Lithography and AI and quantum computing

In two plenary talks on Tuesday at the 2025 Advanced Lithography + Patterning conference, speakers from ASML and IBM explored the lithography challenges associated with AI and quantum computing. Overcoming those challenges would expand the computing toolbox, making it possible to do calculations and solve problems impossible to do with today’s technology.

Christophe Fouquet, president and CEO of ASML, discussed holistic lithography, saying AI demands require an all-encompassing approach to the technology. Chip computing capabilities are doubling every two years and power consumption is also falling. These two lithography driven trends have existed for decades.

However, the recent emergence of ChatGPT and other large language models is driving up the demand for computing and power at rates that far exceed the improvements chip makers can deliver. The only solution right now is to throw more and more hardware at the problem, which means larger and more data centers. But even with that, what AI needs will in a decade or so run up against a hard limit.

“The demand for power for AI training is going to exceed the available power in the world,” Fouquet said.

That, of course, will not happen, he noted. Instead, better models and algorithms will decrease the amount of power needed to compute each parameter in a model.

Semiconductor devices will also improve. Fouquet added that chips can be more efficient by changing their fundamental structure. Today, memory is in one place and data processing is in another on a chip. The brain, on the other hand, does processing and storage at the same place, and this in-memory computing is much more efficient. Memory makers are working to implement this new architecture, with early versions of this method already achieving a significant decrease in power consumption.

As for lithography tool makers like ASML, they are developing simplified process flows that cost less per transistor to carry out. The result will be lower cost for and higher density of transistors.

Part of those improvements involve cost reduction of EUVL (extreme ultraviolet lithography). EUVL is the most advanced semiconductor lithography. Today, though, the dominant lithography is deep ultraviolet, or DUV. Patterning fine features takes two passes through a DUV stepper, with each pass printing part of the pattern. The result is a more complex process with more layers, but this disadvantage is worth it because of the cost differential between the older technology and EUVL. Lower cost EUVL, however, changes that calculation, allowing chip makers to eliminate one pass and save money.

In the next plenary talk, Nelson Felix, director of technical strategy at IBM Semiconductors, agreed that the demand for AI has exploded, reporting that IBM’s customers now view the technology as critically important. Five or even three years ago, there might have been a single AI accelerator in a large computer. Now, there are many of them.

Part of the answer to meeting this need is to move into the third dimension. “We need to aggressively start looking at how we stack transistors,” Felix said.

Putting transistors atop on another will allow more of them to fit into a package. That increase combined with architecture changes will boost computing performance. Still more improvements can come from better measurement and process control of the patterning, Felix noted.

He added, though, that there are problems that neither classical computers nor AI can solve. An example is factoring a number into its components. This is so difficult for classical computers it forms the basis for the encryption used for financial and other highly secure transactions.

Factoring is easy, though, for a quantum computer, which uses the quantum effects of superposition, entanglement, and interference to operate on quantum bits (qubits). This power to do the impossible comes, in part, because instead of being a 1 or a 0, as is the case for bits in a classical computer, a qubit can be a 1, 0, or any value in-between. IBM and others have been working to achieve what quantum advantage – a quantum computer with enough qubits to be able to do what a classical computer cannot on a useful problem.

For its technology, IBM uses superconductors and has been making steady progress. By 2030 or shortly after that, the company expects to demonstrate quantum advantage. The hardware, though, is only part of what’s needed. There’s also the software and all the other elements required for people to use today’s computers.

Looking to the future, Felix mentioned data centers that would consist of a mix of classical, AI, and quantum computing platforms. Having a complete toolbox that doesn’t require extensive technology expertise is what end users really want, he noted.

"You don't have to worry about how AI works. You don't have to worry about how a quantum computer works. You just want answers,” Felix said.

Hank Hogan is a freelance science and technology writer.