4-10-07 economicsofspeed - The Economics of Speed Professor...

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: The Economics of Speed Professor Robert Leachman Director, Competitive Semiconductor Manufacturing Program Engineering Systems Research Center University of California at Berkeley April 10, 2007 4/10/07 Leachman - Economics of Speed 1 Introduction The sales price of virtually every semiconductor device and manufacturing service declines over time ... 4/10/07 Leachman - Economics of Speed 2 "Time is money" DRAM price decline history 100% % of Introductory Price(LOG) 256K 1M 4M 16M 10% 64M 1% 0 4/10/07 2 4 6 8 10 12 14 16 18 20 Leachman - Economics of Speed 22 24 26 months after introduction 3 "Time is money" Intel CPU price decline history 100% % of Introductory Price(LOG) P 166 PII 300 PIII 700 P4 1.4G 10% 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 4/10/07 Leachman - Economics of Speed months after introduction 4 "Time is money" Foundry price decline history 100% 0.13u % of introductory pric e(LOG) 0.15u 0.18u 0.25u 0.35u 0.5u 10% 1 4/10/07 8 15 22 29 36 43 50 57 64 months after introduction 5 Leachman - Economics of Speed Introduction (cont.) This suggests there must be large economic values for increasing the speed of technology development, the speed of manufacturing implementation and the speed of manufacturing execution ... 4/10/07 Leachman - Economics of Speed 6 Introduction (cont.) Most analyses of fab economics concern estimation of wafer costs or die costs Reduction of cycle time reduces finance costs for working capital Reduction of cycle time improves yields Fewer lots at risk during process or equipment "excursions" not detected until end of line Quicker completion of experiments to qualify process or equipment changes that improve yield What is needed in addition is a tool to analyze impacts of manufacturing and R&D speed on sales revenues ("revenue gain model") 4/10/07 Leachman - Economics of Speed 7 Agenda Part 1 - Tools to impute value for manufacturing speed Part 2 Tools to estimate manufacturing speed 4/10/07 Leachman - Economics of Speed 8 Calculating the value ex-ante If we can predict the trend in sales prices, and if we can predict the reduction in elapsed times from a proposed engineering project, then we can predict the revenue gain from the project 4/10/07 Leachman - Economics of Speed 9 Revenue Gain Model Assume the sales market is immediately available for all output that can be completed and that prices are declining according to a known function of time Sales revenues can be increased by reducing durations for process development and qualification, equipment installation and qualification yield and volume ramps manufacturing cycle time (AKA flow time) Revenue gain = (total sales revenue for proposed durations) - (total sales revenue for status quo durations) 4/10/07 Leachman - Economics of Speed 10 Revenue Gain Model Let Pi(t) denote the sales price of device i at time t. Let Wi(t) denote the wafer starts, Yi(t) the yield, CTi(t) the manufacturing cycle time, and Hi the remaining product life, where t is the time of wafer start. Then the remaining lifetime revenue is Hi Pi (t + CTi (t ) )Wi (t ) Yi (t ) dt 0 Leachman - Economics of Speed 11 4/10/07 Ex-ante revenue gain model (cont.) The revenue gain of a project is the total difference in lifetime revenues for all products: Pi t + CTi A (t ) Wi A (t ) Yi A (t ) dt i 0 Hi ( ) - Pi t + CTi B (t ) Wi B (t ) Yi B (t ) dt i 0 4/10/07 Leachman - Economics of Speed 12 Hi ( ) Ex-ante revenue gain model (cont.) A practical approximation is to model sales price as a declining exponential (e.g., prices falling 50% per year) Pi (t ) = P0i e - at This facilitates calculation of the integral expressing lifetime revenue 4/10/07 Leachman - Economics of Speed 13 Ex-ante revenue gain model (cont.) Another practical approximation is to model the yield ramp as one minus a negative exponential: Die Yield YF Mature Yield Phase Ramp Phase Y0 -VT 4/10/07 0 RT Leachman - Economics of Speed H Time 14 Ex-ante revenue gain model (cont.) Die yield is modeled as follows: 1- e , t RT Y (t ) = Y0 + (YF - Y0 ) -bRT 1- e where RT is the time to ramp to mature yield and b defines the shape of the yield curve This also helps to make the integral expressing lifetime revenue easier to calculate 4/10/07 Leachman - Economics of Speed 15 - bt The total lifetime revenue TR = P0YFW { Y 1- 0 - aRT - e - aH Y 1 - e - aRT YF - a (VT + CT ) e e + + 0 -bRT Y a F a 1- e 1 - e - aRT 1 - e -(a +b )RT - a a + b where P0 is the initial selling price, YF is final yield, W is the fab-in rate, VT is development time, CT is cycle time, RT is yield ramp time, H is process life, Y0 is initial yield, a is rate of price decline, and b is yield learning rate. TR is the total revenue over the product life. 4/10/07 Leachman - Economics of Speed 16 Example revenue gains: Compressing the time to first silicon Assume sales price is initially $10,000 per 100%-yielding wafer and declining 25% per year, H = 5 years, Y0 = 0.50, YF = 0.90, W = 25,000 WSPM Consider cases of VT = 4, 3.5, 3, and 2.5 months VT Total revenue Revenue gain compared to (days) for 5 years prod'n VT = 120 days 120 $6.57 billion $0 105 90 75 4/10/07 $6.65 billion $6.73 billion $6.81 billion $80 million $160 million $240 million 17 Leachman - Economics of Speed Example revenue gains: What is one day of cycle time worth? Consider 200mm image sensor fab: W = 10,000 WSPW, YF = 420 DPW, CT = 50 days, P0 = $4.50 declining 25% per year, H = 2 years remaining product life Revenue gain from reducing cycle time by one day: H P(t + CT - 1)W (t )Y (t )dt - P(t + CT )W (t )Y (t )dt = 0 0 H $1.14 million 4/10/07 Leachman - Economics of Speed 18 Ex-post valuations of fab speed If we have actual history of production, sales, selling prices and elapsed times, we can calculate the actual revenue gain Consider a case in which all output is sold immediately after production. Then the revenue gain may be expressed as follows: 4/10/07 Leachman - Economics of Speed 19 Ex-post calculations R = O(t ) P(t )dt - O(t ) P (t + CT (t ))dt 0 0 H H where O(t) is the production output at time t, P(t) is the average selling price at time t, CT(t) is the reduction in cycle time for the output at time t, and R is the gain in revenue over the interval [0,H]. 4/10/07 Leachman - Economics of Speed 20 Step function approximation Suppose cycle time for output in month t was 45 days (compared to 80 days if project had not been undertaken) Then the revenue gain is O(t)*P(t) O(t)*[(25/30)P(t+1) + (5/30)P(t+2)] 0 4/10/07 t 30 t+1 60 t+2 90 21 Leachman - Economics of Speed Example: the SLIM project at Samsung Joint project of Samsung and Leachman & Associates LLC to reduce cycle time Time frame: March, 1996 - December, 2000 Application scope: Two fab lines initially, eventually expanded to all semiconductor factories of the company 4/10/07 Leachman - Economics of Speed 22 Samsung's DRAM cycle times 100 90 Avg. fab cycle tim e (days) 80 70 60 50 40 30 20 Dec-95 Jun-96 Dec-96 SLIM begins SLIM begins on all lines 4M 16M 64M 128M SLIM begins Intrinsic cycle time Jun-97 D ec-97 Jun-98 D ec-98 Jun-99 4/10/07 Leachman - Economics of Speed 23 Samsung's DRAM selling prices 200.00 180.00 160.00 140.00 120.00 100.00 80.00 60.00 40.00 20.00 0.00 M ar-96 Sep-96 M ar-97 Sep-97 M ar-98 Sep-98 M ar-99 Sep-99 4M 16M 64M 128M Average selling price ($) 4/10/07 Leachman - Economics of Speed 24 Financial impact of SLIM Sales revenues for the DRAM output of Samsung's fab lines over the period March, 1996 - December 2000 were tallied: $21.9 billion Sales revenues were re-computed assuming fab cycle times had stayed at 80 days 4/10/07 Leachman - Economics of Speed 25 Financial impact of SLIM Through 2000, SLIM increased DRAM sales revenues by 4.4% or US $954 million Including non-DRAM production, the revenue gain was $1.1 billion Subsequently, Samsung's DRAM market share rose from 18% to 40%. Meanwhile, Hynix went bankrupt, many DRAM makers left the market, others experienced large losses 4/10/07 Leachman - Economics of Speed 26 Summary Part I Ex-ante and ex-post methods have been proposed for calculating revenue gains from cycle time reduction Economic impacts of reductions in time to qualification, time to volume and cycle time may be calculated using spreadsheet tools 4/10/07 Leachman - Economics of Speed 27 General Strategy for Cycle Time Reduction Routinely calculate entitlement cycle times If entitlement is to be reduced, we must devise and evaluate engineering projects that reduce it At the same time, we improve execution tools to close the gap between actual and entitlement 4/10/07 Leachman - Economics of Speed 28 Methods for calculating/estimating entitlement cycle times Discrete-event simulation centralized dept., expert users Queuing theory decentralized spreadsheet tools, every engineer can be a user The application of Queuing Theory to calculate entitlement cycle times is recommended as a more effective strategy organizationally 4/10/07 Leachman - Economics of Speed 29 Part II - Introduction to Queuing Analysis Cycle time at step j on equipment type k = Queue time + Standard cycle time CT jk = QT jk + SCT jk Standard cycle time is time from lot start to lot complete. Queue time is time lot spends waiting to start. 4/10/07 Leachman - Economics of Speed 30 Queuing Analysis (cont.) Basic formula for queue time per lot: Let j index process step, k index equipment type Let u denote utilization of availability, PT denote process time per lot or per batch, m denote no. of qualified machines, A denote availability, ca denote arrival c.v., ce denote effective service time c.v.: 2 ca 2 + cek uk 2 ( m j +1) -1 PTk QT jk = m (1 - u ) A 2 k k j 4/10/07 Leachman - Economics of Speed 31 Queuing Analysis (cont.) Key point: Wait time ~ (u ) m } {Process time/Availability} {Variability} { m (1 - u ) One can reduce cycle time if any of the above terms is reduced 4/10/07 Leachman - Economics of Speed 32 Queuing Analysis (cont.) {Variability} ce = effective service time c.v. (reflects machine down time) Let c0 denote intrinsic process time c.v., cr denote repair time c.v., A denote availability, MTTR denote mean time to repair, PT denote avg. process time MTTRk ce = c + (1 + cr ) Ak (1 - Ak ) PTk 2 k 2 0 2 k Utilization should be set lower for machines with higher variability 4/10/07 Leachman - Economics of Speed 33 Queuing Analysis (cont.) {Utilization} System performance is very sensitive to high utilization levels. Wait Time vs. Utilization 10 9 8 7 6 5 4 3 2 1 0 0. 1 0. 15 0. 05 0. 2 0. 25 Balancing utilization reduces wait time. Increasing the number of qualified machines reduces wait time: Wait Time (1 machine) Wait Time (2 machines) Wait Time (3 machines) Wait Time (8 machines) Wait Time 0. 3 0. 35 0. 4 0. 45 0. 6 0. 65 0. 5 0. 55 Utilization 4/10/07 Leachman - Economics of Speed 34 0. 9 0. 95 0. 7 0. 75 0. 8 0. 85 Queuing Analysis (cont.) General strategy: Compute entitlement cycle times for status quo Compute estimates of cycle time reduction from implementing potential new qualifications, addition of new machines, changes in recipes, assignment of recipes to machines, etc. Determine CT ROI per engineering hour or per unit capital expenditure, rank projects and implement accordingly 4/10/07 Leachman - Economics of Speed 35 Extensions for special cases Batch tools (e.g., diffusion furnaces) Queuing entities in this case are batch recipes rather than product-steps (which share recipes) Extra wait time to build up batch Cycle time for recipe j on equipment type k = Batching time + Queue time + Standard cycle time CT jk = BT jk + QT jk + SCT jk 4/10/07 Leachman - Economics of Speed 36 Batch tools (cont.) Basic formula for batching time: Let j index recipe, k index equipment type, i index product Let bjk denote avg. batch size for recipe j on equipment k, Si denote total weekly lot starts rate of product i, nji denote number of times recipe j occurs in the process flow of product i b jk - 1 168 BT jk = 2 n ji Si i( j ,k ) 4/10/07 Leachman - Economics of Speed 37 Batch tools (cont.) Another key point: The batch size can be set either too big or too small. Best value depends on lot arrival rate and machine utilization: Cycle time vs. Batch size for TEL_OXWETG 5.5 Avg. cycle time (hours) 5.25 5 4.75 4.5 4.25 4 3.75 3.5 1.5 2 2.5 3 3.5 4 4.5 5 5.5 Avg. batch size (Lots) 4/10/07 Leachman - Economics of Speed 38 Extensions (cont.) Machines with setups (e.g., implanters) Queuing entities in this case are setup batches rather than lots (which share setups) Extra wait time for lot's turn within the setup batch Cycle time at step j on equipment type k in batch l = Queue time + Batch time + Standard cycle time CT jk = QT jk + BTl + SCT jk 4/10/07 Leachman - Economics of Speed 39 Setup machines (cont.) Basic formula for wait-within-batch time: Let bl denote avg. batch size for setup type l Let PTl denote the average process time per batch bl - 1 BTl = 2b PTl + st l l 4/10/07 Leachman - Economics of Speed 40 Case-study in Entitlement Cycle Time Planning 4/10/07 Leachman - Economics of Speed 41 Introduction to the Fab Fab A had approx. $650M installed capital, was fabricating approx. 15K 8-inch silicon wafers per month in technologies ranging from 500nm down to 90nm Most volume was 150nm or 130nm 4/10/07 Leachman - Economics of Speed 42 Introduction to Fab (cont.) Wafer fabrication involved building up ~18 layers of circuitry on silicon wafers, ~ 500 total steps Wafers moved from step to step in lots of 25 There were 66 equipment types involved, anywhere from 1 to 15 machines in each equipment type "New" machines cost anywhere from $150K to $15M 4/10/07 Leachman - Economics of Speed 43 Management strategy Carry out queuing analysis to predict changes in cycle time as a function of product mix and volume equipment set and machine qualifications Analyze CT ROI for alternative machine and engineering investments (or dis-investments) Find the cheapest way to maintain good CT performance 4/10/07 Leachman - Economics of Speed 44 Input data 2 ca 2 + cek uk 2 ( m j +1) -1 PTk QT jk = m (1 - u ) A 2 k k j , Arrival rate c.v. (ca): no data, set equal to unity Intrinsic service rate c.v. (c0): no data, set equal to unity Process time PT and SCT: actual statistics for photo, IE stopwatch standards for all other equipment types Actual data for m u computed from fab starts rate and PT Actual statistics for b, MTTR, cr, and A 4/10/07 Leachman - Economics of Speed 45 Implementation The queuing models and heuristics were incorporated into Excel workbooks Automated refresh of most input data from fab databases Every process engineering dept. uses its queuing workbooks 4/10/07 Leachman - Economics of Speed 46 Implementation and results Validation: the fab-wide average cycle time predicted by the model was 87% of actual fab-wide average cycle time The impact of lot and process holds (not modeled) accounted for most of this gap The model was used to plan a volume ramp from 3350 WSPW to 4900 WSPW, determining what tools to buy and what photo qualifications work to complete 4/10/07 Leachman - Economics of Speed 47 Algorithm for capacity planning Determine baseline cycle time (current volume and mix) Determine cycle time for next volume/mix increment if no equipment additions are made Determine cycle time reduction individually for each tool type if one new tool is added Divide this reduction by cost of that equipment to establish a CT ROI Incrementally add equipment in order of decreasing ROI until baseline CT is recovered 4/10/07 Leachman - Economics of Speed 48 Results Ramp plan to increase wafer volume by 45% involved only ~ $20M in non-photo capital and only ~ $20M in photo capital (compared to $650M for existing fab), at the same time the mix was strongly shifted to newer process technologies Cycle time performance was maintained during ramp 4/10/07 Leachman - Economics of Speed 49 Results (cont.) Case-study company uses TSMC foundry as a back-up production source For 150nm devices, TSMC charged $1600 per wafer. Company's wafer cost is $800, i.e., 50% less! TSMC's cycle time is 2.3 DPML (18 layers); Company's cycle time was 1.6 DPML (was 2.3 DPML before project started), i.e., 30% less! 4/10/07 Leachman - Economics of Speed 50 Proposed approach to process engineering Every proposal for an engineering project must quantify changes in fab throughput, yield and cycle time These changes are dollarized and the true economic value of the project is demonstrated Engineering is rewarded for improving entitlement cycle time as well as for yield or throughput improvement 4/10/07 Leachman - Economics of Speed 51 Conclusions Cycle time has a large economic impact that can and should be measured Cycle time can and should be managed with as much discipline and sophistication as has been brought to yield management Effective spreadsheet tools have been developed and demonstrated in actual practice 4/10/07 Leachman - Economics of Speed 52 Textbook reference on Queuing Analysis Factory Physics, by Hopp and Spearman, McGraw-Hill, ISBN 0-256-24795-1 4/10/07 Leachman - Economics of Speed 53 References Economics of Speed Report CSM-47, "Understanding Fab Economics," may be ordered from the CSM website: http://euler.berkeley.edu/esrc/csm The CSM Fab Economics Workbook can be downloaded at no charge from the web site You can e-mail questions to me at leachman@ieor.berkeley.edu 4/10/07 Leachman - Economics of Speed 54 ...
View Full Document

Ask a homework question - tutors are online