Analyzing the furnace area: How equipment and scheduling affect cycle time, COO
10/01/1997
Rien A. van Driel, ASM International NV, Bilthoven, The Netherlands
Boudewijn Sluijk, ASM Europe BV, Bilthoven, The Netherlands
Wafer fab owners strive for minimal cycle time and lowest cost of ownership (COO). The choice of furnace and the scheduling strategy used in the furnace area have a significant effect on both these parameters. Since the scheduling strategy giving the lowest COO (only process fully loaded boats) results in the longest cycle time, a compromise has to be found between COO and cycle time. Using a typical production scenario, we calculated cycle time and COO for two standard equipment types — high productivity 100-wafer-batch furnaces and conventional 150-wafer-batch furnaces — and three scheduling strategies. Equipment with high throughput and moderate load size offered the shortest cycle time and lowest COO for most fab situations.
The ASM Modeling and Simulation Tool used for our analysis is used regularly for analyzing proposed fabs for our customers, and for testing different what-if scenarios. The simulation tool, an implementation of the Process Interaction Approach (PIA) developed by the University of Eindhoven, was written using Smalltalk-80, an object-oriented programming environment. The PIA describes a fab in terms of elements (processes) and relationships (interactions).
An actual simulation run has three phases. In the first phase, either real data from an existing fab or expected data for a fab to be built is used to model the fab being simulated. The model incorporates three types of data: fab-specific data — desired production capacity, number of tubes, device types, and corresponding process flows; equipment-specific data — type of process and process time, cool-down times, operator response times, boat load and unload times, number of wafers/load; and computer-aided manufacturing (CAM)-host/server-specific data — the scheduling rules to be applied.
The second phase is running the simulations and collecting the data, and the third phase involves analyzing the data produced by the simulation runs. After building the model, we ran the simulation program for two equipment types and three different scheduling strategies. This article presents the most important data used for the model and analyzes the results.
The model
The model used fab-specific information from a leading device manufacturer. This mix of measured and historical data from the CAM-host reflects a "true-to-life" fab situation. The equipment-specific data for the small batch, high-productivity furnace came from our own sources; the data for conventional furnaces was from trade literature. The three scheduling strategies reflect typical approaches used in the industry.
Wafer fab scenario. Our model was an accurate reproduction of a BiCMOS production line (Table 1). The line has a capacity of 5000 wafer starts/week (for which 38 tubes are needed) and three process flows: a CMOS flow (70% of the wafer starts), a BiCMOS flow (20% of the wafer starts), and a "miscellaneous" flow (10% of the wafer starts), which includes product qualification, test wafer production, and test runs on equipment.
 |
The furnace area was modeled in detail and included cleaning equipment, furnaces, metrology, and operators. The equipment outside the furnace area — litho, etching, metallization, and ion implant — was included as statistically distributed cycle time data.
Scheduling strategies. Three different scheduling strategies were applied to the model. Scheduling strategy 1 was a theoretical minimum COO schedule; processing was only done if the wafer boat was full. Although this strategy gives the lowest COO, it also leads to the longest cycle time because of the waiting times enforced.
Ideally, all fabs try to run with full boats; in practice, waiting times are constrained by some sort of decision mechanism, either based on a set of rules implemented by the factory control system, or determined by the shift manager based on the situation at hand. Scheduling strategy 2 imposed a fixed maximum waiting time of 12 hr on each furnace step. If, after this time period, not enough material was available to fill a boat, then the process was continued anyway. Scheduling strategy 3 imposed a maximum waiting time of 1.5 × the cycle time for that particular step.
Furnace types. We compared two types of equipment: conventional 150-wafer-batch vertical furnaces and 100-wafer-batch modular vertical furnaces. Good examples of conventional 150-wafer-batch furnaces are the products manufactured by the Japanese vendors Tokyo Electron Ltd. (TEL) and Kokusai. For our analysis, we chose a standard TEL Alpha 8 system (Fig. 1).
 Figure 1. TEL's Alpha 8 furnace. |
The second furnace in this analysis, the ASM Advance 400, is a modular vertical furnace designed for high productivity. The system (Fig. 2) has two reactor modules that share a single wafer-handling robot and work in progress (WIP) station. The furnace can be used as a stand-alone 100-wafer-batch furnace with the two reactors operating independently, or as a clustered 100-wafer-batch furnace where the two reactors handle two subsequent process steps.
 Figure 2. ASM's Advance 400 cluster (top view). |
Our analysis considered both the stand-alone and the clustered mode of operation for the Advance 400; in the modeled fab there are three CMOS and five BiCMOS cases where two successive processing steps can be clustered. Although the Advance 400 can also be run with 125- and 150-wafer batches, these variations were not included in the simulation.
Comparison of logistics. It is important to understand the logistic behavior of the equipment and how it interacts with the operators and the CAM host. The furnace types considered display three major differences. The first difference is load size — the conventional furnace processes 150-wafer batches, and the high-productivity furnace processes 100-wafer batches.
The second major difference is single-boat vs. dual-boat operation. The single-boat design of the conventional furnace requires that all operations from wafer loading to wafer unloading be performed in series. The system is forced to wait for cool-down, boat loading, and boat unloading. In contrast, two boats are present in each Advance 400 reactor module. While one boat is in a reactor (processing), the other can be cooling down, or can be loaded or unloaded.
The last major difference between the two furnace types is clustering capability. Clustering allows the ASM furnace to run two processes directly after each other without the need for unload/load or for the wafer batch to leave the controlled environment of the system, and cannot be accomplished in the conventional system.
The two furnace types also have common features. Both have the same basic logistic flow, where an operator loads a cassette into a WIP storage, the appropriate cassette is presented to the wafer-handling robot that loads the wafers into the boat, and then a boat elevator pushes the boat into the reactor. Both systems have WIP storage capacity that allows loading a batch before the previous batch is completed.
Results
The modeling program predicted a total furnace area cycle time of just over 17 days for the conventional furnaces running on the minimum COO schedule. We used this value as a reference value (100%) for comparison with other results. Combined results of the simulation runs (Fig. 3) show that the clustered 100-wafer-batch furnace gives a 40% reduction in cycle time over the conventional 150-wafer-batch furnace running on the minimum COO schedule. The large batch furnaces require the most stringent waiting time scenario to achieve an equivalent cycle time.
 Figure 3. Relative cycle time vs. schedule. |
The strategies for cycle time reduction work: the large batch furnaces can gain more than 40% in cycle time by adopting the third scheduling strategy (where the maximum waiting time may not exceed 1.5 × the cycle time). The small batch furnaces show a smaller effect, but can still achieve a relative improvement of 20% when applying more stringent scheduling rules. Clustering reduces cycle time by a further 3 hr for each clustered process step.
Cost of ownership
However, such dramatic cycle time reduction has a price. The choice of scheduling strategy affects the final COO. A shorter cycle time leads to a smaller WIP inventory and associated savings on capital costs, but this is more than offset by the costs of an increased number of runs, more operators, and running with incomplete loads. These costs are usually calculated using the Sematech COO method. Running with an incomplete load leads to a proportional increase in cost/wafer.
 |
Table 2 shows average load sizes as a result of scheduling strategies 2 and 3 (schedule 1, lowest COO, always has full boats and therefore gives 100% average load size). With both scheduling strategies and both reactor types, the small furnace's average load size is superior to that achieved by the larger batch furnace. Table 2 also includes the cost/wafer relative to the optimum situation. The small furnace is also superior in this area, with costs/wafer that are typically 10% below conventional systems.
Filler wafers
In addition to the direct extra costs that result from less than optimal utilization of the furnace, incomplete loads also use filler wafers. The cost of filler wafers is neglected in the Sematech standard COO calculation but it is significant: filler wafers can only be used for a limited number of runs before they have to be reclaimed, and have to be discarded after only a few reclaim cycles. Fab logistics must ensure constant availability of filler wafers and provide some mechanism for keeping track of and reclaiming them after the appropriate number of runs. Advanced furnaces not only keep the quantity of filler wafers to a minimum, but also use software to ensure the most efficient use of filler wafers.
 |
We used a $40 estimate for the cost of a 200-mm filler wafer, and assumed that a filler wafer must be reclaimed after a limited number of runs (Table 3) to estimate the impact of filler wafers on the overall cost/wafer (Table 4). Although a detailed cost/wafer analysis is beyond the scope of this article, typical cost/wafer figures are around $1–$2/step. Filler wafers will add up to 10–20% to the cost/wafer. Smaller batches offer a considerable cost advantage.
 |
Both furnaces in this analysis can run atmospheric processes with incomplete loads without compromising film quality and control. Filler wafers must, however, be used in LPCVD processes. A look at LPCVD filler wafer costs in terms of absolute cycle times for the two furnaces is revealing: one can achieve a 40% cycle time reduction in the conventional furnaces at an annual cost of $0.6 million; the small batch furnaces start at 40% cycle time reduction at no additional filler wafer cost.
Summary
It is widely assumed that larger batch sizes give lower COO without compromising cycle times. Our analysis, based on reliable data from the industry, points to a very different conclusion: moderate load size and high throughput (dual-boat) equipment give a 40–55% reduction in cycle time, and this reduction is not greatly affected by the scheduling strategy chosen in the fab.
Lowering cycle time leads to additional costs: if the average load size decreases, then the average cost/wafer increases, the number of runs increases, more filler wafers are required, and more operators are needed. A furnace that offers roughly the same throughput with smaller batches can be used more effectively and requires fewer filler wafers. One can even decide to run these furnaces in a minimum COO scenario, a strategy that would lead to unacceptable cycle times for conventional equipment. Clustering leads to a cycle time reduction of around 3 hr for each clustered process step.
Rien A. van Driel received his PhD degree in physics from the State University Utrecht in The Netherlands. He joined ASM International in 1985, and is presently in charge of worldwide modeling and simulation activities. Prior to joining ASM, van Driel worked on equipment developments and construction for high-energy physics experiments at Stanford University and CERN in Geneva, Switzerland. ASM International NV, Rembrandtlaan 7–9, 3723 BG Bilthoven, The Netherlands; ph 31/30-229-8481, fax 31/30-229-3823, e-mail 100336, [email protected].
Boudewijn Sluijk received a degree in physics from Leiden University in 1985, and joined ASM Europe as a researcher working on the company's first furnace cluster products. He has since held a variety of positions in the US and in Europe; he is presently worldwide marketing manager for the Advance 400 series of clustered modular vertical furnaces at ASM Europe. ASM Europe BV, Rembrandtlaan 7-9, 3723 BG Bilthoven, The Netherlands; ph 31/30-229-8411, fax 31/30-229-3823, e-mail [email protected].