DOI 10.1007/s10845-008-0079-3
Measuring the relative performance for leading fabless firms by using
data envelopment analysis
Mei-Tai Chu · Joesph Z. Shyu · R. Khosla
Received: 1 December 2006 / Accepted: 1 July 2007 / Published online: 16 March 2008 © Springer Science+Business Media, LLC 2008
Abstract IC Design (fabless) is critical for the global
semi-conductor industry. The total revenue of all global fabless firms in 2003 was about US$20 billion, with the top 30 firms earning accounting for 96% of the market share. To examine the leaders in the field, this research analyzes the relative performances of those top 30 fabless firms. Fabless firms are often evaluated based on subjective judgments, and an overall scheme to measure the performance involving objective, multi-input and multi-output criteria is yet to be established. There is also a need for identifying and deter-mining suggestions of how specific firms could improve their performance. Data Envelopment Analysis (DEA) method has been employed in this paper to satisfy the above needs. Using the input and output data of 2003, this study used the DEA method to build a model to evaluate the performance of those global top 30 fabless firms. The current research used four efficiency models: CCR, A&P, BCC, and Cross-Efficiency. To offer a comparison of efficiencies and associated discus-sions, an analysis of the Scale-Return is provided. Finally, the performance of various fabless firms in 2003 is ana-lyzed. According to the CCR and A&P models, the results showed that the top ten Decision Management Units (DMUs) achieved better operation performance among the 30 leading global fabless firms.
Keywords Performance· IC design (fabless) · Data
envelopment analysis (DEA)
M.-T. Chu (
B
)· R. KhoslaBusiness Systems and Knowledge Modeling Laboratory, La Trobe University, Bundoora Campus, Melbourne, VIC 3086, Australia e-mail: debbiechu0421@hotmail.com
J. Z. Shyu
Institute of Management of Technology, National Chiao-Tung University, 1001 Ta Hsueh Rd., Hsinchu 300, Taiwan
Introduction
The total output value of global fabless firms was about US$20 billion in 2003. Of this figure, the top 30 firms had a market share of US$19.2 billion, or 96% of all sales
(Semiconductor Yearbook of 2003). The electronic
compo-nent market will continue to boom in the future since the components for cell-phones, CDs, CMOS, LCD Displays, DRAM, digital cameras, DDS and a host of other products will continue to increase in demand. These new products show that a new phase of the consumer electronics era is coming and that the new battlefield of top global enterprises will be full of challenges. Not only IT but also fabless firms, both of which are the upstream portion of the semiconduc-tor industry, can expect to harvest such growth opportunities from this exciting trend.
There are several key elements explaining why the major players can hope to continuously lead in the fabless field. First, they still prevail with technological innovation and superior patent protection. Second, they build solid supply chains and high entry barriers to keep out competition. Third, they provide total solutions for customers; they provide excel-lent technical specifications, have flexible marketing as well as pricing strategies, and maintain stable chip OEM partner relationships. In other words, even though there are many fac-tors contributing to success, excellent management perfor-mance is always the main key for outstanding fabless firms. Fabless firms, however, struggle with managing extreme boom and bust cycles. Whenever the company finds the right direction for business and has favorable R&D, then the busi-ness becomes more prosperous. These factors, working together, can make the company’s products and achievements greater and greater. On the other hand, once the company’s operation goes poorly and achievement is bad, a negative cycle is set off. The firm’s competitiveness is downgraded in
the market and investment begins to dry up. The company will have a harder time struggling with the dilemma of losing manpower and wealth (Semiconductor Yearbook of 2003). Such are the characteristics in today’s fabless industry.
In the coming years, the global fabless industry will be the scene of frequent integration and the fabless, which has a single and competitive product line, will be the main target for big factories. The top 30 global fabless firms are mainly in the USA and Taiwan (Chang and Tsai 2002;Hung and Yang 2003). It is crucial for every company to find an objective effective evaluation standard based on scientific principles. To date, an overall scheme for measuring the performance of fabless firms involving multi-input and multi-output cri-teria has not been established (Shuai et al. 2004). There also lacks suggestions of how specific firms could improve their performance. Remarkably, this study found that the DEA method meets the above needs. This research employs the DEA method to evaluate the relative performance of the top 30 fabless firms, using input and output data of 2003 (
Semi-conductor Yearbook of 2003).
The DEA relies extensively on both efficiency analysis and operation performance.Chen et al.(2003) discussed the mul-tiplier bounds in DEA.Kleine(2004) mentioned a general model framework of DEA.Lewis and Sexton(2004) used DEA to analyze efficiency of organizations with complex internal structure.Liu et al.(2003a,b), used DEA to assess the efficiency of Taiwan’s colleges.Opricovic and Tzeng(2003) compared DEA and MCDM method.Tzeng et al.(2001) used DEA to evaluate the production efficiency for Taipei city bus company.Yu et al.(2004) analyzed fuzzy multiple MOPA to DEA with imprecise data. The DEA is a unified macro-index that deals with many different inputs and outputs at the same time without the prior knowledge of the function of inputs and outputs by using the Non-Parametric Approach
(Charnes et al. 1985;Chen and Iqbal Ali 2002). DEA can
avoid errors caused by the assumption of productive function in the unclear relationships among input and output. It is not necessary to have the same measurement unit; this flexibility makes it easier to deal with the data (Farrell 1957;Charnes
et al. 1978;Semiconductor Yearbook of 2003;Andersen and
Petersen 1993;Doyle 1992). Therefore, this research adopted
DEA as the analysis tool and it includes CCR, A&P, BBC, and Cross-Efficiency.
The goal of this research is to use the DEA to analyze the top 30 global fabless achievements from the relative opera-tion performance. On the basis of the 2003 data, the index regarding input includes Capital stock, Net Working Capi-tal, and Long-Term Investments. In terms of output, Revenue and Earnings before Taxes (EBT) are included. The authors have tried to solve the problems by using the DEA model in this research and have presented the objective results that they hope will help fabless firms conduct better internal eval-uations.
In sect. “The current status of the global fabless indus-try” of the text, authors introduce the present status of the global fabless industry. Section “Research methods” contains the description of DEA and other evaluation methods for the operation performance. Section “An illustrative exam-ple” includes analysis and discussion. Section “Conclusions and recommendations”, the conclusion, includes recommen-dations for fabless firms to incorporate these findings and help them reach their strategic goals. The limitations of this research are also discussed.
The current status of the global fabless industry
From the revenue of the top 30 fabless corporations, it is easy to notice that there are some key players dominating the whole market. As mentioned above, in 2003, 96% of the global fabless revenue was generated by just 30 firms, primarily in the US, Taiwan, and Canada. Obviously, the fabless industry is not only becoming more concentrated daily, but the competition is getting more intensive. The list of the top 30 global fabless firms is shown in Table1. The product positioning of various fable firms is discussed next. Product positioning of the top 30 fabless firms
Qualcomm occupied 95% of the CDMA chipsets of the 3G wireless communication networks. ATI and NVIDIA occupied 90% of the drawing processors. Xilinx and Altera occupied 80–90% of the logic editor components. MediaTek occupied 50% of the single chips of VCD and DVD play-ers. ATI and NVIDIA have made enormous profits over the past three years by successfully riding the trend of applied multimedia drawers. MediaTek, Sunplus, Ali, ESS, Zoran, and Cirrus Logic have fought fiercely in the storage chip mar-ket. In spite of their already-strong names in the fabless indus-try, they have placed great resources into the DVD±RW, DVD-Recorder, MPEG-4 and other new chip product lines. Sunplus has been very successful with its DSC control chips. Novatek has carved its own place in the Display Driver IC field thanks to its increasing supply of TFT LCD panels and low price. Also, the flash storage chip is one of the nec-essary devices for cell-phones, so SanDisk and SST have also entered into the relevant field of flash IC manufactur-ing. From the above description, one can easily determine the major contours of the global fabless industry as shown in Table2.
Brief summary
The global fabless industry will probably undergo much inte-gration in the coming years. Fabless firms with proven suc-cess will become the main targets for the big factories to take over. Take the leading WLAN chip manufacturer, Intersil,
Table 1 The top 30 fabless firms’ rank/distribution/revenue (unit: US$ million)
Rank Fabless firm Distribution 2003 revenue 2002 revenue Growth rate (%)
1 Qualcomm US 2,466 1,942 27.0 2 Nvidia US 1,823 1,766 3.2 3 Broadcom US 1,611 1,090 47.8 4 ATI Canada 1,511 525 187.8 5 Xilinx US 1,300 1,125 15.6 6 MediaTek Taiwan 1,116 865 29.0 7 SanDisk US 1,080 485 122.7 8 Altera US 827 712 16.2 9 Marvell US 820 480 70.8 10 Conexant US 633 621 1.9 11 VIA Taiwan 598 736 −18.9 12 Qlogic US 516 415 24.3 13 Adaptec US 437 411 6.5 14 Globespan Virata US 379 231 64.1 15 Aeroflex US 341 203 68.3 16 Sunplus Taiwan 325 253 28.5
17 Silicon Lab. Taiwan 325 182 78.7
18 Novatek Taiwan 320 196 63.0 19 SST US 295 275 7.4 20 Realtek Taiwan 272 269 1.3 21 MegaChips Japan 271 345 −21.3 22 ICS US 257 228 12.6 23 PMC-Sierra Canada 249 218 14.4 24 OVTI US 249 82 203.5 25 Zoran US 217 149 45.2
26 Genesis Micro. Canada 213 196 9.1
27 Cirruss Logic US 198 293 −32.3
28 ESS US 195 273 −28.6
29 Semtech US 192 205 −6.4
30 Ali Taiwan 191 178 7.1
Source: IC insights, 2003/12
Table 2 The product positioning of the top 30 global fabless firms
Product Company
Communication ICs Qualcomm, Broadcom, Marvell, Conexant, Q-Logic, Silicon Lab, Realtek, PMC-Sierra, ICS, SMSC, Zarlink, DSP Group
FPGA Xilinx, Altera, Lattice Graphic ICs Nvidia, ATI
Multi-media ICs MediaTek, Sunplus, ESS, Zoran, Cirrus Logic, Realtek
Flash SanDisk, SST
LCD ICs Novatek, Genesis microchip, Zoran PC chipsets VIA, Ali
Ower management Semtech
Source: Topology research, 2004/03
for example, even though Intersil’s market share was more than 50%, it was first bought out by Globalspan Virata, and then Globalspan Virata was purchased by Conexant in late 2003. Such chain-reaction mergers will probably be increas-ingly common in the near future. There are now at least 40 companies in the world that are working on the development of WLAN chips, but indications show that there will be only seven firms remaining after takeovers and mergers in two
years: Intel, Broadcom, Marvell, TI, and three other already big names.
Looking forward, the main factor of success for fabless firms is their quality of operation performance. Under current consumer trends, big firms will combine with small compa-nies that have their own niche, thus achieving the twin goals of system integration and meeting market demand. Major fabless work can be done at the chip factory to take advantage of the big firms’ experiences and to develop the manufactur-ing, but the small companies cannot join this development. In addition, IC products will gradually head for unification, where scale and integration will become the main flashpoints of competition in the field. Finally, the whole semiconductor industry will probably be centralized to IDM and the large-scale fabless firms. Therefore, operation performance will determine the development and survival of small firms in the future.
Research methods
This section discusses this study’s methods of research and development and point out the differences between six traditional research methods. The authors also introduce the investigative elements and the applied methods in this study.
Brown et al.(1998) pointed out that the key to effective evaluation is taking the R&D unit as one part of the entire organization and then emphasizing the inputs, processes, out-puts, and benefits.Poh et al.(2001) addressed six research methods: the Scoring Method, Analytic Hierarchy Process (AHP), Comparative Method, Cost-Benefit Analysis, Eco-nomic Analysis, and Decision Tree Analysis. There are many ways to evaluate the problems of operation performance
(Cavalluzzo and Ittner 2004;Folan and Browne 2005;Hoque
2004;Ittner et al. 2003;Schmitz and Platts 2004). It is
neces-sary to use the multiple indices because a single index has its blind spots; for example, the financial approach cannot solve the input and output aspects. The DEA, in contrast, takes multiple criteria, multiple indexes, and the idea of relative importance as its core principle. DEA can overcome the tra-ditional evaluation shortcomings by using weights, normali-zation, and comparison (Liu et al. 2004;Ohta and Yamaguchi
1995;Zimmermann 1978;Charnes et al. 1985). Therefore,
this research used the DEA method, a input and multi-output approach. DEA is one kind of cost-benefit analysis that is used for evaluating the operation performance of the top 30 global fabless firms. There are five main merits associ-ated with DEA. First is to take the business cost-benefit ratio method and use the DEA to identify the operation perfor-mance of every company. Second, scholars have improved the DEA. There are four kinds of evaluation models com-monly used, and DEA makes it convenient to identify the operation performance and the comparative analysis of every company. Third, from the analysis of internal and external periodical databases, the DEA is not only suitable for eval-uating the achievement of nonprofit organizations but also for commercial enterprises. Fourth, fabless revenue in the semiconductor industry plays a crucial role, so fabless firms have an enormous influence on the global economy. Thus, there is an enormous need to find an evaluation model for the top 30 fabless firms so one can discover firms’ competitive advantages and thereby further develop the semiconductor industry. Fifth, analysis of the input and output data of the top 30 firms are based on fabless annual reports as found in the ITRI (Industrial Technology Research Institute) 2003 Semiconductor Yearbook and Dataquest. The following sec-tion is mainly about the characteristics of the DEA, basic assumptions, and the application of different models. Method of DEA
The DEA used in this research is a kind of mathematics pro-gramming model. It first applies the observed information into the model and then finds a DEA efficient frontier to cal-culate relative efficient values of each DMU among its group. Farrell (1957) first addressed the concept of the Determin-istic Non-Parametric Frontier. The determinacy means that the engineering level of all DMUs is the same and faces the
common production frontier. The non-parametric frontier is a pattern of non-preset production function; the multi-input efficiency evaluation established the foundation of DEA
the-oryFarrell(1957). This pattern has some basic assumptions.
First, the production frontier is formed by the most efficient DMU, and the ineffective DMUs are below the frontier. Sec-ond, Fixed Scale Returns are assumed. Third, the production frontier protrudes to the origin, so the slope of every dot is smaller than or equal to zero.
In DEA theory, when the combination of input and output of a certain DMU has fallen at the border of DEA, authors assume it is an efficient DMU. On the contrary, if the DMU has fallen out of the border, then this DMU is relatively inefficient. Many scholars have proposed and proven the analysis of the DEA model (Farrell 1957; Charnes et al.
1978;Semiconductor Yearbook of 2003). Basically, DEA is a
non-parametric analytical method with the following main characteristics. First, this approach is one estimating non-parametric maximum production. It is not necessary to set the relationship between previous inputs and outputs in the target function and therefore avoid the risk of wrong func-tion assumpfunc-tions. Second, the DEA model can calculate the relative efficiency values of the specific individual and the rel-ative group. Third, the DEA model sets up a comprehensive index by mathematical programming. It can measure the rel-ative efficiency among the different inputs and outputs. The DEA model can solve the problem of the different units of measurement caused by the evaluation of multiple inputs and outputs. Fourth, DEA is more objective and fair than the gen-eral questionnaires and the decisions of policymakers (such as in AHP). And fifth, DEA method can provide efficiency scores for multi-inputs and multi-outputs in one single step. This is similar to the relative analysis method with single input and single output.
The CCR model of DEA
The CCR model used in this research, a tool for measuring an organization’s efficiency, was created by Charnes et al. (1978). It supposes there are s kinds of output items, and n pieces of DMU using m kinds of input items; the k piece of DMU’s efficiency value can then be calculated by using Fractional Linear Programming (Charnes and Cooper 1984). A sample equation of CCR DEA linear programming for-mulation is as follows: Max Zk = s j= ujYj k Subject to: s j=1 uj.Yj k − m i= vi.Xi k ≤ 0, k = 1, . . . , n,
m i= viXi k = 1, uj ≥ ε > 0, r = 1, . . . , s, vi ≥ ε > 0, i = 1, . . . , m, uj= t·Uj; vi = t·Vi; t−1=Vi · Xi k BCC model of DEA
Banker et al.(1984) addressed the BCC model, the input and
output cost-benefit analysis. According to the BCC defini-tion, the Scale Efficiency (SE) is the quantity of input under the fixed output standard, while the ratio of the quantity of input is under the best production scale. The Technical Effi-ciency (TE) is the quantity of input under the fixed output standard and the ratio of the quantity of input under the fixed output standard. The assumption of the CCR model is that the Scale Returns are fixed to estimate the whole efficiency. If an inefficient situation occurs, it might be partly influenced by the scale factor but not by the inefficient technique. Thus, Banker et al. revised the CCR model to create the BCC model and examined the Technical Efficiency under dynamic scale condition.
A&P model of DEA
While working on the efficiency analysis, it may happen that the DMU efficiency value is 1 by the calculation of CCR. That is a lone outlier without sufficient discriminating power. For the efficient DMU,Andersen and Petersen(1993) proposed an advanced model which would not influence the ineffi-cient DMU, but the efficiency value of effiineffi-cient DMU will be greater than 1 after recalculation. One can then rank the efficient DMUs in order. The way to calculate efficiency is to eliminate the efficient DMUs from the reference set of the
CCR model, remove the B dot that originally lies on the fron-tier line by using the A&P model, and finally the production frontier turns into A BC, so the efficiency of dot B will be greater than 1 (Table3).
Cross-efficiency model of DEA
Doyle and Green (1994) explained the concept of
Cross-Efficiency. Compared to self-appraisal, it is a kind of peer-appraisal model. In the CCR model, if the efficient DMU from self-appraisal has few references, it shows the high pos-sibility of departing from groups and the Cross-Efficiency value will thus have a greater decrease in peer-appraisal. In the Cross-Efficiency matrix table, the Cross-Efficiency value (ek) of the k piece of DMU is the average of DMUk’s effi-ciency by using the virtual multiple calculation of the other DMU.
Analyses and discussion of various DEA models
This research adopts the CCR, BCC, A&P, and Cross-Effi-ciency models of DEA to evaluate the operation performance of the top 30 global fabless firms with the same input and output values. The authors have simply used the different theoretical foundations and different relative efficiency stan-dards to evaluate the companies and provide suggestions for improving their operation performance (Table4).
Scholars have improved the original CCR model so that the Cross-Efficiency model is now more objective. The authors took the result of the CCR and BBC models for the analysis while evaluating the scale efficiency values even though four models of DEA are calculated. As for the per-formance evaluation of the top 30 fabless firms, authors used the Cross-Efficiency model for analysis and comparison.
An illustrative example
This section is based on the characteristics, restrictions, procedures, measurements and models of DEA in sect. “Research methods”. The first step of the DEA model is to establish the index for analysis. In this context as shown in sect. “The current status of the global fabless industry”,
Table 3 The cross-efficiency matrix table
Source:Doyle and Green(1994)
DMU of peer-appraisal/DMU 1 2 . . . n 1 E11 E12 E1n 2 E21 E22 E2n .. . n En1 En2 Enn
Table 4 The comparison sheet of various DEA methods
Event/DEA model CCR model BBC model A&P model Cross-efficiency model
Time 1978 1984 1993 1994
Relative efficiency standard Specific DMU Specific DMU Adjacent DMU Peer DMU Weigh measurement Single favorable Single favorable Average of two Average of peer Relative efficiency character Subjective Subjective Subjective Objective
Method type Original type Improving type Improving type Improving type
Source: Xu, Ji Sheng, etc., 2002, using the data envelopment analysis to evaluate the research performance of ITRI
Remark: CCR model (total efficiency value) and BBC model (technological efficiency value) can achieve the scale efficiency value
authors collected data on the revenue, ranking, location, and product positioning of global top 30 fabless firms. It is very important to know what field of improvements that fabless firms should concentrate on. Spotting these trends will influ-ence the future competitiveness of the fabless industry.
Firstly, authors chose the top 30 fabless firms from “2003 IC Insights” statistics as DMU. Secondly, referring to the rel-evant documents, “2003 ITRI Semiconductor Yearbook” and each company’s annual report, authors found proper input and output items as parameters and conducted relationship analysis. Thirdly, authors selected appropriate DEA mod-els and adopted four modmod-els for analyzing and comparing the real examples. Fourthly, authors compared the efficiency analysis between CCR and BCC models. The primary inef-ficient sources of DMU came from a lack of Pure Techno-logical Efficiency (BCC efficiency) or Scale Efficiency, and recommendations for improvement have been made. Finally, authors included the rewards scale analysis to discuss the Scale Efficiencies of the top 30 fabless firms.
Selection of DMU
Golany and Roll(1989) thought that DMU must be
homo-geneous, which means the evaluating targets need to have the similar operation characteristics. These are outlined as follows:
The same internal essence: All of the evaluated firms are IC designs, and then transfer the products to the packag-ing factory for packagpackag-ing and testpackag-ing the products. Since the problems facing each company are similar, maximizing the operation value is their common goal.
The same external environment: Even though the semi-conductor companies are distributed across the world, their industrial environments are roughly the same. The operation inputs and output items of every company are the same, for instance: the biggest expenses are the costs of R&D and the fixed assets such as equipment, etc.
According to the aforesaid section, the DMU of this research, the name of the companies, the world distribu-tion, the profits, and the growth rates are stated as Tables1 and2. As shown in our results, the companies are listed in descending order according to revenue: Qualcomm, Nvidia,
Broadcom, ATI, Xilinx, MediaTek, SanDisk, Altera, Marv-ell, Conexant, VIA, Qlogic, Adaptec, Globespan Virata, Aeroflex, Sunplus, Silicon Lab. Novatek, SST, Realtek, MegaChips, ICS, PMC-Sierra, OmniVision, Zoran, Genesis Micro, Cirruss Logic, ESS, Semtech, and Ali.
Selection of input and output items
When using DEA to actually weigh criteria by priority, one cannot consider too many input and output items. Other-wise the efficiency value of every DMU will be 1 because of the idea based on Pareto Optimality criterion, and this goes against the original idea of weight efficiency (Lee and Li 1993). It is thus necessary to merge similar items or adopt factor analysis. As for the restriction of precise item quantity, authors considered the geometry room dimension of DEA is counted with a sum of input and output of DMU. When the input and output both increase, the number of DMUs must be increased correspondingly, and then one can use the envel-opment line principle to search for the most efficient DMUs. Authors referred to the Rule of Thumb for the item selection that the DMU should be at least twice the sum of input and outputBanker et al.(1984).
The selected input and output items from 30 companies are:
– Input items: From Table1, there were eight input items: R&D Expenses, Fixed Assets, Intangible Assets, Capital Stock, Cash, Net Working Capital, Long-Term Invest-ments, and Debt Ratios.
– Output items: From Table1, there were seven output items: Revenue, Earnings before Taxes (EBT), Net Income after Taxes, Earning per Share (EPS), Return on Common Equity (ROE), Return on Assets (ROA), and Turnover Ratios.
Financial experts’ selection process for inputs and outputs The 15 aforesaid inputs and outputs were subjected to another advanced selection process to see if they fit with the follow-ing principles. (1) data origins are credible; (2) it can be controlled; (3) it conforms to the current period relationships of input and output; (4) they have the same evaluation basis;
(5) they have the definite relationship with the operation per-formance.
As a basis of annual reports and public financial informa-tion, this research collected fifteen preliminary data (Appen-dix TableA.1) to evaluate the index via the five selecting principles. The data was then ranked in order and summed up five inputs and five outputs as shown in Appendix TableA.2. Correlation analysis and the double regression of input and output
The authors examined the input and output relationships by the Pearson correlation coefficient. The analysis results and variables are shown in Appendix TableA.3. One must then observe whether the relationships conform to isotonicity or not. In other words, if the input quantity is increased, then the output quantity cannot be reduced. The item must be rejected if there is a negative correlation. The input and output data of this research are all Ratio Scales, so one can adopt the Pearson Production-Moment Correlation for examination.
From Appendix TableA.3, one can find the correlation coefficients of items. The following ones are apparent and
positive: Capital Stock, Net Working Capital, and Long-Term Investments among inputs; Revenue and Earnings before Taxes (EBT) among outputs. These four above-mentioned values all meet the requirements of isotonicity and signifi-cance. The input and output material comes from the com-panies’ annual reports, experts’ appraisals, and correlation analysis. Therefore, authors adopted their inputs (Capital Stock, Net Working Capital, and Long-Term Investments) and two outputs (Revenue and Earnings before Taxes) as the indices of evaluation as shown in Table5.
Election and application of DEA model
After inspecting the isotonicity of the selecting inputs and outputs, authors chose the CCR, A&P, BCC, and Cross-Effi-ciency models of DEA for the effiCross-Effi-ciency value. Authors also examined the Integrated Technical Efficiency, Pure Techno-logical Efficiency, and Scale Efficiency of each DMU using BCC and CCR models. Since the calculating course is mis-cellaneous, one may utilize different PC software to calculate the efficiency value with LINGO, the coefficient correlation Table 5 Global top 30 fabless inputs-outputs data sheet in 2003 (Thousand US$)
Fabless/Item Input Output
Capital stock Net working capital Long-term investments Revenue EBT
Qualcomm 789,586 2,624,559 1,120,927 2,466,331 1,285,147 Nvidia 153,513 328,979 190,029 1,822,945 86,673 Broadcom 292,009 444,931 41,097 1,610,095 (934,738) ATI 237,227 92,600 711,100 1,510,992 (280,200) Xilinx 337,069 458,805 1,091,697 1,299,900 350,544 MediaTek 192,856 673,456 347,266 1,115,931 500,669 SanDisk 144,781 812,977 185,062 1,079,801 241,881 Altera 381,387 111,771 14,451 827,207 212,501 Marvell 127,456 287,499 0 819,762 63,352 Conexant 303,488 133,734 119,230 633,100 23,433 VIA 379,295 175,804 368,543 597,664 (50,016) Qlogic 103,473 206,342 0 516,200 215,601 Adaptec 106,772 170,487 6,346 437,200 (189,160) Globespan 268,586 133,734 119,230 379,100 23,433 Aeroflex 60,193 161,556 0 341,028 12,883 Sunplus 192,856 116,249 127,658 325,349 60,817 Silicon 48,850 81,138 0 325,305 66,196 Novatek 102,535 80,454 18,933 319,706 64,414 SST 94,723 175,866 83,046 295,041 (38,751) Realtek 197,597 170,718 117,630 272,005 84,613 MegaChips 246,610 115,178 1,248 271,297 2,795 ICS 67,898 163,687 32,000 256,900 71,541 PMC-Sierra 173,568 198,327 52,905 249,483 (15,843) OVTI 22,678 254,761 7,110 249,400 89,008 Zoran 33,231 67,954 0 216,528 (66,615) Genesis 31,248 48,670 0 213,400 (5,268) Cirruss 83,445 143,199 6,996 198,200 39,444 ESS 39,517 79,313 9,076 195,273 40,894 Semtech 73,013 124,048 86,119 192,079 42,718 Ali 51,594 (21,439) 41,730 191,082 152
with SPSS, and use EXCEL to find the fuzzy DEA model and the fuzzy multi-programming calculation.
CCR, A&P, and cross-efficiency value analysis
In the CCR model, if the value equals 1, this means the result is relatively efficient; in contrast, values smaller than 1 are relatively inefficient. The A&P model eliminates the efficient DMU itself from the reference set of CCR model to make the efficiency values equal to 1 become greater than 1, in order to further differentiate among the efficient DMUs. Both A&P and CCR models are self-appraisal. The reason they have high efficiency is because they offer higher virtual multiple value in accordance with the favorable inputs and outputs, so the evaluations are rather subjective. On the other hand, Cross-Efficiency is a peer-appraisal method. In sum, each model has its own function and influence on the analysis and explanation. The Cross-Efficiency model has been proved by this research and is comparatively objective.
Assisting the efficiency value of analysis by reference set frequencies: In order to enhance the discriminating power
of the CCR model, and to avoid non-discrimination situa-tion,λjderived from duel model is often used for assistance. Whenλjis> 0, all the related DMUj will be the reference set for the assessing units. Thus the higher number of times a DMU efficiency appears in the reference set of other DMU, the higher is its robustness of efficiency. If a DMU efficiency has not shown up in the reference set of other DMU, it will be an outlier (Andersen and Petersen(1993)).
The results of this research are reported in Table6. The CCR and A&P models using the 2003 data are both the results of self-appraisal. The better DMU among the top 30 fabless firms are Nvidia, Broadcom, ATI, Altera, Marvell, Qlogic, Silicon, OVTI, Genesis, and Ali. The poorer ones with effi-ciency values below 0.6 are Mega Chips, Conexant, ICS, SST, Xilinx, Cirruss, VIA, Globespan, Sunplus, Qualcomm, and Realtek. The more often the DMU appears in reference lists, the more robust that DMU’s efficiency is. The number of times each company was listed is thus: Genesis (8), Ali (8), Nvidia (7), ATI (7), Altera (7), OVTI (6), and Silicon (5). The comparison of ordinal scale efficiency: From the aver-ages of different efficiencies in Table6, the A&P and the Cross-Efficiency models scored highest while the CCR model
Table 6 CCR efficiency, A&P efficiency, and cross-efficiency
Number DMU CCR efficiency Reference set Cross-reference times A&P efficiency Cross-efficiency
1 Qualcomm 0.3026 1, 2, 4, 7, 24 0 0.3026 3.4954 2 Nvidia 1 2 7 1.8221 2.3285 3 Broadcom 1 3 1 1.0372 1.7767 4 ATI 1 4 7 1.1742 1.8704 5 Xilinx 0.4305 5 3 0.4305 1.8338 6 MediaTek 0.6456 2, 4, 6, 7, 24 0 0.6456 1.6768 7 SanDisk 0.6678 7 2 0.6678 1.5567 8 Altera 1 8 7 2.0289 1.2688 9 Marvell 1 9 1 1.5881 1.2143 10 Conexant 0.5838 2, 4, 8, 10, 30 0 0.5838 0.9960 11 VIA 0.3505 2, 4, 8, 11, 30 0 0.3505 0.9341 12 Qlogic 1 12 0 1.1501 0.8869 13 Adaptec 0.6989 3, 8, 9, 13, 17 0 0.6989 0.6987 14 Globespan 0.3404 2,8,14,17,30 0 0.3404 0.7157 15 Aeroflex 0.9225 15 0 0.9225 0.6586 16 Sunplus 0.3231 2, 8, 16, 17, 30 0 0.3231 0.6680 17 Silicon 1 17 5 1.1043 0.6472 18 Novatek 0.7202 8, 17, 18, 26, 30 1 0.7202 0.6518 19 SST 0.4362 4, 19, 24, 26 0 0.4362 0.6036 20 Realtek 0.2780 2, 4, 18, 20, 26, 30 0 0.2780 0.6259 21 MegaChips 0.5916 21 1 0.5916 0.5716 22 ICS 0.5498 4, 5, 22, 24, 26 0 0.5498 0.6046 23 PMC-Sierra 0.2656 8, 17, 23, 26, 30 0 0.2656 0.5605 24 OVTI 1 24 6 1.5652 0.5490 25 Zoran 0.9607 25 0 0.9607 0.4934 26 Genesis 1 26 8 1.3605 0.4845 27 Cirruss 0.4172 21, 26, 27, 30 0 0.4172 0.4943 28 ESS 0.8761 5, 24, 26, 28 0 0.8761 0.4577 29 Semtech 0.6567 5, 24, 26, 29 0 0.6567 0.4882 30 Ali 1 30 8 1.0701 0.4753
scored the lowest. Regarding the discriminating power, the efficiency value of 10 DMUs were 1 by the CCR model, so the efficiency value is unable to differentiate which one is better; some other reference sets are needed. However, the discrimination of the other two efficiency evaluations was better than that of the CCR model. Among them, A&P was the extension of the CCR model, so it had better discrimina-tion. Table7lists all DMU ranks under each model and the relevant analysis for discussing the efficiency influence of the three models. The data after ranking is an ordinal scale type, so it is suitable to use the Spearman Rank-Order Correla-tion so that its coefficient correlaCorrela-tion can show the consistent degree between four ranking groups.
There is high degree of correlation between rating by CCR efficiency and by A&P efficiency, since A&P model is an extension of the CCR model. For DMUs with an efficiency score less than 1, their ratings are the same either by CCR effi-ciency or A&P effieffi-ciency. Cross effieffi-ciency is a peer-evaluation type, their efficiency scores were obtained through subjective
assessing and there was a high degree of correlation related to the efficiency score rating among DMUs by this model. The correlation coefficient can be as high as 0.959. If one consid-ers three models together, the lowest correlation coefficient is 0.023, which implies that all kinds of models have positive correlation. Consequently, those DMUs that perform well will have higher ranking regardless of the models used.
Comparing the efficiency of the CCR and BCC models If one subdivides the Integrated Efficiency (CCR efficiency), one can find that inefficiency comes from a lack of Pure Tech-nological Efficiency (BCC efficiency) or Scale Efficiency. It means that the Integrated Technological Efficiency is the product of Pure Technological Efficiency and Scale Effi-ciency, representing the whole R&D efficiency of the top 30 fabless firms. Pure Technological Efficiency means the efficient application of inputs from every firm in the cur-rent year, so that it can reach the goal of minimum input Table 7 DMU ranking based on efficiency scores by different models
DMU Reference times of CCR efficiency A&P efficiency Cross-efficiency
Qualcomm 28 28 1 Nvidia 3 2 2 Broadcom 8 9 5 ATI 3 7 3 Xilinx 23 23 4 MediaTek 18 19 6 SanDisk 16 16 7 Altera 3 1 8 Marvell 8 3 9 Conexant 20 20 10 VIA 25 25 11 Qlogic 10 5 12 Adaptec 15 15 14 Globespan 26 26 13 Aeroflex 12 12 16 Sunplus 27 27 15 Silicon 7 8 17 Novatek 14 14 20 SST 22 22 18 Realtek 29 29 21 MegaChips 19 17 19 ICS 21 21 22 PMC-Sierra 30 30 23 OVTI 6 4 25 Zoran 11 11 26 Genesis 1 6 24 Cirruss 24 24 30 ESS 13 13 27 Semtech 17 18 28 Ali 1 10 29
Correlation coefficient (based on CCR efficiency & referenced times)
1 0.959 0.023
Correlation coefficient (based on A&P efficiency)
0.959 1 0.095
Correlation coefficient (based on cross-efficiency)
and maximum output. Its value shows the applied efficiency of input, and the Scale Efficiency represents the appropriate proportion of input and output of every firm in every year to reach maximum productivity. Greater values indicate more suitable scales and hence greater productivity.
Table8shows that inefficiency comes totally from a lack of Pure Technological Efficiency. For instance, Realtek, PMC-Sierra, Qualcomm, Sunplus, Globespan, VIA, SST, Cirruss, ICS, Conexant, MediaTek, and Novatek 13 are thir-teen companies with an inefficient application of inputs. Authors found the scale inefficient companies were Xilinx, MegaChips, Semtech, SanDisk, Adaptec, Cirruss, ESS, Aeroflex, SST, Zoran, VIA, Conexant, PMC-Sierra, and Globespan 14. In addition, six unfortunate companies lacked both the Pure Technological Efficiency and Scale Efficiency: Cirruss, SST, VIA, Conexant, PMC-Sierra, and Globespan. The reason behind their lack of Integrated Technological Efficiency came more from the degree in which they lacked Pure Technological Efficiency than from their scale inefficiency. This inclination showed that the low
technology inefficiency did not affect production scale. However, there were five companies—Xilinx, MegaChips, Semtech, SanDisk, and Adaptec—that had more serious problems of scale inefficiency. In scale inefficiency, Scale Returns of these five firms might increase or decrease pro-gressively. The suggested improvements are listed in the fol-lowing analysis.
Scale return analysis
Calculation of DMU efficiency scores by CCR model is based on the assumption of fixed scale return. In this sit-uation, the DMU inefficiency operation might come from a different Scale Return. When the Scale Efficiency value is equal to 1, it is a Fixed Scale Return. However, when it is not 1, the Scale Return increases or decreases accord-ingly. The larger the deviation, the greater is the increase or decrease in scale return. As was explained in the second para-graph of Sect. "Research methods", authors found all DMUs
Table 8 CCR efficiency, BCC efficiency and scale return of each fabless
DMU/Efficiency CCR model BCC model Scale model
Qualcomm 0.3026 0.3026 1 Nvidia 1 1 1 Broadcom 1 1 1 ATI 1 1 1 Xilinx 0.4305 1 0.4305 MediaTek 0.6456 0.6456 1 SanDisk 0.6678 1 0.6678 Altera 1 1 1 Marvell 1 1 1 Conexant 0.5838 0.6035 0.9674 VIA 0.3505 0.3648 0.9608 Qlogic 1 1 1 Adaptec 0.6989 1 0.6989 Globespan 0.3404 0.3415 0.9968 Aeroflex 0.9225 1 0.9225 Sunplus 0.3231 0.3231 1 Silicon 1 1 1 Novatek 0.7202 0.7202 1 SST 0.4362 0.4554 0.9578 Realtek 0.2780 0.2780 1 MegaChips 0.5916 1 0.5916 ICS 0.5498 0.5498 1 PMC-Sierra 0.2656 0.2696 0.9852 OVTI 1 1 1 Zoran 0.9607 1 0.9607 Genesis 1 1 1 Cirruss 0.4172 0.5113 0.8160 ESS 0.8761 1 0.8761 Semtech 0.6567 1 0.6567 Ali 1 1 1
Reference: The discrimination of scale returns: In the above Table, when p<0, the scale returns decrease progressively; when p=0, the scale returns
had only two kinds of Scale Returns - fixed and progres-sively increasing – instead of progresprogres-sively increasing only. Among these 30 firms, the ones with progressively increas-ing included Globespan, PMC-Sierra, VIA, Conexant, SST, Zoran, Ali, Semtech, Adaptec and Genesis. For DMU with sharp increase in Scale Return, the output increase rate is greater than the input increase rate, so as to expand Capi-tal stock, Net Working CapiCapi-tal, and Long-Term Investments Table9.
Discussion
The 2003 performance of the top 30 fabless firms used the ratio of input and output to find the relative efficiency of each company. The selection of the input and output values in this research was evaluated by experts’ discussions and the two-stage coefficient correlation of input and output data that fit in with the DEA analysis. The evaluation of operation performance was divided into three stages: the evaluation of operation, the evaluation of outputs, and the evaluation of benefits. This research relies on the reference set and the
four improvement types of DEA evaluations for differentiat-ing the operation performance of the 30 fabless firms in 2003 (Tables4and5) and probes into and compares the four DEA methods at the same time.
After analyzing and comparing the results, this research can give us the result of the competitive orientation and rel-ative operation performance; moreover, it provides fabless industry the reference and basis for the development of spe-cific policies. The performance evaluation of a fabless should not use a single financial input and output index. Using a sin-gle output weight value, or making comparisons with just a few companies, would result in inaccuracy. Therefore, our approach can obtain the weight of each standard objectively by using the DEA model of multi-input and multi-output in this research while at the same time find the objective analy-sis and orientation to evaluate the operation performance of global fabless firms. The author’s hope that, on the strength of these results and the relevant business efficiency, this infor-mation could provide decision-makers with the references needed to improve their operation efficiency and wise allo-cation of resources.
Table 9 Scale return assessment
DMU/Efficiency CCR Scale efficiency P Scale return
Qualcomm 0.3026 1 0 Fixed Nvidia 1 1 0 Fixed Broadcom 1 1 0 Fixed ATI 1 1 0 Fixed Xilinx 0.4305 0.4305 0 Fixed MediaTek 0.6456 1 0 Fixed SanDisk 0.6678 0.6678 0 Fixed Altera 1 1 0 Fixed Marvell 1 1 0 Fixed Conexant 0.5838 0.9674 0.3739 Increase VIA 0.3505 0.9608 0.2267 Increase Qlogic 1 1 0 Fixed Adaptec 0.6989 0.6989 5.1775 Increase Globespan 0.3404 0.9968 0.1204 Increase Aeroflex 0.9225 0.9225 0 Fixed Sunplus 0.3231 1 0 Fixed Silicon 1 1 0 Fixed Novatek 0.7202 1 0 Fixed SST 0.4362 0.9578 0.6479 Increase Realtek 0.2780 1 0 Fixed MegaChips 0.5916 0.5916 0 Fixed ICS 0.5498 1 0 Fixed PMC-Sierra 0.2656 0.9852 0.1398 Increase OVTI 1 1 0 Fixed Zoran 0.9607 0.9607 1.1807 Increase Genesis 1 1 5.3252 Increase Cirruss 0.4172 0.8160 0 Fixed ESS 0.8761 0.8761 0 Fixed Semtech 0.6567 0.6567 5.1336 Increase Ali 1 1 3.8388 Increase
Remark: Scale return discrimination: p< 0 shows a decrease in scale return; p = 0 shows a fixed scale return; p > 0 shows an increase in scale
Conclusions and recommendations
Fabless firms play an important role in the global semicon-ductor industry, an industry that has been booming over the past 20 years. Authors used the DEA to evaluate the opera-tion performances of the top 30 fabless firms and also used the DEA efficiency value to evaluate cost effectiveness. This model can not only compare and assess efficiency, but it also offers relevant resource allocation and improves manage-ment. DEA is different from the traditional financial achieve-ments or the single comparison index. In rest of this section the authors conclude based on an objective comparison of the top 30 fabless firms and follow that with suggestions and recommendations to improve the performances by reviewing inputs and outputs.
Conclusions
Results of various efficiency rank analyses: The DMU rank coefficient correlation of the A&P and CCR models was 0.959. The DMU rank of the A&P and Cross-Efficiency model was 0.095. The DMU coefficient correlation rank of the Cross-Efficiency model was 0.023. From the lowest value 0.023 of the Spearman rank coefficient correlation of the three models, the rank results of A&P and CCR showed the positive correlation. Regarding discrimination, the CCR model must have the reference times while A&P found the solution just once to obtain the same discrimination ability that the CCR model had.
Results of operation performance evaluations: Both the A&P and CCR models are the results of self-appraisal. The better DMUs among the top 30 fabless firms were Nvidia, Broadcom, ATI, Altera, Marvell, Qlogic, Silicon, OVTI, Genesis, and Ali. The weaker firms with efficiency values smaller than 0.6 were Mega Chips, Conexant, ICS, SST, Xilinx, Cirruss, VIA, Globespan, Sunplus, Qualcomm, and Realtek. The more often a company appears as a reference, the more robust the DMU efficiency gets. The rankings of the reference times were Genesis (8), Ali (8), Nvidia (7), ATI (7), Altera (7), OVTI (6), and Silicon (5).
The Cross-Efficiency model was the result of peer-appraisal. The better DMUs among the top 30 fabless firms were these nine: Qualcomm, Nvidia, ATI, Xilinx, Broadcom, MediaTek, SanDisk, Altera, and Marvell. The worst DMUs (with efficiency values lower than 0.6) were MegaChips, the PMC-Sierra, OVTI, Cirruss Logic, Zoran, Semtech, Genesis Micro, and Ali.
Inefficiency source: From the results of CCR and BCC analysis, one could learn that the inefficiency came totally from a lack of Pure Technological Efficiency. Realtek, PMC-Sierra, Qualcomm, Sunplus, Globespan, VIA, SST, Cirruss, ICS, Conexant, MediaTek, and Novatek are 13 companies with an inefficient application of inputs. The 14 scale
inefficient companies were Xilinx, MegaChips, Semtech, SanDisk, Adaptec, Cirruss, ESS, Aeroflex, SST, Zoran, VIA, Conexant, PMC-Sierra, and Globespan. In addition, authors found six companies, Cirruss, SST, VIA, Conexant, PMC-Sierra, and Globespan, in the case that lacked both Pure Tech-nological Efficiency and Scale Efficiency, The reason behind their lacking of Integrated Technological Efficiency came more from their lacking of Pure Technological Efficiency than from their scale inefficiency; this implied that low tech-nology inefficiency would not affect production scale. How-ever, there were five companies—Xilinx, MegaChips, Sem-tech, SanDisk, and Adaptec—that had more serious problems with scale inefficiency. In the scale inefficiency, their scale returns might increase or decrease progressively.
Analysis of the Scale Returns: From the Scale Returns of the top 30 fabless firms in 2003, authors found that all DMU fall into two types: fixed or progressively increasing. This was surprising because authors expected only one type: progressively increasing. Among the firms studied, the com-panies that increased progressively were Globespan, PMC-Sierra, VIA, Conexant, SST, Zoran, Ali, Semtech, Adaptec, and Genesis. Since the increasing rate of output was greater than input, these firms should expand their Capital Stock, Net Working Capital, and Long-Term Investment expenditures.
Recommendations
From the relevant documents and interviews with business leaders, authors know that every manager pays much atten-tion to evaluating operaatten-tion performance, both internally and externally. On the basis of every company’s orientation and mission, the evaluation also involves the progressive effi-ciency and deferred performance. Thus, it is hard to set up an effective integrated evaluation model and select a suit-able index and fair weight of evaluation. Authors have tried to solve the above-mentioned problems by using the DEA model in this research and have presented objective results that authors hope will help fabless firms conduct better inter-nal evaluations.
In order to promote the operation performance of 20 firms with Fixed Scale Returns, it is necessary to strengthen the efficient application of the input Capital Stock, Net Working Capital, and Long-Term Investments to expand the Revenue and Earnings before Taxes of every company.
Authors found 13 companies with inefficient technology that should strengthen their technical management and devel-opment. Especially regarding input resource application, they should finish the collection and the establishment of input and output data for every company as quickly as possible. This would help those companies carry out further research and evaluation and more accurately adjust their resource appli-cations and distributions.
A ppendix T a ble A .1 Preliminary selection o f input/output items Assessing items/Criteria Data source and reliability Controllable S ame p eriod of time Assessing criteria Performance correlation R&D expense ◦◦ ◦◦ Fix ed assets ◦◦ ◦◦ Intangible assets ◦ Capital stock ◦◦ ◦◦ Cash ◦◦ ◦◦ Net w orking capital ◦◦ ◦◦ Long term in v estments ◦◦ ◦◦ Debt ratio ◦ Re venue ◦◦ ◦◦ Earnings before tax es (EBT) ◦◦ ◦◦ Net income b efore p referred d iv i-dends ◦ ◦◦ Earning p er share ◦◦ ◦◦ Return on common equity(R OE) ◦◦ ◦◦ Return on assets(R O A ) ◦◦ ◦◦ T u rn ove r ra ti o ◦◦ ◦◦ Remark :◦= 100 % m atched, = av erage, X = non-matching those listed abo v e are m atched items
T a ble A .2 2003 global inputs–outputs d ata sheet (5 × 5) Unit: U S$ Thousand F abless /items Input Output R&D expense F ix ed assets Capital stock Net w orking capital L ong-term in v estments R ev enue EBT E PS R O E (%) T u rno v er ratio (%) Qualcomm 523,267 622,265 789,586 2,624,559 1,120,927 2,466,331 1,285,147 1. 05 10.48 27 .96 Nvidia 269,972 190,029 153,513 328,979 190,029 1,822,945 86,673 0. 59 7. 08 130 .27 Broadcom 434,018 142,113 292,009 444,931 41,097 1,610,095 (934,738) (3 .29 ) − 64 .43 79 .80 A T I 248,800 711,100 237,227 92,600 711,100 1,510,992 (280,200) 0. 15 − 180 .08 80 .16 Xilinx 247,609 335,114 337,069 458,805 1,091,697 1,299,900 350,544 0. 37 12.20 44 .25 MediaT ek 119,697 31,775 192,856 673,456 347,266 1,115,931 500,669 0. 79 46.06 90 .66 SanDisk 84,200 59,470 144,781 812,977 185,062 1,079,801 241,881 1. 17 11.25 53 .36 Altera 178,543 160,924 381,387 111,771 14,451 827,207 212,501 0. 41 14 .07 55 .61 Marvell 213,740 149,705 127,456 287,499 0 819,762 63,352 0. 09 2. 08 33 .66 Cone xant 159,354 36,310 303,488 133,734 119,230 633,100 23,433 0. 09 − 422 .94 67 .95 VIA 72,856 57,986 379,295 175,804 368,543 597,664 (50,016) (0 .04 ) 55 .18 68 .34 Qlogic 87,755 67,224 103,473 206,342 0 516,200 215,601 1. 11 15 .41 55 .59 Adaptec 123,022 58,435 106,772 170,487 6,346 437,200 (189,160) (0 .14 ) − 30 .42 41 .59 Globespan 159,354 36,310 268,586 133,734 119,230 379,100 23,433 0. 09 − 422 .94 40 .69 Aerofle x 31,102 69,080 60,193 161,556 0 341,028 12,883 0. 14 2.47 103 .15 Sunplus 45,928 41,883 192,856 116,249 127,658 325,349 60,817 0. 79 14.71 66 .76 Silicon 48,296 30,712 48,850 81,138 0 325,305 66,196 0. 86 15.57 165 .07 Nov atek 17,952 18,130 102,535 80,454 18,933 319,706 64,414 0. 19 39.10 122 .63 SST 43,144 11,325 94,723 175,866 83,046 295,041 (38,751) 0. 06 − 19 .66 74 .44 Realtek 42,984 40,594 197,597 170,718 117,630 272,005 84,613 0. 13 16.23 1. 33 Me g aChips 10,275 7,203 246,610 115,178 1,248 271,297 2,795 0. 07 1.24 163 .08 ICS 35,006 15,749 67,898 163,687 32,000 256,900 71,541 0. 87 22.55 81 .52 PMC-Sierra 119,473 20,750 173,568 198,327 52,905 249,483 (15,843) (0 .05 ) − 3. 53 45 .12 O V TI 15,500 20,622 22,678 254,761 7,110 249,400 89,008 0. 68 19.58 72 .12 Zoran 40,402 20,029 33,231 67,954 0 216,528 (66,615) (2 .05 ) − 12 .72 35 .04 Genesis 30,983 17,257 31,248 48,670 0 213,400 (5,268) (0 .47 ) − 1. 09 51 .96 Cirruss 76,168 22,663 83,445 143,199 6,996 198,200 39,444 (2 .39 ) 21.72 62 .99 ESS 33,184 24,629 39,517 79,313 9,076 195,273 40,894 0. 64 11.14 55 .38 Semtech 30,371 49,579 73,013 124,048 86,119 192,079 42,718 0. 47 8.55 47 .02 Ali 41,512 42,125 51,594 (21,439) 41,730 191,082 152 0. 001 17.00 115 .43
Table A.3 Correlation between input/output items using Pearson correlation coefficients
Output/CC/Input R&D expense Fixed assets Capital stock Net working capital Long-term investments
Revenue 0.907 0.783 0.711 0.731 0.722
EBT 0.183 0.305 0.579 0.709 0.535
EPS −0.166 0.116 0.095 0.210 0.227
ROE −0.208 −0.110 −0.189 0.108 −0.032
ROA −0.206 −0.147 −0.230 −0.239 −0.215
Table A.4 Glossary
Fabless: IC design company DEA: Data envelopment analysis
CCR: DEA model created by Charnes, Cooper and Rhodes A&P: DEA model proposed by Andersen and Petersen BCC: DEA model addressed by Banker, Charnes and Cooper DMUs: Decision management units
OEM: Original equipment manufacturer MCDM: Multi-criteria decision making MOPA: Multi-object programming approach SE: Scale efficiency
TE: Technical efficiency CDMA: Code division multiple access FPGA: Field-programmable gate array WLAN: Wireless local area network IDM: Integrated device manufacturer
CMOS: Complementary metal-oxide-semiconductor LCD: Liquid crystal display
DRAM: Dynamic random access memory DDS: Digital data storage
EBT: Revenue and earnings before taxes DSC: Digital still camera
References
Andersen, P., & Petersen, N. C. (1993). A procedure for ranking efficient units in data envelopment analysis. Management Science, 39(10), 1261–1264.
Banker, R. D., Charnes, A., & Cooper, W. W. (1984). Some models for estimating technical and scale inefficiencies in data envelopment analysis. Management Science, 30(9), 1078–1092.
Brown, M. G., & Svenson, R. A. (1998). Measuring R&D productivity.
Research Technology Management, 41(6), 30–35.
Cavalluzzo, K. S., & Ittner, C. D. (2004). Implementing performance measurement innovations: Evidence from government. Accounting,
Organizations and Society, 29(3–4), 243–267.
Chang, P.-L., & Tsai, C.-T. (2002). Finding the niche position—compe-tition strategy of Taiwan’s IC design industry. Technovation, 22(2), 101–111.
Charnes, A., Clark, T., Cooper, W. W., & Golany, B. (1985). A devel-opmental study of data envelopment analysis in measuring the efficiency of maintenance units in the U.S. air forces. Annals of
Oper-ation Research, 2(1), 95–112.
Charnes, A., & Cooper, W. W. (1984). The non-Archimedean CCR ratio for efficiency analysis:A rejoinder Boyd and Fare. European Journal
of Operational Research, 15, 333–334.
Charnes, A., Cooper, W. W., & Rhodes, E. (1978). Measuring the effi-ciency of decision making units. European Journal of Operational
Research, 2, 429–444.
Chen, Y., & Iqbal Ali, A. (2002). Output–input ratio analysis and DEA frontier. European Journal of Operational Research, 142(3), 1, 476–479.
Chen, Y., Morita, H., & Zhu, J. (2003). Multiplier bounds in DEA via strong complementary slackness condition solution. International
Journal of Production Economics, 86(1), 11–19.
Doyle, J. (1992). MCC:Multiple correlation clustering. International
Journal of Man-Machine Studies, 37, 751–765.
Doyle, J., & Green, R. (1994). Efficiency and cross-efficiency in DEA:Derivations, meanings and uses. Journal of the Operational
Research Society, 45(5), 567–578.
Everitt, B. (1980). Cluster analysis (2nd ed.). London: Halsted Press (Wiley).
Farrell, M. J. (1957). The measurement of productive efficiency. Journal
of the Royal Statistical Society, Series A, General, 120(3), 253–281.
Folan, P., & Browne, J. (2005), A review of performance measurement: Towards performance management. Computers in Industry, 56(7), 663–680.
Golany, B., & Roll, Y. (1989). An application procedure for DEA.
OMEGA, 17(3), 237–250.
Hoque, Z. (2004). A contingency model of the association between strategy, environmental uncertainty and performance measure-ment: Impact on organizational performance. International Business
Review, 13(4), 485–502.
Hung, S.-W., & Yang, C. (2003). The IC fabless industry in Taiwan: Current status and future challenges. Technology in Society, 25(3), 385–402.
Ittner, C. D., Larcker, D. F., & Randall, T. (2003). Performance impli-cations of strategic performance measurement in financial services firms. Accounting, Organizations and Society, 28(7–8), 715–741. Kleine, A. (2004). A general model framework for DEA. Omega, 32(1),
17–23.
Lee, E. S., & Li, R. J. (1993). Fuzzy multiple objective programming and compromise programming with pareto optimum. Fuzzy Sets and
Systems, 53(3), 275–288.
Lewis, H. F., & Sexton, T. R. (2004). Network DEA: Efficiency analy-sis of organizations with complex internal structure. Computers and
Operations Research, 31(9), 1365–1410.
Liu, L.-C., Lee, C., & Tzeng, G.-H. (2003a). DEA for evaluating the current-period and cross-period efficiency of Taiwan’s upgraded technical institutes. In T. Tanino, T. Tanaka, & M. Inuiguchi (Eds.),
Multi-objective programming and goal-programming: Theory and applications (pp. 353–360). Springer.
Liu, L.-C., Lee, C., & Tzeng, G.-H. (2003b). Using DEA of REM and EAM for efficiency assessment of technology institutes upgraded from junior colleges: The case in Taiwan. In T. Tanino, T. Tanaka, & M. Inuiguchi (Eds.) Multi-objective programming
and goal- programming: Theory and applications (pp. 361–366).
Springer.
Liu, L.-C., Lee, C., & Tzeng, G.-H. (2004). DEA approach of the current-period and cross-period efficiency for evaluating the voca-tional education. Internavoca-tional Journal of Information Technology
and Decision Making, 3(2), 1–22.
Ohta, H., & Yamaguchi, T. (1995). Multi-goal programming including fractional goal in consideration of fuzzy solutions. Journal of Japan
Opricovic, S., & Tzeng, G.-H. (2003). Comparing DEA and MCDM method. In T. Tanino, T. Tanaka, & M. Inuiguchi (Eds.),
Multi-objec-tive programming and goal-Programming: Theory and applications
(pp. 227–232). Springer.
Poh, K. L., Ang, B. W., & Bai, F. (2001). A comparative analysis of R&D project evaluation. R&D Management, 31(1), 63–74. Schmitz, J., & Platts, K. W. (2004). Supplier logistics performance
measurement: Indications from a study in the automotive industry.
International Journal of Production Economics, 89(2), 231–243.
Semiconductor Yearbook of 2003. International Technology Informa-tion Services. Industrial Technology Research Institute in Taiwan, pp. 52–61.
Shuai, J.-J., Tzeng, G.-H., & Li, H.-L. (2004), The multi-source fabless-foundry partnership selection model, International Journal
of Manufacturing Technology and Management, 6(1/2), 137–154.
Tzeng, G.-H., Feng, G.-M., & Kang, C.-C. (2001). The fuzzy set theory and DEA model for forecasting production efficiency: Case study for Taipei city bus company. Journal of Advance Computational
Intel-ligences, 5(3), 128–138.
Yu, J.-R., Tzeng, Y.-C., Tzeng, G.-H., Yu, T. Y., & Sheu, H. J. (2004). Fuzzy multiple objective programming approach to DEA with imprecise data, International Journal of Uncertainty, Fuzziness
and Knowledge-Based Systems, 12(5), 591–600.
Zimmermann, H. J. (1978). Fuzzy programming and linear program-ming with several objective functions. Fuzzy Sets and Systems, 1(1), 45–55.
