Indian-origin researchers Subhasish Mitra from Stanford University and Tathagata Srimani of Carnegie Mellon University have played a central role in creating the first monolithic 3D artificial intelligence chip fabricated in a US commercial foundry. Collaborating with teams from Stanford, Carnegie Mellon, the University of Pennsylvania, and MIT, the duo demonstrated a design that significantly enhances AI system performance while reducing energy consumption. This achievement was recently presented at a major semiconductor conference and represents a milestone in AI hardware development.
Subhasish Mitra, the William E. Ayer Professor of Electrical Engineering and Computer Science at Stanford, has long been a leader in advanced electronics, focusing on nanoelectronics, hardware security, and three-dimensional chip integration. His decades of research aim to make complex hardware faster and more reliable. Tathagata Srimani, an assistant professor at Carnegie Mellon, previously worked as a postdoctoral researcher under Mitra and specializes in energy-efficient computing and AI hardware architectures. Together, they translated a theoretical concept of vertically stacked memory and computing units into a manufacturable chip.
The innovation addresses the “memory wall” challenge in modern AI systems, where data movement between memory and processors slows operations and consumes excessive energy. By stacking memory and processing units vertically and connecting them with ultra-dense vertical links, the chip allows data to travel shorter distances, increasing speed and lowering power requirements. Unlike conventional “3D” chips that stack pre-fabricated components, this monolithic design integrates layers at a microscopic level, enabling unified operation and substantial performance improvements.
Prototype testing revealed roughly four times higher throughput than comparable flat chips, while simulations indicate that scaled-up versions could achieve 100–1,000× gains in energy-delay efficiency for intensive AI workloads, such as large language models. Notably, the chip was fabricated entirely at SkyWater Technology, the largest US-based pure-play semiconductor foundry, demonstrating domestic scalability and reducing dependence on overseas production.
This breakthrough offers a practical solution for next-generation AI, enhancing efficiency, reducing energy costs, and strengthening US semiconductor capabilities. Future work aims to scale the design further, potentially establishing monolithic 3D architectures as the foundation of advanced AI hardware.
