How to weigh variables that will affect long-term yields
PATRICK FOLMAR
In advanced packaging, the increased attention on flip chip (FC) and chip scale packages (CSP) – fueled by the need to lower signal path transmission delays and reduce circuit board real estate – is only part of the story. How precisely these packages are placed now poses a dramatic impact on the process owner`s bottom line. As flip chip pitches move down to 150 microns and lower, and the use of non-self-aligning processes grows, many current surface mount technology (SMT) types of machines will be eliminated from handling these accuracy-dependent processes.
To be sure, as semiconductors become more complex and their packages become smaller, the accuracy of the machines placing these semiconductor die into packages or substrates is key. It has become imperative for manufacturers to understand accuracy fully, and that includes how a machine`s ability to maintain accuracy will affect yields in the long term. In addition, end users will be challenged enough in the future just dealing with semiconductor production yields because of die size and complexity. They should not have to worry about a die placement machine adding to yield losses due to misplacements.
Defining the Terms
Inconsistencies occur because accuracies stated by some equipment suppliers are really “sales” accuracies. They are merely stated to generate sales and are not achievable in a production environment. “Real accuracy” is measured by a machine`s performance a year or two after installation, after having performed 5 million to 10 million pick-and-place cycles. “Placement accuracy” is defined, in general terms, as how close the placement came to where it was supposed to be. Thus, it is the difference between the actual placement location and the desired placement location. For a pick-and-place machine, which uses encoder feedback to determine the proper location, this translates to how well it can move to a location by calculating the encoder counts it takes to arrive there. This is opposed to the term “repeatability” which, for a machine, is defined as how well it returns to a taught (known encoder count) location. Therefore, any machine should be far more repeatable than accurate.
Accuracy comes into play when a machine is programmed to move 100 mm from its current location. Or, as with most equipment today, it uses a vision system to generate offsets to known data for a placement location. Accuracy, as defined above, is the most critical parameter for pick-and-place equipment. Quantifying placement accuracy becomes even more critical. This is where statistics come in and why machine specifications should use terms like “±25 µm @ 3s.” If specifications are not described this way, buyers should be aware that the accuracy of the machine is not really known or is not a selling point. Without a statistics course, this accuracy nomenclature can be difficult to understand.
After a machine picks a part, uses vision to determine the offsets and places the part, accuracy is then measured and this process is repeated many times. The machine does not always return to the same location, but will form a random pattern around the ideal placement location. Most placements will be very near ideal and a few could be quite far (in relative terms) from it. When this data is charted to count the number of placements within concentric rings around the ideal placement location (Figure 1), a histogram graph results (Figure 2). If a line is drawn through the tops of each of the histogram bars, a shape known as a bell curve (Figure 3) results (also known in statistics as a normal graph). From this data, standard deviation can be calculated; this indicates how tightly bunched the locations are around an average location. The standard deviation is also known as sigma (s). Thus, a standard deviation of 10 µm (the same as saying ±10 µm @ 1s) would be a narrow bell curve (and a better accuracy number) than a standard deviation of 50 µm (a much wider bell curve).
The standard deviation number is linear when extrapolating out. Thus, ±10 µm @ 1s is the same as ±20 µm @ 2s and the same as ±30 µm @ 3s. This also means about 68 percent of placements are within ±10 µm, about 96 percent are within ±20 µm, and about 99 percent are within ±30 µm. If a machine has stated ±30 µm @ 3s accuracy, then a majority of its placements should be within ±10 µm. If a manufacturer only states an accuracy of ±10 µm without a sigma number, it can be assumed the reference is to 1 sigma at best (or, worst case, it may mean at least one part has been placed in a lab within those specifications).
Going Smaller, Faster
Two words sum up why it`s so important to understand a placement machine`s accuracy: Smaller, faster. Flip chip bumps are getting smaller to allow for either more I/O per die or for smaller die. Many current bump diameters are down to 75 to 80 µm at 150 µm pitch. New packaging techniques, such as silicon on silicon, are using bumps as small as 50 µm. Bumping houses have already produced flip chip bumps down to 30 µm and the market is heading in this direction. The performance of the die, especially in the RF communications field, needs to be faster to move data efficiently. Because transmission speed from the die to the substrate is up to 10 times faster than a wire bonded package, it is the only way to go for RF devices.
One reason many SMT machine manufacturers have been able to get into the flip chip market is because of the self-alignment of the solder bumps on the flip chips. This means that a bump can theoretically be off the pad by as much as 50 percent, and then, during reflow, the surface tension of the many solder bumps will pull the die to its proper location. A great majority of the flip chips being placed today use the C4 process with bump diameters in the 100 µm to 125 µm range and bump pitches between 225 µm and 250 µm. A machine with an accuracy of ±25 µm @ 3s can perform this application without yield problems, but a pitch of 225 µm is the lower limit for the C4 process. Therefore, as pitches get smaller, other processes must be used. Also, with smaller bumps, the self-alignment principle becomes far less of a factor because of less surface tension. Self-alignment is not a factor with many new connection techniques, such as conductive epoxy, anisotropic adhesive, and thermo-compression bonding. Since the die stays where it is placed, accuracy again is the key issue.
Why COO Matters
For an end user of placement machines, what matters most is cost of ownership (COO) and process yield. COO should be as low as possible, while the yield should be as high as possible. These are not mutually exclusive. Higher yields mean lower COO. The initial cost of a machine can sometimes pale in comparison to the future cost of an inaccurate machine. Assuming that a misplacement costs $10 (or much more for high-performance flip chip packages) and the machine is capable of producing 7 million placements/year, then each 1 percent drop in yield can cost $70,000 a year; the lost revenue would be much higher. End users need to be aware that it is very possible for a new placement machine to meet specifications initially, but may not after years of continuous use and wear and tear.
Accuracy Scenario
In a seven-day production environment (at 20 hours to 21 hours a day), many things can happen that affect accuracy over which the machine has no control. Thus, the first defense is to have a placement machine that operates at a high-production rate for high accurate placements (> 1200 uph), is accurate at the desired rate (to minimize the placement machine inaccuracy variable) and is built to minimize external variables it cannot control.
Raw product variations: Raw product variations can occur in the substrates, which are notorious for not being within specification. Their manufacturing process can make the pads` grid patterns intolerant of each other. Often the fiducials, which are used by the placement machine vision system to align the part to the pads, are printed on the substrate in a separate operation from the pads being put on, which adds yet another error variable.
Temperature variations: Ambient temperature surrounding a machine can vary. When machine manufacturers publish an accuracy specification, it is usually based on a machine tested at a stable ambient temperature. This is logical because many machines are used in some type of cleanroom environment where the temperature is expected to be stable. But in actual production, this cannot be ensured. If there is a 5°F temperature variation, you do not want to have to shut down production because your placement machine is placing parts 50 µm from the ideal. The machine itself can change the temperature through motor heating, friction in the bearings, electronic board heat, and process heating or cooling. Temperature variations usually have a greater effect on the mean of the bell curve than on the standard deviation. While placements may continue in a tight bunch around the mean, the mean may shift away from the ideal placement location.
Machine wear and tear: The final variable is simply the wear and tear occurring on all machines. Where there is friction, there is wear. If this friction is associated with the surfaces that are responsible to hold accuracy, you can expect a degradation of accuracy over time.
Machine Materials
A machine`s very construction, which must be stable mechanically, thermally and electrically, has a significant impact on its real accuracy. Another important requirement is floor support, with machines usually moving a large mass at speeds up to and more than 1 meter/sec. Pick-and-place machines fall into three construction categories based on the materials that make up the machine frame and its moving axis: aluminum-based, steel-based and granite/ceramic-based platforms.
Granite/ceramic-based machines are generally more accurate than aluminum- and steel-based platforms. Aluminum-based machines are not really considered high-accuracy platforms because of aluminum`s light weight and overall instability; these platforms can generally only achieve in the 50 µm to 100 µm @ 3s accuracy level, at best. Steel-based platforms, the most commonly built today, are much more stable than aluminum because of their extra weight and rigidity. However, the theoretical limit for accuracy on any steel-based platform – using the standard technique of having one camera looking at the substrate and another looking at the part after it has been picked – is about 20 µm @ 3s. Machine manufacturers have had to do a lot of work to achieve this. It is not possible to machine steel to the desired flatness over the length that the X-Y tables need to travel to achieve improved accuracy. The steel-based machine can result with waves or a bowed surface, when it ideally should be flat. As a result, steel-based machines have to be mapped to reach even a 20 to 25µm accuracy level. Mapping involves determining offsets that must be applied along the axis that does not fit the desired straight line.
Compensating for Temperatures
Temperature also impacts these platforms. Even ambient room temperature changes of only 5°F can cause a 20 to 30 µm shift in machine placement. To overcome this, equipment manufacturers have developed thermal compensation techniques. One method is to analyze one or more points on a table at repetitive times during a machine`s operation to determine what offsets need to be applied. Another is to actually make the encoders out of the same material as the rest of the machine so that it grows and shrinks at the same rate. However, even the best compensation method can only approximate the placement location, which represents another source of inaccuracy.
Finally, steel-based machines have a hard time reaching below the 20 µm @ 3s level because of long-term wear and tear. Steel machines generally use high-precision, well-lubricated linear bearings. But they do produce friction and therefore wear. Two to three years after use in a full-time production environment, for example, a machine that produced 20 µm @ 3s may now be at 30 or 40 µm @ 3s. These may be small numbers, but what is insignificant for the majority of pick-and-place equipment doing SMT and other low-accuracy applications can be detrimental for high-accuracy applications.
The Most Accurate Machines
Granite/ceramic-based machines are the most accurate. Granite is used in the base platform to provide rigidity with weight and a low center of gravity, a flat reference surface that the machine peripherals are precisely mounted to and a thermally stable reference surface.
Ceramic is used for the positional surfaces of the X and Y beams because ceramic can be machined to a flatness of ±2 µm over a 1-meter length. Thus, a maximum of only 2 µm of error is caused by the X or Y positioning system. Temperature has less effect on a granite/ceramic system. The 5°F differential that can move a steel-based system 20 to 30 µm would only affect the granite/ceramic by 2 to 3 µm. However, because temperature does affect placement, temperature compensation hardware and software are used on these systems, as well.
The main reason that granite/ceramic equipment is practically guaranteed to hold its stated accuracy for the life of a product is lack of wear and tear. These machines use frictionless air bearings that ride on the ceramic beams. Without friction on the positional surfaces, the machine will return to the exact same place two years from now. This has been shown in production by major microprocessor manufacturers who use this type of equipment for high-end microprocessor flip chip production. They choose this platform because it eliminates many of the variables that a steel-based machine brings to the equation and also brings a guarantee that accuracy will not degrade with time.
The theoretical accuracy limit for this type of platform built to run at production speeds (>1,200 uph) would be 5 to 8 µm @ 3s. Thus, for any application that must have a placement accuracy below 20 µm @ 3s, the only choice at this time is a granite/ceramic-based system. Otherwise, manufacturers risk never achieving desired accuracy levels or experiencing decreased yields in the long term. Table 1 shows different machine capabilities and the approximate throughput using a typical flip chip operation. These numbers represent what would happen in production on a long-term basis (real accuracy), not initial machine testing when new.
Bottom Line
In today`s marketplace, manufacturers evaluating pick-and-place machines must understand the impact of accuracy and how this measurement impacts yields, especially of high-end devices. An equipment manufacturer may state an accuracy specification, but the buyer must understand the physical limits that come into play in trying to get a production machine to these stated accuracy levels. In any production environment requiring high accuracy, it`s key that a buyer evaluate the equipment`s physical limitations, deciding amongst aluminum, steel and granite/ceramic equipment to ensure the gain in long-term yield.
PATRICK FOLMAR, technical manager, can be contacted at ESEC (USA) Inc., 9830 South 51st Street, Suite B-111, Phoenix, AZ 85044; 480-893-6990; Fax: 480-893-6793; E-mail: [email protected].
 |
 |
Figure 2. Histogram showing quantity of placements per circle.
 |
 |
Figure 3. Bell curves (left one has larger sigma).