The 2026 SPIE Advanced Lithography + Patterning conference highlighted AI, both as a challenge and a solution. A case in point was the opening plenary session, which featured presentations on high performance memory and diversified manufacturing.
The challenge of AI played a large role in the first talk. The existence of ultra large AI models with trillions of parameters, up from billions a few years ago, is improving AI capabilities. Those enormous models, though, pose a problem because they demand higher performing chips. At one time, the limitation was processing power, but that's no longer the case.
"The bottleneck is shifting from compute to memory," said Unoh Kwon, vice president at high performance memory maker SK hynix .
The company concentrates on high bandwidth memory, chips that can shuttle data at many gigabits per second to and from processors. Kwon said SK hynix is increasing bandwidth by a factor of 1.5x every two years, meaning that if it were 10 Gb/s now, it would be 15 Gb/s in two years. The memory intended for AI use must also have large storage capacity and be low power.
To achieve the needed improvements, the semiconductor industry is moving the memory closer and closer to the processor. That lessens the distance the signal needs to travel, boosting speed and reducing power. On the horizon is memory that not only stores data but also processes it – an approach similar to what neurons do.
Another method deployed to meet these AI-driven challenges involves stacking memory chips atop one another. SK hynix, for instance, builds a tower that may be as high as 20 chips. Such stacking increases memory density but comes at a cost – any warpage of the die causes problems. Such warpage happens because die are thin and under mechanical stress from the manufacturing process. What's more, the warpage grows along with the stack.
"Every time we stack them, the warpage becomes amplified," Kwon said of memory chips.
This 3D approach is also a challenge because it makes it hard to spot defects. Device killers may hide in a stack, invisible to most detection methods.
A way to potentially use AI as a solution came from the next presentation, where GlobalFoundries Vice President Hui Peng Koh talked about diversified manufacturing. She pointed out that AI is helping the semiconductor industry overcome manufacturing hurdles.
GlobalFoundries make a mix of products at their chip factories. They churn out millions of memory, logic, silicon photonics, and other chip types using various technologies, producing an array of products over a period of years. This diverse manufacturing is all done on the same set of equipment, with various specification requirements. This situation creates a manufacturing issue.
"The first thing that will break is not the process but the control," Koh said of such a chip manufacturing setup.
Silicon photonics chips, she pointed out, move photons around instead of electrons. Thus, their feature sizes tend to be in the hundreds of nanometers, not the handful of nanometers needed for the most advanced electronic chips. But silicon photonics are very sensitive to line edge roughness and so demand exceptionally smooth features. If line edges are the least bit jagged on the scale of nanometers, then photons will get lost and a silicon photonics chip will not work as it should.
As this example shows, a lack of control of any of a number of parameters may make a process yield underperforming or even defective chips. But discovering the problem may take anywhere from minutes to months, which means that a great deal of production may be at risk of being scrapped.
Human-based methods, such as charting process measurements to spot problems, do not work well in such an environment, Koh said. With training, AI, on the other hand, can ingest data and look for patterns. A machine-based approach can even handle low volume products, which run infrequently and those do not produce much measurement data. In that case virtual measurements created to augment the actual ones can provide a high enough volume to improve process control, according to Koh.
In summing up, she said AI methods have changed how semiconductor process engineers operate. "Engineers are no longer managing charts," Koh said. "Instead, they are managing the system."
