The 2026 SPIE Advanced Lithography + Patterning conference highlighted AI as both a challenge and a solution. A great example is the opening plenary session featuring presentations on high-performance memory and diversified manufacturing.
The challenges of AI played a big role in the first talk. The existence of very large-scale AI models with billions to trillions of parameters a few years ago has increased the capabilities of AI. However, these huge models require more powerful chips, which poses a problem. Processing power used to be the limit, but no longer.
“The bottleneck is shifting from computing to memory,” said Uno Kwon, vice president of high-performance memory maker SK Hynix.
The company is focused on high-bandwidth memory, chips that can transfer data to and from processors at speeds of several gigabits per second. Kwon said SK Hynix increases its bandwidth by 1.5 times every two years, which means that if it is 10 Gb/s now, it will be 15 Gb/s in two years. Memory for AI applications is also required to have large capacity and low power consumption.
To achieve the necessary improvements, the semiconductor industry is moving memory closer and closer to the processor. This shortens the signal’s reach, increases speed, and reduces power. Memory that not only stores data but also processes it is on the horizon. This is a similar approach to neurons.
Another method introduced to address these AI-driven challenges involves stacking memory chips on top of each other. For example, SK Hynix builds towers of up to 20 chips. Such stacking increases memory density but comes at a cost. Warped dies can cause problems. This warpage occurs because the die is thin and subject to mechanical stress from the manufacturing process. Moreover, as they are stacked up, the warpage increases.
Regarding memory chips, Kwon said, “Every time you stack them, the warpage gets amplified.”
This 3D approach is also a challenge because it makes it difficult to find defects. Device killers can be hidden in the stack and are invisible to most detection methods.
The potential use of AI as a solution was highlighted in the following presentation by GlobalFoundries Vice President Hui Peng Koh, who talked about diversified manufacturing. He noted that AI is helping the semiconductor industry overcome manufacturing hurdles.
GlobalFoundries manufactures a variety of products in its chip factories. The company has been producing a variety of products over the years, using a variety of technologies to mass produce millions of memory, logic, silicon photonics, and other chip types. All of this diverse manufacturing takes place on the same set of equipment with varying specification requirements. This situation creates manufacturing problems.
“The first thing that breaks down is the control, not the process,” Koh said of these chip manufacturing systems.
Silicon photonics chips move photons instead of electrons, she noted. Therefore, their feature sizes tend to be hundreds of nanometers rather than the few nanometers required for state-of-the-art electronic chips. However, silicon photonics is highly sensitive to line edge roughness, requiring extremely smooth functionality. If the line edges are jagged on the nanometer scale, photons will be lost and the silicon photonics chip will not work properly.
As this example shows, not being able to control any one of many parameters can result in low process yields or defective chips. However, discovering problems can take minutes to months, meaning large amounts of production can be at risk of being scrapped.
Coe said human-based methods, such as graphing process measurements to identify problems, don’t work well in this environment. Training, on the other hand, allows AI to take in data and look for patterns. Machine-based approaches can also process low-volume products that are operated infrequently and do not generate much measurement data. In that case, Koh says virtual measurements created to augment real measurements can provide enough to improve process control.
In her summary, she said that AI techniques have changed the way semiconductor process engineers work. “Engineers are no longer managing charts,” Ko says. “Instead, they manage the system.”
/Open to the public. This material from the original organization/author may be of a contemporary nature and has been edited for clarity, style, and length. Mirage.News does not take any institutional position or position, and all views, positions, and conclusions expressed herein are those of the authors alone. Read the full text here.
