Building Efficient AI Hardware for a World Without the Cloud
November, 2026
Dr. Mircea Stan and Dr. Ethan Ahn first met at a Virginia Microelectronics Consortium (VMEC) gathering, where a conversation about overlapping research interests quickly turned into a plan to collaborate. Dr. Stan is a Professor of Electrical and Computer Engineering at the University of Virginia (UVA), and Dr. Ahn is an Associate Professor of Electrical Engineering at George Mason University (GMU). With 4-VA funding, the pair is now building efficient AI hardware that can run without relying on the cloud, a project that has also drawn students into hands-on research from the start.
The two groups meet every two weeks to review progress and plan next steps toward publication. The division of labor plays to each side's strengths. Dr. Stan's group at UVA designs and simulates the circuits and architectures, while Dr. Ahn's group at GMU fabricates and tests the resistive memory devices in the clean room. The pairing is productive precisely because of the gap between how a device behaves in simulation and how it behaves once it actually exists; each round of fabrication sharpens the next round of design. Both teams hope to build on the partnership with larger joint proposals in the future.
The partnership also created opportunities for students. UVA undergraduate Noreen Hossain interned in Dr. Ahn’s lab at GMU, where she learned to fabricate microelectronic devices. Her internship was part of the Virginia Microelectronics Consortium's Summer Scholar Program, which places undergraduates from Virginia's public universities into summer internships with industry partners like Micron and BAE Systems or in universities other than their home institution. Noreen started in research even earlier through USOAR, a UVA program that pairs first- and second-year students with faculty. The work she did in USOAR sat outside her electrical engineering major. Even so, it taught her how research is run and helped set her on the path to the VMEC internship at GMU.
In Dr. Stan's lab, Noreen modeled resistive RAM devices in simulation. She ran the simulations and pulled out each device's material properties. She then turned those properties into settings for predictive models and checked them against how the real devices behaved. At the end of the year, she presented the work at UVA's Undergraduate Research Symposium, tracing how her fabrication work at GMU informed the modeling she did at UVA.
Melika Morsali, the Ph.D. student leading UVA’s portion of the project, mentored Noreen, walking her through the unfamiliar simulation software until she could run it independently and serving as her first point of contact whenever a result did not make sense. Along the way, Noreen came to understand these devices down to their fundamentals: how they behave, why they misbehave, and how they fit into larger circuits. This experience pointed her toward a career in the semiconductor industry. She plans to gain industry experience after graduating this year, with the possibility of returning for a master's degree at UVA or elsewhere.
The 4-VA project is part of a larger three-year effort funded by the Semiconductor Research Corporation (SRC), an industry consortium whose members include many of the leading companies in the semiconductor field. The larger project focuses on building AI hardware that runs efficiently at the edge, away from the cloud's massive servers. A sensor on a GPS in a remote field, a car in a connectivity dead zone, or a military system all need to make fast decisions locally. At the edge, being fast and accurate is not enough. The hardware also has to be efficient enough to run without kilowatts or even megawatts of power.
Nearly every AI system depends on an operation called multiply-and-accumulate, or MAC. Most computers handle it digitally, using the ones and zeros of standard Boolean logic. Dr. Stan's group is trying a different route by doing the math using a component called resistive RAM. This approach is fast because the calculation happens in a single physical step and compact because resistive RAM devices are small. It is also low power because the work is just electric currents combining rather than a chip switching billions of times a second.
The payoff is AI that can operate “at the edge”, where the cloud cannot reach. If complex computations can run on a small, low-power chip, processing moves out of distant servers and into the device itself. For Dr. Stan's team, the work is a bet on where AI hardware is headed.