以非線性成長曲線模型預測全球Fabless之晶圓需求量

Abstract

半導體是一個廣為周知的高度資本和技術密集的產業。因此，對於擁有晶圓製造產能的半導體公司，特別是晶圓代工廠，準確的需求預測模型是至關重要的決定性成功關鍵因素。在本篇論文中，我們利用廣泛應用於各種產業研究卻鮮少應用在半導體產業的預測工具「非線性成長曲線預測模型」，作為主要的研究方法以預測晶圓代工產業中來自Fabless晶圓需求。 根據過去2002年〜2008年的平均比例，晶圓代工產業的營業收入來自於Fabless(63％)，整合元件製造廠(IDM，31％)和系統公司(6％)。因此，一個好的Fabless預測模型己經決定了晶圓代工產能的大部份需求。 本研究重點在Fabless的晶圓需求預測。首先，我們回溯到1980∼2009年全球半導體每年營收，及相對應1988∼2009年Fabless年營收以計算Fabless對於半導體產業的滲透率。接下來，我們採用三種不同的非線性成長曲線模型(Logistic，Gompertz和Extedned Logistic)和四個不同的測量方法─（決定係數(R2)、合計平方誤差(SSE)、誤差平方根(RMSE)，及平均相對誤差 (MRE)）來評估實際值及模型的配適能力。然後，分別利用過去1988〜2003年和1988〜1999年間的歷史資料來預測5年(2003〜2009年)和10年(2000〜2009)的結果。最後，我們採用誤差平均絕對比率(MAPE)來衡量各模型預測的能力。依照上述的法則，驗證後找到最佳的模型，我們可以用來針對Fabless給予晶圓代工的需求得到更佳的預測結果。 本論文主要研究結論：1. Gompertz是最適用於半導體產業的預測模型。2. Logistic是最適用於Fabless在半導體產業滲透率的預測模型。根據上述研究結果，可以合理的預測在2020年，全球半導體營收將達到3,450億美金，複合成長率為4% (2009~2020)。而Fabless營收則達到 910億美元，複合成長率達 5％ (2009~2020)，代表了Fabless對於晶圓代工的需求是3700萬片8吋約當量。

Semiconductor is well known as highly capital and technology intensive industry. Therefore, a reasonably accurate demand forecast modeling is vitally decisive factor for the success of a semiconductor company with owned wafer manufacturing capacity, especially for semiconductor foundry industry. In this thesis, we tentatively adopted “Non-linear growth curve prediction model”, a widely applied forecasting tool in variety industries while rarely chosen for semiconductor, as major methodology to forecast Fabless wafer demand to foundry industry. According to year 2002~2008 historical average ratio, source of foundry revenue consist of fabless (63%), IDM (Integrated Devices Manufacturer, 31%) and System Companies (6%). Therefore, a good forecast model for fabless has decided most of foundry wafer capacity requirements. This research focus on aggregate fabless companies’ wafer demand forecast. Firstly, we back traced global semiconductor annual revenues from 1980 to 2009, and corresponding annual fabless revenues from 1988 to 2009 to calculate annual fabless share or penetration rate of semiconductor. In the next, we applied three different non-linear growth curve models (Logistic, Gompertz and Extended Logistic), and four different measurement methods (Coefficient of Determination (R2), Sum Squares Error (SSE), Root Mean Squares Error (RMSE), and Mean Relative Error (MRE)), to assess fitting quality between actual and model. Then, we developed 5-year (2003~2009) and 10-year (2000~2009) forecasts based on historical period of 1988~2003 and 1988~1999 respectively. Finally, we applied Mean Absolute Percentage Error (MAPE) to measure predicting quality of this model. With discovery of the besting fitting model, we can better forecast fabless wafer demand forecast to foundry. Our key findings are: 1. Gompertz model is the best for semiconductor forecast, 2. Logistic model is the best for fabless penetration rate of semiconductor forecast. According to our findings, we can forecast global semiconductor revenue will reach US$345 billion in 2020, with 4% compound annual growth rate from 2009 to 2020. Also, Fabless revenue will reach US$91 billion in 2020 with surpassing compound annual growth rate of 5% in the same period, and this represents 37 millions of 8” equivalent wafers demand to foundry in 2020.