Prize-Winning Planning at Harris Semiconductors

This article is based on the paper by Robert Leachman and his colleagues at the University of California at Berkeley which won the 1995 Franz Edelman prize for Management Science Achievement.

Harris Corporation

Harris Corporation is a US electronics company, one of whose divisions manufactures semiconductors at plants in the US and Asia.

Through a series of takeovers in the late 1980s, it tripled in size and moved out of a profitable niche in defense into a wider and more competitive market where on-time delivery was crucial. In 1989 only 75% of items were delivered on time, 60% of customers wished to replace Harris as a vendor and $100 million of potential sales were lost as a result.

Senior executives took action. They commissioned a survey which found that production planning and delivery quotation was a mess. Two large MRP (Manufacturing Resource Planning) systems were in use along with many other smaller spreadsheet analyses. Delivery quotations were made largely by judgements using incomplete data. Rival semiconductor manufacturers had company-wide production planning systems based on MRP technology but adapted for the special features of the semiconductor industry.

Since 1984 Harris had been a member of a consortium of semiconductor manufacturers which had been supporting the development of an optimization-based production planning system at the University of California at Berkeley. Harris had implemented a prototype at its Florida plant and had been pleased with the results. Now it decided to expand the system company-wide. Thus IMPReSS was born: the Integrated Manufacturing Production Requirements Scheduling System.

Why MRP Won't Work

The problem which Harris faced is normally tackled using MRP technology. There are two special features of semiconductor manufacturing which make classic MRP inadequate:

  • 100% utilisation of key equipment on repeated visits;
  • binning and substitution.

Semiconductor manufacturing is capital-intensive and plant operates 24 hours a day, 365 days a year. Demand often exceeds production capacity. It is not possible to overcome capacity constraints by scheduling overtime. Instead orders for the less-valuable products must be declined or given long delivery times.

The production process itself involves repeated visits (as many as 20) to the key pieces of equipment over a period of several weeks. These visits are interspersed with other manufacturing processes. The scheduling of the key equipment, which is running at 100% utilisation, is challenging, particularly when combined with deadlines for finished products.

Binning is a practice peculiar to semiconductors. We are all familiar with different speeds of CPU chips. How, we might ask, does Intel manufacture 166 MHz Pentiums as opposed to 133 MHz or 120 MHz chips? The answer is: it tries to manufacture 166 MHz chips and then tests chips individually to see how well they work. If they work perfectly at 166 MHz they are put in the 166 MHz bin. If they fail at 166 MHz but work perfectly at 133 MHz they are put in the 133 MHz bin; and so on.

A corollary to binning is that if there is a surge in demand for 133 MHz chips while demand for 166 MHz chips remains low (because the price is too high), this can be met by substituting 166 MHz chips for the 133 MHz chips requested. This is shown in Figure 1.

Figure 1: Binning and Substitution of Chips

The proportion of chips which will be assigned to each bin can be predicted with a fair degree of confidence. In principle, therefore, one could work back from the demand to calculate the number of chips to be manufactured and apply classic MRP to that. But this is to ignore the economics and the capacity constraints on the plant. Binning and substitution must be taken into account in deciding whether to accept an order and what delivery time to quote. There is a genuine optimization problem in deciding how to maximize revenues given the capacity limits of the plant and the prices and binning characteristics of the chips concerned.

How IMPReSS Works

The full-scale problem which Harris faces extends to 10,000 product lines using 200 types of equipment and 200 categories of raw material over a period of 18 months. If this were formulated as a single optimization problem it would require more than a million variables and half a million constraints, even using a gradation of weekly, monthly and quarterly time periods.

To make the problem tractable, it is decomposed in a way which reflects the manufacturing process and where the bottlenecks lie. The planning process is driven by demands, which have to be estimated for products where there are not firm orders.

The stages of semiconductor manufacturing are shown in Figure 2 along with a schematic of how IMPReSS works.

Figure 2: Semiconductor Manufacture and IMPReSS

There are five modules which are run in the following order:

  1. The Test Requirements Planner considers each product family separately. It uses MRP techniques to work back from demands to calculate the number of final test starts for each product at each plant and their average expected revenue.
  2. The Die Requirements Planner also considers each product family separately. It performs a mix of Linear Programming (LP) and MRP calculations to work out how many dies are required from each die fabrication area and the average expected revenue of each die type.
  3. The Front End Load Planner is the heart of the system. It uses LP to solve the capacitated load-planning problem across all the wafer fabrication plants and allocates the output of dies to assembly and test sites. Two separate manufacturing networks are represented, the larger giving rise to a matrix with about 160,000 rows.
  4. The Back End Loader uses LP to solve the capacitated load planning problem for each of five assembly and test sites.
  5. The Availability Calculator assembles the results and nets out the demands which are forecasts in order to calculate the availability schedules for products with firm orders.

Note the way in which the revenues from the finished products feed through modules 1 and 2 to the expected revenue from each die so that they can be used to drive the decisions about how the wafer fabrication facilities are to be used.

Practical Considerations

A system such as IMPReSS consists of much more than its algorithms. It also includes a worldwide database of factory status and capabilities, order status and marketing demands.

But to say this is to miss one of the greatest difficulties in implementing IMPReSS. The various parts of Harris defined their data entities in different ways. A common planning system required a standardised set of data entities. These new definitions raised conflicts with long-held conventions and extant factory-floor systems. It was a Herculean task to overcome these difficulties and put in place interfaces which mapped data from extant systems into the centralised new ones.

As further parts of the IMPReSS project, Harris installed a commercial demand forecasting system, implemented a Bill of Materials database and upgraded the order entry system to provide on-line delivery quotation and reservation capability.

Costs and Results

Althogether IMPReSS has cost about $4 million and has annual recurrent costs of $600,000. But it has succeeded. It has enabled Harris to achieve 95% on-time delivery, one of the best scores in the industry. Sales have increased to $700 million despite a decline in defense business. Losses, which reached $75 million in 1991, have been turned into profits of $42 million in 1995.

Related articles include Aluminimum Smelter Benefits from MP Consultancy and Optimizing the Supply of Bulk Gases. To find other articles, refer to the MP in Action page.