First monolithic 3D chip built in US foundry delivers major AI speed gains
First monolithic 3D chip built in US foundry delivers major AI speed gains
Story by Ian Scheffler • 19h
A collaborative team has achieved the first monolithic 3D chip built in a U.S. foundry, delivering the densest 3D chip wiring and order-of-magnitude speed gains.
Subscribe to our newsletter for the latest sci-tech news updates.
Engineers at Stanford University, Carnegie Mellon University, University of Pennsylvania, and the Massachusetts Institute of Technology have collaborated with SkyWater Technology, the largest exclusively U.S.-based pure-play semiconductor foundry, to develop a novel multilayer computer chip whose architecture could help usher in a new era of AI hardware and domestic semiconductor innovation.
Unlike today's largely flat, 2D chips, the new prototype's key ultra-thin components rise like stories in a tall building, with vertical wiring acting like numerous high-speed elevators that enable fast, massive data movement. Its record-setting density of vertical connections and carefully interwoven mix of memory and computing units help the chip bypass the bottlenecks that have long slowed improvement in flat designs. In hardware tests and simulations, the new 3D chip outperforms 2D chips by roughly an order of magnitude.
While academic labs have previously built experimental 3D chips, this is the first time such a chip has shown clear performance gains and been manufactured in a commercial foundry. "This opens the door to a new era of chip production and innovation," said Subhasish Mitra, the William E. Ayer Professor in Electrical Engineering and professor of computer science at Stanford University, and principal investigator of a new paper describing the chip presented at the 71st Annual IEEE International Electron Devices Meeting ( IEDM 2025), held in San Francisco Dec. 6–10. "Breakthroughs like this are how we get to the 1,000-fold hardware performance improvements future AI systems will demand."
The challenges for flat chips
Modern AI models, like ChatGPT and Claude, must move staggering amounts of data back and forth between memory, which stores information, and the computing units that process it.
On conventional 2D chips, components sit on a single, flat surface with limited, spread-out memory, so data must travel across a few long, crowded routes. Because the computing elements run much faster than the data can move—and because the chip can't store enough memory close by—the system ends up constantly waiting on information. Engineers call this bottleneck the "memory wall," the point at which processing speed outpaces the chip's ability to deliver data.
For decades, chipmakers addressed the memory wall problem by shrinking transistors—the tiny switches on a chip that perform computations and store data—and squeezing more of them onto each chip. But that strategy, too, is approaching hard, physical limits, which researchers refer to as the "miniaturization wall."
The new chip climbs these walls by literally rising above them. "By integrating memory and computation vertically, we can move a lot more information much quicker, just as the elevator banks in a high-rise let many residents travel between floors at once," said Tathagata Srimani, assistant professor of electrical and computer engineering at Carnegie Mellon University, the paper's senior author, who began the work as a postdoctoral fellow advised by Mitra.
"The memory wall and the miniaturization wall form a deadly combination," said Robert M. Radway, assistant professor of electrical and systems engineering at the University of Pennsylvania and a co-author of the study. "We attacked it head-on by tightly integrating memory and logic and then building upward at extremely high density. It's like the Manhattan of computing—we can fit more people in less space."
How the new 3D chip is made
Until now, most attempts at 3D chips have relied on stacking separate chips. That approach works, but the connections between layers are coarse, sparse, and prone to bottlenecks.
Instead of fabricating separate chips and then fusing them, the team builds each layer directly on top of the last in one continuous process. This "monolithic" method uses temperatures low enough to avoid damaging the circuitry below, allowing the researchers to stack components more tightly and connect them far more densely.
Perhaps even more notably, the process was completed entirely in a domestic commercial silicon foundry. "Turning a cutting-edge academic concept into something a commercial fab can build is an enormous challenge," said co-author Mark Nelson, vice president of technology development operations at SkyWater Technology. "This shows that these advanced architectures aren't just possible in the lab—they can be produced domestically, at scale, which is what America needs to stay at the forefront of semiconductor innovation."
The chip's performance and potential
Early hardware tests show that the prototype already outperforms comparable 2D chips by roughly a factor of four. Simulations of taller, future versions—with more stacked layers of memory and compute—point to even greater gains. Designs with additional tiers show up to a 12-fold improvement on real AI workloads, including those derived from Meta's open-source LLaMA model.
Most strikingly, the researchers say the design opens a realistic path to 100- to 1,000-fold improvements in energy-delay product (EDP), a key metric that balances speed and energy efficiency. By drastically shortening data movement and adding many more vertical pathways, the chip can achieve both higher throughput and lower energy per operation, a combination long viewed as out of reach for conventional, flat architectures.
The researchers emphasize that the long-term significance of this research goes beyond performance. By proving that monolithic 3D chips can be built on U.S. soil, they say, the work establishes a blueprint for a new era of domestic hardware innovation, one in which America can design and manufacture the most advanced chips.
Just as the integrated-circuit revolution of the 1980s was fueled by students who learned to design and build chips in U.S. labs, the researchers say the shift to vertical, monolithic 3D integration will require a new generation of engineers fluent in these technologies. Through collaborations and funding sources, students and researchers are already being trained to advance American semiconductor innovation.
"Breakthroughs like this are of course about performance," said H.-S. Philip Wong, the Willard R. and Inez Kerr Bell Professor in the Stanford School of Engineering and principal investigator of the Northwest-AI-Hub. "But they're also about capability. If we can build advanced 3D chips, we can innovate faster, respond faster, and shape the future of AI hardware."
Provided by Stanford University |
