Adding more layers in packages is making it difficult, and sometimes impossible, to inspect wire bonds that are deep within the different layers.
Wire bonds may seem like old technology, but it remains the bonding approach of choice for a broad swath of applications. This is particularly evident in automotive, industrial, and many consumer applications, where the majority of chips are not developed at the most advanced process technologies, as well as for various memories.
But wire bonds also have a host of problems, and those problems are becoming more pronounced. They lack enough I/Os for increasingly heterogeneous devices, and they seem fragile and overly complicated compared with flip chips. And that’s just for starters. In addition, solder balls may be misshapen, in what are dubbed “golf ball” defects. Wires may be “depressed,” meaning they don’t conform to normal shapes, or they may not even reach their intended pads. Worst of all, defects can be hidden in the interwoven complexity of hundreds of overlapping wires, causing opens or failures that may go undetected.
“It’s virtually impossible to rework layers once you have layers on top,” said Jeff Schaefer, senior process engineer of Promex Industries. “We’ve seen things where it’ll be six rings on a substrate and two rings on the dies. The outer ring on the die usually goes to the first three rings on substrate and then the inner ring on the die goes to the outer rings on the substrate, and all of these things are going over the top of each other.”
Traditionally, there have been two approaches to discovering flaws.
“From the day when this technology started, they’ve been doing the same kinds of tests,” said Per Viklund, director of IC packaging and RF product lines at Siemens EDA. “It’s destructive testing, in which they do a test design and try to pull the wires off, and measure how firmly they grab. While they still do that, the interesting question is, ‘Can we inspect something that’s intended to be shipped to the customer and validate that it’s okay?’”
Indeed, destructive methods are being challenged by smaller and more complex configurations, said Chris Davis, product line manager for semiconductor products at Nordson. “Testing can be even harder than placing the wires down. There needs to be a tool or an implement to fit within the pitch of the wires in order to do a mechanical test on them. There are changes in the way that the bonding process is taking place, so that bonds become almost untestable in that kind of form.”
The non-destructive approach was visual inspection, which was relatively simple when there was just a single layer to examine. All it took was looking at wire bonds either with the naked eye, or as they got smaller, with an optical microscope. The biggest challenge to accuracy was throughput: There could be thousands of wires on a single chip, but an operator would have to make a judgment call in a few seconds to keep up with production runs.
In many cases, the complexity now far exceeds what the human eye can grasp at any speed. In some instances, there can be a dozen or more layers of wire bonds. A challenging example is in memory, said Frank Chen, director of applications and product management at Bruker. “Each of the NAND/DRAM layers require wire bonds, which results in many wires flying over each other. Previously, when it was just a single layer of wire bonds, it was relatively easy to tell it was shorting. Now, with multiple wires overlapping, it’s harder to determine if the wires are shorted or connected to a different layer.”
During wire bonding, it is also critical to monitor the distance between wires, because if they are too close together they eventually can short under mechanical or thermal stress. The challenge lies in monitoring this margin between wires in multiple dimensions.
“In functional tests, they’re testing through the wire bond, and would obviously find if there’s an open connection or a short circuit,” said Viklund. “They’re looking at things like wire sweep. Many of these wire bonded designs are covered in epoxy. There is a risk that when the epoxy flows in that it drags the wires with it, and the wires then bend sideways to the level where they get too close to each other or are touching.”
The solution for these issues can go beyond optical microscopy into X-ray microscopy. “A nuisance with optical inspection is handling the reflections from the metallic wires during 3D analysis,” said Chen. “X-ray microscopy can avoid this issue and determine the unique wire paths using just a few angles to maintain high throughput.”
While some chips are assembled in such a way that all the wires are visible, some are so complex that even 2D X-ray won’t be enough, said Viklund. “If you have eight levels of wire loops on top of each other and you don’t do 3D X-ray, it’s hard to tell where things are overlapping in the same plane. Your standard 2D X-ray is just a 2D projection of your design and you cannot really tell if there are conflicts in the depth of the picture.”
X-ray inspection can go beyond what a human operator can catch, which has led to the automation of X-ray inspection. “When the dies are on top of each other connected by wires, it’s is very challenging to inspect with manual X-ray, especially typical defects, like sweeps or saggings,” said Margareta Popovics, product line manager of the automated X-ray inspection system product line at Nordson.
For X-ray, there are also the challenges of different materials and thicknesses, which are often sector dependent. “For automotive, most of the controllers are still bonded with gold wires,” said Popovics. “While that inspection is less challenging, the complexity is rising. So they’re less of a problem because of the material and more of a problem because of the complexity of the sample. Oftentimes, we have to combine inspection methods from 2D and 2.5D in order to cater to all that could happen during wire inspection. We also have to examine for die shifts or poor wettability and similar aspects.”
By contrast, consumer electronics is generally the realm of copper bonds, which presents its own problems for X-ray inspection. “It’s a less visible, low-contrast material, which makes it hard to get the dimensions of those wires. It can go down to 0.6 mil, which is about 20 microns in diameter,” said Popovics. “So it requires high resolution and really low power and a little bit longer inspection type compared with the gold wires. But the complexity comes in again. I would say the most challenging products that we have to inspect are usually the copper-wired products with high complexity from consumer electronics.”
As dimensions get smaller, X-ray metrology also can run up against physical limits, said Davis. “For the resolution size, we’re limited by the wavelength of the photon that we’re using to inspect the product. So while X-ray is good, it does have an efficiency limit, because we need to be able to pass the X-ray through the product and turn it into something that we can measure on the other side,” said Nordson’s Davis. “We’re trying to improve quantum efficiency to detect those X-rays at lower and lower energies, and get more out from the X-ray source detector so that we can properly inspect at much higher resolutions.”
AI/ML
As X-ray becomes automated, AI can help with accuracy. An AI algorithm can be written to validate against a canonical shape or learn from imperfect shapes what to flag. An automated system with an AI program can detect most of the defects even in 2.5D stacks, said Popovics.
However, AI has its risks and frustrations, warned both Viklund and Chen.
“You rely on the test equipment’s algorithms to detect correctly,” said Viklund. “But you don’t want any false positives. If you pull a perfectly good design off the belt and don’t ship it, that’s a cost.”
One company created an algorithm that helped to automate inspection, according to Chen. But in the end, it became a hopeless game of catch-up. “As soon as they
got a little bit more complex, then the algorithm broke, and they had to spend more effort to develop it again. It can become unending. You solve the problem, then the parts get more complex and it breaks again.”
Optical
Optical inspection, in the form of confocal microscopy, is also advancing to the point that it can be competitive in this new world, said Oliver Schulz, business development manager, semiconductors at Precitec. “The cheap solution in the beginning was just the camera, but the camera has a problem. It’s just two-dimensional. You don’t have image depths. You need to have a depth of measurement range.”
His company, for example, makes a high-speed confocal line sensor, which can be integrated into the wire bonding machine itself or used as a standalone system after wire bonding. “Confocal microscopy has existed for decades,” said Schulz. “Because a wire bond has multiple points, you need to get a three-dimensional picture, so one confocal sensor is not enough. If you just had one point, it would be necessary to move the sensor along the whole wire. So you would need to measure every wire separately, which would take too much time and the sensor would be too slow. If you have multiple points, you just move the sample below the line so it’s really like a scanner.”
Conclusion
Ultimately, the choice of wire bond inspection method comes down to use cases and cost. “It depends on what the end price of your device is,” said Schulz. “Here’s a simple example. A DVD player on a laptop may have one laser inside with one or two bond wires to connect to the power, and the laser may cost about a dollar. There will never be a return on investment for buying an advanced optical system to inspect it. It’s way too expensive, especially with machine learning software.”
By contrast, he said, what if your end product is an undersea cable, where a laser diode could cost $5,000. “You need to have a much higher yield. And the cost to replace it is much higher because it’s undersea. Those people are much more willing to invest into metrology systems,” said Schulz. “In every process, it’s the same. The R&D guys have many cool ideas about how to measure and what is probably necessary. But at the end of the day, there’s somebody calculating the cost of the improvement in yield, and he’s coming back and asking, ‘Is there any return on investment?’”