Lessons From the Semiconductor Industry
By Arnold B. Maltz, William L. Grenoble, Dale S. Rogers, Robert M. Baseman, William Grey, and Kaan K. Katircioglu -- Supply Chain Management Review, 11/1/2000
Semiconductor products are a fundamental building block of the new information economy. In 1998, semiconductor sales were $126 billion on 257 billion units shipped, and the industry experienced double-digit growth rates in 1999 and 2000. Semiconductors now are used in a wide variety of industries—from computers to communications equipment to motor vehicles and industrial controls. As the importance of being "connected" continues to grow, semiconductors will be needed literally everywhere.
Semiconductors are also prototypical new-economy products for supply chain managers. Chips are small, extremely valuable, quick to become obsolete, and globally produced and distributed. Customers measure product lifecycles in months. Inventory is precious early in a product's lifecycle, but its contribution to profit can drop by 80 percent in a year. Faced with an absolute need for speed, chip manufacturers use air freight as the default transportation mode, except where surface delivery is as fast or faster. Shipping across oceans, accommodating extraordinary growth, and supplying key customers overnight are routine challenges in the high-technology world of semiconductor logistics.
Until recently, these supply chain concerns were not on the executive radar screen. From its inception, the semiconductor business has been engineering and product driven. The essence of the business is Moore's Law—that is, every 18 months performance capabilities double, putting continuous pressure on prices for older chips. Cutting-edge technology and "design wins" traditionally have been regarded as the keys to success. But now chip customers also are insisting on better service. Having the latest, fastest chip is not the only key to competitive advantage.
Consider the following examples. Dell has built its business around expecting suppliers to deliver, and other users are following suit. Gateway recently penalized Intel for non-performance by shifting major amounts of business to Advanced Micro Devices (AMD).1 Some companies report that automotive customers like Ford expect just-in-time (JIT) delivery schedules and refuse to share the risk of product obsolescence. Supply chain realities have descended on the semiconductor world with a vengeance.
Faced with these pressures, the International Sematech Semiconductor Logistics Forum, a group of semiconductor manufacturers that meets to address industrywide issues, initiated a study of logistics practices throughout the manufacturing and delivery process. Our research team, made up of professionals from three universities and IBM's Thomas J. Watson Research Center, analyzed detailed service, location, and shipment data for seven global semiconductor companies.2 We looked at the challenges affecting semiconductor logistics, how the industry was meeting those challenges, and what strategies companies were using to improve performance. The findings from this work can be condensed into six key lessons that are applicable to any company serving the fast-moving global marketplace. (The sidebar on page 49 summarizes these lessons.)
The Operating EnvironmentThe operating environment for semiconductor logistics is influenced by three elements—the increasing level of customer expectations, the dispersed nature of the manufacturing network, and the sky-high cost of inventory. Understanding the interplay between these factors is fundamental to grasping the problems and potential of supply chain management for semiconductor manufacturers.
Customer Expectations—High and Rising
One of the research team's first tasks was to investigate customer expectations. We probed both manufacturers' perceptions of their customers and how those perceptions translated into planned service levels.
Lesson 1: Where customer service is involved, industry matters. Customers that use chips in their products insist on high service levels even at high cost; resellers accept lower service levels for lower cost.
The service model offered to customers is only part of the story, however. Different customers within the same classification may receive different service levels. For example, although one manufacturer has several "preferred" communications customers within the United States, these customers do not all receive the same service levels. Instead, delivery times range from one to four days. Service levels to consumer-electronics customers also see regional variations. Most of the customers within this sector are in Asia, meaning that even non-preferred Asian customers receive next-day service because they are close to semiconductor manufacturers.
Geography also plays a role in service to component customers. For two of the manufacturers, component demand is in the United States, but the components are shipped directly from Asian plants. For these non-preferred customers, even air freight requires a minimum of three days. (By contrast, some non-preferred customers in Asia routinely get overnight service.) Another manufacturer rates most component customers "preferred" but uses surface transport in Europe and the United States, which results in three- and five-day delivery cycles. For most manufacturers, preferred customers get one- to two-day service. These variations lead to our second lesson:
Lesson 2: Global companies may have uniform customer classifications, but these do not necessarily translate into global service standards.
A Far-Flung Manufacturing Network
The second lesson shows that the distance between the semiconductor facility and the customer significantly affects the service level. To understand why semiconductor companies locate their facilities where they do and how that affects their logistics service, we looked at the manufacturing networks. Semiconductor manufacturing and delivery is a three-stage process, typically referred to in the industry as wafer fab, bond/assembly/test, and product distribution.
Wafer fab (short for wafer fabrication) is the process wherein multiple integrated circuits are fabricated on each raw silicon wafer. This highly repetitive process also is referred to as the front end. A layer of material is deposited on the wafer, a specific pattern is created in the layer, another layer is added and patterned, and so on. Wafer fab plants are extremely expensive, costing $2 billion each. This level of capital investment mandates round-the-clock operation for these installations. Even so, the physics of the fabrication process sets a lower limit of 30 days for cycle time from raw wafer to finished wafer. More typical times are 60-plus days to make sure the fabs always are operating at optimum load.
Bond/assembly/test (BAT) facilities dissect finished wafers into separate chips, assemble the devices into various packages, and make the appropriate electrical connections. The facilities then test the packaged devices to make sure that they are operating according to customer specification. The "assembly" stage of this process has a significant labor component, while final testing is performed using sophisticated and expensive automatic test equipment. Bond/assembly/test often is referred to as back-end processing because the output of this phase is a finished product ready for shipment to customers. The output unit of BAT typically is a single packaged integrated circuit.
Product distribution in the semiconductor industry refers to the warehousing and shipping of finished goods. Because the product is light, small, and of relatively high value (selling prices range from $500 per unit for microprocessors to $0.05 per unit for discrete components, with an average of $0.49 per unit overall in 1998), many semiconductor distribution centers have automated storage and retrieval systems. The short life cycles and high cost of inventory push most manufacturers to minimize final storage locations, and outsourcing of warehousing is common. On the other hand, certain customers demand just-in-time delivery on four hours' or less notice. These customers are serviced using small satellite locations close to the customer assembly plants.
Given this production process, the location decisions for these networks reflect the interplay between three factors—the need for technical expertise, the labor-intensive nature of assembly and test, and the importance of being close to customers. Local incentives and favorable tax treatment have also influenced assembly and test location choices.4 So although logistics and transportation costs factor into these decisions, they typically are not large enough to outweigh the main drivers of the decision. The location decision turns out to be different, based on the type of facility being considered—wafer fabrication, BAT, or product distribution. (Exhibit 3 shows the site-selection criteria used for different types of facilities.)
Most wafer fabrication plants are located in relatively advanced economies—the United States, Europe, and Japan. Fabs are true high technology with extremely high capital costs and a level of sophistication that makes technically trained personnel a necessity. Historically, the great majority of fabs have been located in countries that have large supplies of engineers, technicians, and other technologically proficient workers as well as a reliable physical infrastructure. In some cases, the companies profiled here have fabs in less economically developed areas, but this remains the exception. All of our study companies maintain some wafer fabrication operations in the United States.
BAT is concentrated in Asian countries other than Japan. The bond, assembly, and test process requires significant amounts of labor, which can range as high as 20 percent of the total cost. Chip companies early on went to low labor cost areas for these operations. Recently some companies also have located BAT facilities in such places as South America and North Africa.
Product distribution activities are evenly spread among the three geographies. The decision here is how much to pay in extra transportation and inventory cost to be close to the customer. The industry operates both Tier 1 and Tier 2 distribution centers. Tier 1 sites receive finished products from BAT facilities and ship to customers for final use. Tier 2 sites receive inventory from Tier 1 centers and ship to close customers as necessary. Tier 2 locations are frequently dedicated to important JIT customers. The number of Tier 2 sites varies depending on inventory availability and customer demands. Where scarce new products are involved, both buyer and supplier prefer not to spread finished-goods inventory too thin.
In summary, semiconductor manufacturing is truly a worldwide enterprise, with the heaviest concentration of activity in North America, Western Europe, and Japan. The need for expertise and a reliable infrastructure means that fabs historically were located in advanced countries. These fabs would be ruinously expensive to move. Bond/assembly/test activity, historically difficult to automate and relatively labor intensive, is now entrenched in southern Asia. Finished-goods inventory is located consistent with service requirements. Powerful customers like Dell can dictate that so-called Tier 2 facilities be located near their plants. In a few cases, the chip manufacturers will literally put the semiconductors in the assembler's plant.
Lesson 3: For high-tech industries in the new global economy, logistics may have relatively little influence on location decisions. But these decisions can drive companies to adopt high-cost transportation alternatives.
The Soaring Cost of Inventory
The semiconductor industry is not the only one with demanding customers and a globally dispersed production process. But semiconductors are almost as perishable as fresh food or designer fashions, a fact that makes inventory expensive. To understand semiconductor logistics, it is important to realize how costly holding inventory can be.
Semiconductors are good examples of the "innovative products" Marshall Fisher wrote about in his 1997 article, "What Is the Right Supply Chain for Your Product?"5 Fisher's examples were drawn from consumer goods such as ski fashions. He pointed out that the cardinal sin for those selling innovative products was to allow a stockout to occur in the early stages of product introduction, when prices were high and demand was strong. He recommended a flexible supply chain with an emphasis on improved forecasting and quick replenishment of stocks when consumer fashion choice revealed itself.
That prescription holds when retailers are handling fashion goods and customers are disappointed one at a time. But what happens when your customers are manufacturers practicing mass customization, offering thousands of configurations on demand? Now the consequences of a stockout are not a few lost sales but rather whole factories brought to a halt. Worse, industrial purchase decisions are made infrequently and involve large commitments. Just ask Intel about Gateway's decision to second source with AMD. Being out of stock is not a viable option; in these markets, business is lost in chunks, not one sale at a time.
Still, being "in stock" is not cheap. Semiconductor prices depend on many factors, in particular the balance between capacity and demand and the position of the specific product in its lifecycle. The example in the sidebar on page 47 assumes that chip prices decrease at the same rate as higher-performance chips become available—in other words, Moore's Law. Although this is clearly a gross simplification, it illustrates the double impact of holding products with short cycle times. Not only are there costs to finance the inventory, but the potential return from the inventory investment also decreases on a daily basis.
Going beyond inventory, how much can the industry afford to spend on freight? A typical semiconductor company spends up to two percent of sales on freight and warehousing. In the example shown in the sidebar, reducing cycle time by one week would pay for a 35-percent increase in freight and warehousing expense. Specifically, gross profits decrease weekly by ($0.80/52) = $0.015. Weekly finance costs amount to ($0.15/52) = $0.003. At two percent of sales, freight and distribution comes to $0.0434. Reducing inventory by one week saves $0.0153. Because $0.0153/$0.0434 = 35 percent, one week of inventory savings offsets a 35-percent increase in logistics costs.
So a careful look at inventory costs leads to the following lesson:
Lesson 4: In fast-paced global industries, using inventory to improve service is prohibitively expensive and, for new products, simply impossible. Fast transportation, improved information, and better processes are the only realistic alternatives.
A Performance Report Card
Semiconductor logistics, then, is not for the faint of heart. Short product lifecycles, rapid price declines, demanding customers, highly dispersed facilities ... any one of these would cause logistics problems. All four together make for an extremely challenging environment. How well are the semiconductor manufacturers managing their supply chains in the face of these challenges? As usual, the answer depends on your perspective. We collected data on two aspects of supply chain performance—on-time delivery and cycle times. Mindful of the importance of sheer distance, we gathered the data at the origin-destination pair level. The results reflect the global nature of company supply chains, the mobility (that is, high value/light weight) of semiconductor materials, and customer demands for speed of delivery.
On-Time Delivery
Exhibit 5 shows the industry's delivery performance against two accepted standards—customer request date and manufacturer commit date. In most cases, late deliveries resulted from the company's not having the product in time to fulfill the customer's request. As one executive noted, the tough part of semiconductor logistics is not moving the product; it's planning enough capacity to make the product.
Where customers have choices, larger customers (for example, Compaq, Dell, and Gateway in PCs; and Qualcomm, Cisco, Nokia, and Ericsson in communications) may apply real penalties for non-performance. These customers also are in "technology/fashion" businesses with short lifecycle products and high growth curves. They increasingly demand just-in-time deliveries or even line-side stocking. In some cases, customers have contracted manufacturing to outside specialists operating on very thin margins. If several chipmakers appear to offer equivalent technology, either the customers or their manufacturing partners will press to switch business to suppliers that can perform on service. In fact, one semiconductor maker stated that "good/bad service ... may be the primary reason for customer turnover."
The mismatch between manufacturing cycle times and customer planning horizons is a familiar problem for logistics professionals. There are two general approaches to "fixing" the resulting customer-service problems—holding higher inventories or operating on shorter cycle times. As we noted above, holding more inventory/safety stock is prohibitively expensive for the semiconductor industry. So the industry is working hard to emulate customers, shorten cycle time, and improve process flexibility.6 Semiconductor companies are striving to pattern themselves after high-profile success stories such as Dell's assemble-to-order operations and Hewlett-Packard's postponement program for printers.
Cycle Times and the Geography of Global Service
Our data suggest that shorter cycle times will have to come from design or manufacturing, the parts of the supply chain that precede the final shipment to the customer. Transportation times to customers already are near the physical limits for speed. What logistics cycle times can a global industry deliver? What cycle times should it deliver? In semiconductors, this is determined by a combination of geography and customer requirements. We first looked at overall performance levels, then separated out the influence of geography, customer type, and service model variations on industry performance.
Semiconductors are delivered fast, if they are in stock. In our sample, 40 percent of all shipments were delivered the next day for data-processing, consumer electronics, and transportation customers. At least two-thirds of all customer shipments were delivered within two days, with the exception of shipments to resellers. Only 5 percent of shipments (again excepting reseller customers) had delivery times as long as five days.
Manufacturers have set up distributed networks to support this service level. Overnight shipments originate in the customer's country 90 percent of the time. European companies buying within the European Union also can get overnight service, as can U.S. customers buying in Canada (and vice versa). These service levels hold whether air or surface transportation is used. Conversations with semiconductor manufacturers confirm that if a border has to be crossed, one-day service is not a realistic option.
Locating bond/assembly/test facilities in Asia has had an interesting result in terms of service. Because BAT facilities are closer to them, Asian buyers of semiconductors enjoy, on average, one-day-shorter delivery times compared with U.S. and European customers. For example, Japanese data-processing customers get their chips in 1.1 days, U.S. data-processing customers wait 2.1 days, and European customers wait 2.9 days. Similarly, Asian component and consumer electronics companies also enjoy 1 to 1.5 days' faster delivery than the rest of the world. The only exceptions are (1) transportation customers, where cycle times are roughly equal across geographies, and (2) European communications customers, which get almost the same service as their Asian competitors do.
One other related point should be noted. Terms of sale for domestic North American shipments differ from those in Europe and Asia. In North America, freight is paid by the customer, which may be sensitive to the extra cost of overnight delivery. In Europe and Asia, the manufacturer pays for delivery, and overnight service is standard practice. Reseller customers, which are concentrated in the United States, need to budget for and pay transportation costs. This is probably one of the reasons they accept longer transportation times.
Lesson 5: Geography matters, even when high-velocity transportation is used. In particular, next-day service usually requires in-country (or at least in-customs-union) inventory.
Both geography and customer expectations play a role in determining delivery times for semiconductors. Most customers receive two-day service or better, although component customers get significantly worse service. Asian semiconductor users have a one-day advantage over the rest of the world, except for automotive customers.7 In general, one-day service requires a shipping point within the customer's nation, suggesting that borders remain a hindrance to fast delivery. Customers working under the highest service model—enhanced service—get the fastest service. But the actual service levels also depend on the extent of the manufacturer's network and the details of both shipper and receiver location.
Shortening the Semiconductor PipelineOf course, fast service is costly. But in the semiconductor world, inventory and obsolescence are dominant parts of the cost mix. Faster deliveries may actually be cheaper on a total cost basis. As noted above, reducing leadtimes by a week may save enough to finance a 35-percent increase in logistics costs.
How much room do the manufacturers have for improving cycle times? The manufacturers indicated that front-end (wafer fab) cycle times were very dependent on the specific types of chips being made. Cycle times at back-end operations—bond/assembly/test and product distribution—are not as affected by product mix. In other words, decreasing fab cycle times will probably require new product designs, while changes in logistics processes by themselves can lower post-fab inventory and throughput times. Accordingly, we estimated the average cycle time for each manufacturer based on reported finished-goods inventory at each site and customer shipments from each site. We found that front-end cycle times varied from 35 to 93 days depending on the product and the age of the fab technology. The industry composite was 76 days. Post fab, or back-end, cycle times varied by company from 20 days to 51 days, depending on the network structure, market volatility, and the availability of capacity. The industry composite for back-end time to customer was 41 days.
The majority of cycle time is taken up in the complicated front-end processes where the wafers are processed to create chips from raw materials. But there is clearly room for improvement in the three to four weeks of processing time used by the bond/assembly/test and product distribution activities.
The participants in our study indicated at least three major approaches to shortening logistics cycle time. First, companies are working on postponement strategies. Borrowing techniques from their customers such as Dell and Hewlett-Packard, these companies are looking for ways to avoid committing finished inventory until firm orders are in hand. Second, companies are reorganizing to eliminate process delays that add to cycle time. Finally, semiconductor manufacturers are adopting information technology, including enterprise resource planning (ERP) systems and other tools that provide inventory visibility and better planning capabilities. Examples of each strategy follow.
Postponement
One semiconductor manufacturer operates four regional distribution centers for shipping final product to customers. These end-shipping points have become more involved in customizing outbound shipments by furnishing and applying unique shipping labels and bar codes. Because nearly every customer has different requirements, this customization activity amounts to postponing the finishing process until firm orders are available.
Another manufacturer has changed bond/assembly/test procedures to postpone inventory commitment and thus reduce inventory and cycle times. This company has set up three "die banks" where integrated circuits are stored in the Far East close to assembly and test operations. Undifferentiated product is "pushed" from U.S. fabs into these storage facilities and is "pulled" by customer orders when required. This staging of the input to BAT has resulted in reductions of up to five days in "back-end" total processing time.
Reorganization
Pressure to reduce order-cycle time is constant in today's semiconductor marketplace. One producer has addressed this pressure by adding logistics managers to cross-functional customer teams. Supported by top management, the teams are encouraged and empowered to negotiate and work out process improvements with customers. Recently, this company received a request from an important original equipment manufacturer (OEM) customer to reduce leadtime from 100 days to 70 for a class of high-tech products. Reducing manufacturing time was difficult, so the team attacked distribution time first. It found several places to shorten the process, including making direct deliveries to a contract manufacturer used by the OEM. Although the 30-day target has not yet been achieved, the company has made significant progress and is continually seeking ways to get faster.
Information Technology
Semiconductor manufacturers believe that better information technology will improve customer service and reduce logistics costs. When asked to identify the most important future logistics trends, four companies emphasized the demand for more information, automation, and connectivity throughout the supply chain.
Unfortunately, logistics and supply chain management are not always priorities for IT investment. But there have been successes that can point the way to improved global logistics capabilities. For example, one company implemented several modules of an ERP package to cope with both growth and Y2K issues. Implementation of the modules for order processing, inventory management, and production planning took approximately two years. The implementation was phased in by product line and location. Systems supporting transportation and warehousing initially were left untouched except for interfaces to the corporate ERP package and necessary adjustments such as naming conventions and lot-adjustment capabilities. Nevertheless, this company found that customer service (as measured by product availability) increased from the low 70 percent range to the high 80s, based on line-fill rates. Inventory visibility really does matter.
The initiatives from these three companies all share a common goal: better control of scheduling to shorten the semiconductor supply chain. This can be summarized in the final lesson learned.
Lesson 6: For short-lifecycle products that are manufactured and sold globally, scheduling discipline is critical at all points in the supply chain all the way through to the customer.
Logistics in the New Economy
Logistics and supply chain managers are learning to cope with demands that would have been unthinkable 10 or 15 years ago. Four-hour response times, contractual requirements for upside flexibility, and calls for vendor-managed inventory are becoming routine if you are to do business with electronics manufacturers. Geography still matters, even when airfreight transportation is automatic. Global reach notwithstanding, overnight delivery across borders is still problematic. Companies segment customers by service model, but this does not translate into global service standards. Inventory costs and price erosion, already nearly 80 percent annually, will increase as product lifecycles get shorter. Recognizing the physical limits to manufacturing and transportation speeds, companies have to count on planning and scheduling innovations for further supply chain improvements.
In the final analysis, semiconductor logistics is a prototype for what may very well be the dominant industrial model in the coming century. Small, high-value, short-lifecycle items will be produced wherever the total costs are lowest and sold on demand to customers that are directly fronting the final user market, be that consumer or business.
Our study suggests that some gaps will need to be filled in before this vision can be realized. Better data, especially cost data, will be critical. Models will encompass traditional manufacturing and logistics costs as well as true inventory costs. This should lead to intelligent location decisions. Activities will be spread throughout the supply chain based on cost and real needs for proximity and technology. Postponement strategies will be used to minimize the problems caused by long production times vs. short customer leadtimes. Border frictions have to be eliminated, and manufacturers must stand ready to manage their customers' component inventory, if necessary.
Underlying all of these efforts will be information—and that is perhaps the most important lesson we learned. For all their sophistication and positioning in the forefront of the technology revolution, the semiconductor manufacturers participating in this study did not have the information or the processes in place to monitor network and allocation decisions continuously.
In a number of cases, accurate shipment counts by site were unavailable. One company accepted a single number for worldwide operations because it had a single worldwide delivery process. For this company, finding regional details proved to be difficult. Most companies were working to improve their logistics analyses through implementing decision-support tools, and some now are testing these tools or actually using them. But supporting a global customer base through a worldwide supply chain remains a goal to be pursued even for those leading the race toward a high-tech world.
Authors' note: The authors would like to thank the International Sematech Semiconductor Logistics Forum and the Thomas J. Watson Research Center at IBM for their support of this work. All authors contributed equally to the research. The authors also acknowledge the research assistance of the following students at the Center for Logistics Research at the University of Nevada, Reno: Yi Zhang, Kris Narayanan, Sam Anderson, Shannon Thawley, and Tim Johnson. Questions should be directed to Arnold Maltz, Associate Professor, Supply Chain Management, Arizona State University. Telephone: (480) 965-9768; E-mail: arnie.maltz@asu.edu
| Author Information |
| Arnold B. Maltz is an associate professor of supply chain management at Arizona State University. William L. Grenoble is the administrative director and research associate for the Center for Logistics Research at Pennsylvania State University. Dale S. Rogers is a professor of supply chain management and the director of the Center for Logistics Management at the University of Nevada, Reno. Robert M. Baseman, William Grey, and Kaan K. Katircioglu are with IBM's Thomas J. Watson Research Center. |
| Footnotes |
| 1 Takahashi, Dean, "Gateway to Use Athlon Processors of AMD in PCs," Wall Street Journal, Jan. 10, 2000, p. 1. |
| 2 The participating members of the Semiconductor Logistics Forum were Advanced Micro Devices, Intel, National Semiconductor, and the semiconductor divisions of Texas Instruments, IBM, Hewlett-Packard (now Agilent), and Motorola (including the since spun-off On Semiconductor group). |
| 3 Line-side stocking is a practice whereby the supplier delivers components directly to the point of use, while maintaining ownership until the goods are drawn into the production process. |
| 4 Wilson, Peter G., "The Role of Taxes in Location and Sourcing Decisions," in Studies in International Taxation, Alberto Giovanni, R. Glenn Hubbard, and Joel Slemrod, ed., National Bureau of Economic Research, 1993. |
| 5 Fisher, Marshall, "What Is the Right Supply Chain for Your Product?" Harvard Business Review, May–June, 1997. |
| 6 Fisher, op. cit. |
| 7 Although we did not ask for individual customer identification, it appears that manufacturers in our sample did not ship to Japanese auto companies in Japan. |
| |||||||||||||||||||||||||||||||||||||||||||||||||||||
View All Blogs
