The Devil in Foundry APC

Transcription

1 The Devil in Foundry APC Keung Hui, Jason Mou TSMC It is a unique benchmark that more than 2850 different types of semiconductor devices are being concurrently fabricated with a total volume of over 1 million pieces of 8 equivalent wafers every month in the leading foundry. If including active devices not making deliveries within the same accounting periods as well, this number would be magnified a few times more. In addition, this benchmark is being constantly rewritten since the rebound from the financial crisis in With so many different products being fabricated at the same time, it is a formidable logistical nightmare for the process control engineers attending to the daily operations orchestrated by myriad components of the control hierarchy and its infrastructure. Still, offering the same service levels of advanced process controls (APC) is of paramount importance in quality assurances to every customer, big and small. Achieving this operational objective necessarily imposes on the foundry services providers some unique problems not encountered by integrated device makers (IDM). In summary detail below, we examine four critical aspects: high-mix, lowvolume, run-length, group-size, in the search of practicable solutions, the uniquely proven sets of pillars in foundry APC. After two decades of rapid evolutions, what sets foundry services apart from IDM in semiconductor manufacturing today lies in its widely diverse scopes of product coverage, instead of on the levels of technological sophistications, as it once was. As it currently stands, the fundamental difference in scope coverage constitutes a severe logistical problem to foundry services. In the pursuit of its own goals of premium revenue and sustainable development, a foundry operates to satisfy five constraints from its customer expectations: prompt response, ample capacity, flexible capability, excellent quality and minimum cost. Subsequently, effective management of the highly severe logistical problem brings about a tremendous amount of complications and unique difficulties regarding quality performances of APC in foundry. Quantification of the operational differences starts with the numbers of active devices concurrently being fabricated (Figure 1). As the general trend of asset reshuffling away from in-house manufacturing keeps accelerating in the semiconductor business, the total number of devices and combined volumes of delivered wafers keep increasing in the leading foundry. This curve was compiled from historic records over the last decade and it closely reflected FUTURE FAB International Issue 37

2 this business movement. In contrast, prior reports from IDMs listed the number of active devices at around 150,[1] less than 6 percent of the current benchmark of 2850 products. Such a huge gap in the number of products necessitates a structural change in the optimal designs of the control hierarchy and its infrastructures. A successful solution to this high-mix problem holds the key to satisfy the customer expectation of prompt responses. The second aspect of the operational differences penetrates into the distribution profiles of the product-volume compositions (Figure 2). The foundry business model enables extensive realizations of rampant innovations of IC designers and intellectual properties (IP) providers, in all sizes of economic scales. Subsequently, volume of each product comes in drastically differing scales, ranging from a few dozens to thousands of pieces of wafers (Figure 2). Tracking the chronological evolutions of the product-volume compositions over a span of accounting periods, the distributions roughly maintain the same silhouettes. While there are a number of devices finding widespread applications, the absolute majority of all productions are of low-volume types: Around 75 percent of Figure 1. Distributions of Delivered Volumes Against the Number of Products in the Leading Foundry (active devices without delivery within sampling period were not counted)

3 The Devil in Foundry APC all products bring about a final volume of less than 150 pieces of 8 equivalent wafers (six lots at 25 pieces). Only less than 3 percent of all products fill up sufficient volumes to those levels fabricated by IDM. The immediate implication of this structural difference is that foundries effectively operate in continual transient states, instead of reaching some dynamic forms of steady states as with IDMs. The latter condition plays a fundamental role for most process controllers to function properly; subsequently, methods of realizing APC other than the conventional design philosophies are needed for costeffective process controls. Another cross section of the productvolume compositions is the volume-percentage distribution of the devices. The top 100 products alone contributed over 52 percent of all the historic volumes of wafers ever fabricated. Two distinct modes of delivery profiles are observed: large volumes concentrated in short periods, and low volumes intermittently spread over long accounting spans. Clearly, the highrate products are for applications in fashion and the low-rate products are for durables of long life spans. The former products temporarily drive the foundry, such as that of the IDMs, but the majority of intermittent deliveries set the norm of operations in foundries. Being capable of switching between operation modes smoothly becomes crucial to meet the customer expectation of ample capacities. The third difference exposes the wide ranges of the semiconductor devices Figure 2. Chronological Evolutions of Product-Volume Compositions With Low-Volume Dominance FUTURE FAB International Issue 37

4 intended for diverse spectra of analogue and digital applications: clock rates, frequency bands, voltage ratings, power consumptions and other parameters of importance. Subsequently, many more types of processing tools with different engineering capabilities need to be installed than most IDMs need, as the latter are likely to focus on some market niches rather than comprehensively manufacturing all-encompassing products. Compounded with the complexities of reentrant flows of different products, utilizations of the processing tools vary greatly, with non-uniform run lengths as dictated by the high-mix, lowvolume compositions. Snapshots of run lengths of products processed by different tools are pulled over selected periods (Figure 3). Even for high-volume devices, the number of tools available to process an operation step in the lengthy sequences depends critically on the tool constraints, among other factors. Worse still, this situation of severely non-uniform loadings occurs to tool sets even though they may belong to the same tool groups. Subsequently, the APC hierarchy must inherit multiple sets of tactics responding to non-uniform loadings and yet maintain uniform behaviors to guarantee performances. Obviously, non-uniform operating conditions are highly undesirable from the viewpoint of throughputs; nevertheless, having achieved the economies of scale complements the engineering requirements among the tool sets and creates significant imprints on the customer expectation of flexible capabilities. Figure 3. Non-uniform Run Lengths of Control-Threads for High-Volume Products

5 The Devil in Foundry APC The fourth aspect fundamentally differentiates the tool-level realizations of APC in foundries from those in IDMs, rendering many of the conventional practices pioneered by the latter no longer optimal in the former. With high levels of work-inprogress to maximize utilizations for throughputs, metrology delays normally take up to around six lots before measurements become available for feedback to the process controller. Given composition profiles of products shown in Figure 2, clearly, methods other than the conventional thread controls[2] employed by IDMs must be adopted for over 75 percent of those devices having a volume of less than six lots (Figure 4). Practical constraints render sampled measurements useful only to long run lengths of threaded controls in conventional feedback loops. For raging transients due to the high-mix, low-volume situation, some other forms of data feedback or uses of the metrology measurements are needed for these 75 percent of short run-length products, whether or not the control is threaded. A primary and logical strategy to simultaneously resolve the three issues of non-uniform tool loadings, short runlengths and sparse metrology feedbacks tries to superimpose grouping controls of tools, products or a mixture of both. Optimizing the management practices and controls of the functional groups proves as Figure 4. Application Ranges of Thread-Controls Using Conventional Process Controllers FUTURE FAB International Issue 37

6 the most cost-effective to efficiently fulfill the customer expectation of excellent qualities. Lastly, the concerted efforts in constructing the optimization hierarchy with inherent characteristics of high-mix productions, the infrastructure to slide into operation modes accommodating loading surges, the comprehensiveness of engineering diversity in catching technological demands, and the optimal realizations of process control systems in escorting quality assurances aligns synergistically toward the business goal of premium revenues to the foundry and the customer expectation of minimum costs. The construction process, however, is full of details with too many uncharted territories where the devil looms. At the lower spectrum in realizing APC in foundries and for which gargantuan volumes of process and tool data are crunched at high frequencies, the most hideous pitfall is the assumption of ergodicity, one detail too often enshrouded behind thick veils and harmfully taken for granted by most data analysts in the designs of APC applications.[3] In conclusion, we tried to quantify the operational differences between IDMs and semiconductor foundries over APC applications from four aspects in relation to the five customer expectations. The original article was motivated by an exposition on temptations of elaborate statistical constructions without cautious alerts to the underlying assumptions for APC applications to function properly. The revelation hopes to draw the attention of practitioners on building more robust process and equipment models for quality control purposes. Otherwise, the devil of foundry APC lurks in the absence of ergodicity in the system dynamics, rendering the logistical problem not controllable in effective and efficient manners. References 1. John Schmitz, View on Advanced Control Techniques from a Wafer Fab Manager s Perspective, 3rd European AEC/APC Conference, Dresden, Germany (2002) 2. C.A. Bode, J. Wang, Q.P. He, T.F. Edgar, Run-to-run control and state estimation in high-mix semiconductor manufacturing, Annual Reviews in Control v. 31, pp (2007) 3. K. Hui, J.I. Mou, The Devil in Foundry APC, AEC/APC Symposium XXII, Austin, Texas (2010) About the Authors Keung Hui Keung Hui is a technical manager with TSMC. He holds a Ph.D. in control engineering with interests in modeling, simulations and optimizations. Dr. Hui is the program chair of many process control symposia in Taiwan. Jason Mou Jason Mou is a deputy director responsible for the development of APC solutions in the Manufacturing Technology Center of TSMC. He holds a Ph.D. in engineering and has held various educational and executive positions.