IBM has the blueprint for the next generation of tiny, but mighty, artificial-intelligence chips.
Big Blue's research division announced Thursday that it had produced the world's first sub-1 nanometer chip technology, featuring a transistor architecture at the 0.7 nanometer node.
It is in character for the tech giant -- its research division produced the first prototype for a 2 nanometer node chip back in May 2021.
With that development, "we highlighted the research, and now all leading foundries are manufacturing these," IBM Research Director Jay Gambetta said in an interview with Barron's. "I think this is going to be as big as that. It's a completely new paradigm."
Today, most AI chips are built on 5 nanometer and 4 nanometer fabrication processes and almost exclusively produced by Taiwan Semiconductor Manufacturing. Nvidia's Blackwell GPU, for instance, uses a customized version of TSMC's 4nm node.
However, the latest generation of premium AI hardware has migrated to 3nm processes. By squeezing more transistors into tighter spaces, these advanced nodes boost processing speeds while cutting down on power consumption.
Nvidia's coming Rubin platform, slated to ship in the third quarter, leverages TSMC's 3 nanometer family for precisely this reason. Looking ahead, TSMC announced it has already begun mass-producing next-generation 2 nanometer chips to replace aging technology.
Now, IBM is taking it one step further. "Our goal at IBM Research is to never be done," Gambetta told Barron's. "And so we asked the question, how do you go beyond nanosheets?"
Nanosheet technology is the latest breakthrough in semiconductor manufacturing, succeeding the FinFET architecture that powered most advanced chips over the last decade. It was designed to stop electricity from leaking as transistors shrink to microscopic levels, ultimately allowing chips to become smaller, faster, and more efficient.
IBM is embracing so-called nanostacks to push logic scaling beyond traditional nanosheets -- by building upward, rather than parallel, to allow more transistors in a given footprint.
Gambetta emphasized that the announcement purely pertains to research, as the real challenge lies in producing chips at scale. "We aren't claiming we've got to manufacturability," he said. "When it's going to have an impact is going to be when the manufacturing costs are equal to the performance improvements. And the initial data say this is very likely."
The announcement doesn't signal a new direction for IBM, which has worked with silicon since the 1950s. However, Big Blue no longer manufactures chips itself -- it exited physical fabrication in 2015 by transferring its semiconductor plants to GlobalFoundries.
While a pair of lawsuits fractured their relationship between 2021 and 2023, IBM and GlobalFoundries have since buried the hatchet. Today, they are independently working on advancing quantum computing, another one of IBM's big initiatives.
In May, the Commerce Department and IBM each committed $1 billion to create Anderon, a stand-alone quantum chip foundry. As part of the same funding package, GlobalFoundries secured a $375 million infusion to establish a business unit focused on domestic manufacturing for utility-scale quantum computing.
"We see quantum computing as the next technology. We're going to manufacture quantum computers and Anderon is our foundry to scale that," Gambetta told Barron's. "But we're trying to make sure classical computing hasn't come to an end as well."
As he sees it, this philosophy is what drives IBM's ongoing work in transistors. "It continues to be an important part of our research mission to understand the limits of silicon technology," Gambetta explained. "We can leverage that foundation to be the fastest mover in the quantum space as well."
IBM stock climbed 5.8% in premarket trading Thursday following the announcement. Futures tracking the S&P 500 were up 0.8%.
Comments