 7 –

Incorporating Diverse Cycle Times

  • A myriad of cycle time expectations

– Technologies have very different processing requirements (e.g., 8 days vs. 30 days) – Within technologies, some customers may purchase preferential treatment

  • G arrival processes to the system

– Route through the fab is common (may be Bernoulli) – Expected remaining cycle time at each stage of production is common

  • Multiple cycle time FSVCT slack

– Let hg(l) = expected remaining cycle time for lot l from arrival process g – Let CTg = total expected cycle time for lots from arrival process g – Let CTNOR = arbitrary normalization constant > 0 – For a lot l from arrival process g define

  • Alternate approaches

– Scaled lateness (from the expected cycle time) – For scaling constants Kg (capturing the relative importance of lateness for each arrival process g)

g NOR g

CT CT l l a l Slack )] ( ) ( Now [ : ) ( h    

g g g

K CT l l a l Slack ] ) ( ) ( Now [ : ) (      h

slide-8
SLIDE 8

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 8 –

Estimation of Expected Cycle Times

  • Fundamental parameters for multiple cycle

time FSVCT:

– CTg = total expected cycle time for lots from arrival process g – hg(l) = expected remaining cycle time for lot l from arrival process g

  • Absence of plant model with routes,

capacities and cycle times

  • How to determine cycle times?

– Simulation is out – no model, costly to create – Measure and use existing CT (update as needed) – Set some CTs (preferred customers)  determine CT imposed on remaining lots

Tool Group 1 Tool Group 2 Tool Group 3 Tool Group 4 Tool Group 5 Tool Group 6 Tool Group 7 Tool Group 8 Tool Group 9 Tool Group 11 Tool Group 12 Tool Group 10

Lots enter

slide-9
SLIDE 9

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 9 –

Estimation of Expected Cycle Times

  • Resort to a Corollary of Little’s Law:

– For each arrival process g:

  • Denote by Pg the number of stages of production
  • Denote by lg the throughput rate for lots from g

– Let Lg = aggregate rate at which stages of production are completed

= lgPg

– Little’s Law: lg = Ng / CTg

  • Ng is mean number of g lots
  • CTg is mean cycle time for a g lot

– Aggregate rate of completion of stages of production Lg = Pg Ng / CTg

slide-10
SLIDE 10

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 10 –

Estimation of Expected Cycle Times

  • Measure the historical rate of completion
  • f stages of production

LH

g

= Pg Ng / CTg

  • Assume (independent of control policy)

that

– There exist collections of arrival processes with

  • We can employ historical throughput rates

to predict cycle times

– Fix Ng – Fix Pg – Choose/find CTg satisfying assumption

 

 

 L  L

c g g g g c g H g c

CT N P :

  • EXAMPLE:
  • Suppose for two arrival processes

– N1 = 100 lots, P1 = 300 stages – N2 = 400 lots, P2 = 300 stages

  • Assume (independent of control policy)

that

– Lc = 3000 lot*stages/day

  • If CT1 = 40 days

– LH

1

= P1N1/CT1 = (300 stages)(100 lots)/(40 days) = 750 lot*stages/day

  • Lc = LH

1 + LH 2

LH

2 = Lc – LH 1 = 2250 = (300*400)/CT2

 CT2 = 53.3 days

slide-11
SLIDE 11

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 11 –

Estimation of Expected Cycle Times

  • Employ this approach for

– CTg by application to the fabricator – hg(l) by application to each toolset

  • Determine the consequences of prioritization

– P1 = P2 = 300 stages – N1 = 100 lots, N2 = 400 lots – Cycle time targets: CTT

1 = 20 days, CTT 2 = 50 days

– Lots from arrival process 1 are for preferred customers – To ensure CT for preferred customer lots:

  • Devote L1 = (300 stages)(100 lots)/(20 days) = 1500 lot*stages/day

– The remaining Lc – L1 aggregate throughput is devoted group 2 lots

  • Let CT2 = e*CTT

2

  • Then

T c T

CT N P CT N P

1 1 1 2 2 2

/ /  L  e

slide-12
SLIDE 12

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 12 –

Implementation Challenges

  • Cycle time targets must be mutable

– Management directed reprioritization of WIP – Update: CTg  CT’g – Update: hg(l)  h’g(l) – Arrival dates must be updated: a(l)  a’(l) –

  • Patience and maintenance of focus
  • Programming resource allocation

)] 1 ( / ) 1 ( ' [ * ] ) ( [ ) ( '

g g

CT CT Now l a Now l a   

Initial Stage Final Stage Current Stage Expected Stage Initial Stage Final Stage Current Stage Expected Stage

slide-13
SLIDE 13

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 13 –

Performance Evaluation

  • Multiple cycle time FSVCT

– Implementation date: April 2005

  • Performance evaluation challenges

– No fabricator cycle time model – Ever changing load and mix – Loading reduction prior and subsequent to implementation – Queueing time consists of roughly 30% of the cycle time

  • Alternative considerations

– Cycle time variation reduction – Variation as a function of cycle time – Throughput as a function of WIP

Cycle Time Performance Curves for Manufacturing Systems

0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0.2 0.4 0.6 0.8 1 Loading (Utilization) Normalized Cycle Time M/M/1 - Cycle Time M/D/1 - Cycle Time Practical System - Approximate Cycle Time

slide-14
SLIDE 14

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 14 –

Performance Evaluation: Standard Deviation

  • For the same cycle time

– M/M/1 has greater standard deviation of cycle time – M/D/1 has lesser standard deviation of cycle time

  • Does our change in control policy shift the standard deviation

performance of the fabricator?

M/D/1 Standard Deviation is Less than M/M/1 Standard Deviation at Constant Cycle Time

1 2 3 4 5 1 2 3 4 5

Normalized Cycle Time Standard Deviation of Cycle Time

M/M/1-STD M/D/1-STD

slide-15
SLIDE 15

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 15 –

Performance Evaluation: Standard Deviation

  • Reduction in fabricator loading
  • Queueing time about 30% of the cycle time
  • Reduction in standard deviation of cycle time

– Apparent immediate improvement – Does the trend represent operation on a new performance curve?

Normalized Cycle Time (X Factor) Standard Deviation of Normalized Cycle Time After FSVCT - Standard Deviation Before FSVCT - Standard Deviation

Monthly Standard Deviation of Cycle Time: Before and After Implementation of FSVCT

slide-16
SLIDE 16

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 16 –

Performance Evaluation: Throughput and WIP

  • For the same fixed level of work in process (WIP)

– Exponential server has less throughput – Deterministic server has ideal throughput

  • Does our change in control policy shift the throughput

performance of the fabricator?

Throughput as a Function of WIP: Consequences of Randomness

0.2 0.4 0.6 0.8 1 1.2 2 4 6 8 10 Lots in System Mean Throughput (Normalized) Ideal Throughput Mean Throughput: Exponential Service To achieve 90% of maximum throughput, 9 lots in the system

slide-17
SLIDE 17

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 17 –

Performance Evaluation: Throughput and WIP

  • Reduction in work in process (WIP) within fabricator
  • Throughput and WIP reduction observed
  • In fairly low loading regime (30% of cycle time due to

queueing)

– No observable throughput performance change

Wafers in the Fabricator/Tool Completed Daily (Throughput) Total Stages of Production Throughput - Pre-FSVCT Throughput - Post-SVCT Estimated Ideal Throughput

Throughput Performance as a Function of WIP Weekly Data for IBM BTV 200mm

Post-FSVCT

slide-18
SLIDE 18

IBM Implementation of Fluctuation Smoothing Morrison, Campbell, Dews and LaFreniere CDC – ECC 2005 – 18 –

Concluding Remarks

  • Multiple cycle time FSVCT production control policy implemented in

IBM’s 200mm semiconductor wafer manufacturing facility April 2005

  • Provides:

– Variation smoothing heuristic – Capability to adjust cycle time targets during production – Approximate tool to assess the consequences of prioritization – Control requiring only simple fabricator measurements

  • Dramatically increased functionality coupled with variation reduction
  • Future directions:

– Capacity–state dependent control (incorporate status of bottleneck tools) – WIP–state dependent control