Manufacturing Next Gen Microchips in the Age of AI

Because the surge in AI-enabled electronics is pushing the limits of the current manufacturing capabilities, Williams kicked off the conversation by asking Leirer and Schmucker of SCHOTT for their perspectives on the role that advanced materials take in meeting these challenges. “We’re at the end of Moore’s law, and with that coming to an end, we need to find new ways to boost semiconductor performance,” Leirer said. Moore’s law is the observation from Gordon Moore’s 1965 paper that postulates that circuit density doubles every two years. But technology is such that we’re approaching a point where, because of the size of an atom, semiconductors are as small as they can get in their current form.

That’s putting more emphasis on not only how chips are fabricated (the frontend), but the backend packaging—how chips are combined to function. So while chip manufacturers are limited in how much smaller transistors can be shrunk, performance bumps are achieved through packaging and combining various functions and specifications. “It’s an interesting time,” said Schmucker. “We’re obviously a glass company, so we’re approaching this from a very specific material mindset, but in terms of spend and advising our clients, the focus is really on backend packaging.”

Other experts in the room tackled the question from a different perspective, with the group noting that much of the meta discussion around processors has been on the final product: cutting-edge frontier AI models. However, out in the marketplace, attendees are seeing increasing demand for “n-1” technologies (products that are just below the top of the line) that are less resource intensive to build and deploy than some of the “flashier” marquee products, like the GPUs used in training AI models. After all, when the initial exuberance of new technology fades, organizations that have boards with a fiduciary responsibility to shareholders will be interested in how these products have been implemented and their relative return on investment.

The discussion highlighted that in the interim, one of the best use cases has been implementing AI in data centers, but in that case, there is the question of energy management. Optimizing for better heat management and energy efficiency goes all the way down to the processors, and when you’re trying to fit an exponential amount of processing power into a vanishingly small amount of square millimeters, it causes other issues related to memory access, power, and more. Once again, the business opportunity comes back to packaging those semiconductors.