Tradeoffs in multichip module yield and cost with known good die probability and repair

Abstract

The influences of known good die (KGD) probability, repair, and module testing on multichip module (MCM) yield and cost have been modeled and systematically analyzed. The current work extends our previous efforts on MCMs with single (one KGD probability) and dual (two KGD probabilities) populations to modules containing complex, multiple chip populations. Most of the analysis is performed on modules with multiple populations in the range of three to five (i.e., three to five distinct chip types). In order to develop a total cost picture for an MCM versus the respective KGD probabilities of the underlying chip populations, it was necessary to develop new algorithms or modify previously developed algorithms for the following items: number of modules (necessary to ensure at least one MCM works), KGD chip cost, and chip repair/replacement costs. In order to visualize the results and simplify calculations, averages over the respective subpopulations have been employed. The combination of these models and algorithms produces cost values in multiple (KGD) probability space that contain optimum or minimal points. Associated with the cost minimums are specific KGD probabilities for each chip type in the module population. Thus, one only pays for improved KGD probability up to the values that minimize overall module cost. Repair has a direct and significant impact on overall module yield and cost, with the first repair providing the largest improvement in both yield and cost reduction.

Introduction

The production of working multichip modules (MCMs) requires the confluence of three major elements: defect-free (tested) substrates, known good die (KGD), and high quality assembly procedures (error free). Defect-free substrate validation is usually accomplished by a series of design rule checks to ensure compatibility with the MCM fabrication processes [1] followed by electrical testing using bed of nails (probe card) or flying probe testers. The electrical tests verify that all networks on the substrate are connected in the appropriate manner (consistent with the schematic diagram) and that there are no opens, shorts, or nets with resistances (impedances) outside of the prescribed ranges. High module yields require testing of the chips prior to assembly. Much time and effort over the last few years [2], [3] has been expended on the mounting of bare die for full functional testing (at speed and temperature) and then demounting them for actual attachment to the MCM.

Following assembly [1], the resultant MCM must be tested to ensure a working module, thus validating the defect-free nature of the assembly process and certifying the previous results of chip (KGD) and substrate testing. In an MCM, if one die fails, the whole module fails or its performance is so reduced that the MCM cannot be used for its intended purpose or sold at a price consistent with full cost recovery. Since chip and substrate pre-testing alone cannot ensure that all dies will perform as planned after assembly, MCMs must either be built so inexpensively that they are disposable (throwaways) (i.e., using low cost, high yield, high volume manufacturing processes) or they must be able to be repaired (i.e., the replacement of defective chips).

Testing to find the defective chips is a key element in the repair process. Without appropriate procedures to locate defective die, repair can be extremely costly or impossible. MCMs must be designed from the beginning for testability and repair, including the complete test protocol necessary to locate defective die. The need to test and repair influences substrate design including extended bond pads; room for die attach and removal tools; robust board metallizations; and, perhaps, extra test points, traces, and/or chips to support the testing process. The chips themselves may require extra circuit elements and contacts to facilitate the testing process. The design for testability [4] of both substrates and integrated circuits is beyond the scope of this work.

This study focuses on module yield and repair, making the assumptions that the defective chips can be located, removed, and replaced – thus producing working MCMs. In our modeling, we account for the difficulty in locating and replacing defective chips by a repair complexity factor, which directly reflects the cost of those operations. Our goal is to minimize overall chip and module cost as a function of KGD probabilities and the amount of repair (number of repairs and their associated costs). In this work, we will consider: (1) modules with all the same chips or modules with different chips all having the same KGD probability [5], (2) modules containing chips with two different KGD probabilities [6], [7], and modules containing more than two distinct chip populations [8] with different KGD probabilities. Results are presented for various chip populations (including one or more different KGD probabilities) for MCMs of fixed size and varying amounts of repair. Module testing and the influence of underlying chip-processing yields on test results are brought into perspective. In addition, attention is given to estimation of the number of modules required to ensure the production of at least one good module, as well as refinements in previously presented [5] chip cost and repair cost algorithms. Graphical displays allow the reader to make tradeoffs between KGD probability and the cost of both chips and modules.