Paradigm Shifts in Semiconductors: Lessons from the Past and the Emerging Playbook

For decades, computing has leapt forward by turning barriers into breakthroughs - the demise of Dennard scaling and the end of planar transistor scaling represent two such barriers. Each challenge sparked a paradigm shift, redefining what was possible. Drawing on my three decades at Intel, I will share lessons from these inflection points and discuss how they shaped the future of semiconductor innovation.

Today, the next frontier is energy efficiency. Performance per watt now dictates progress across every compute domain. I propose a “Call-to-Action” to develop a new transistor capable of operating at ultra-low operating V_DD (~300 mV) to dramatically improve compute energy efficiency while alleviating the significant performance loss. The device structure, key features, desired operating characteristics, and potential material options of this transistor will be presented. Achieving acceptable performance at ultra-low V_DD requires deep collaboration across device physics, circuit design, and microarchitecture??"a holistic approach to redefine the energy-efficiency curve.

AI accelerators face both “Energy and Memory Walls,” where data movement between the GPU and HBM (not computation) dominates power consumption and system performance. Enlarging on-die L2 cache is critical, yet conventional SRAM is constrained by large area and leakage. My vision introduces a hybrid gain-cell architecture using IGZO selectors and silicon read circuitry as a 6T SRAM replacement, delivering ultra-low leakage and ~50% area savings. Such a breakthrough can increase L2 cache capacity, thereby reducing HBM traffic, and boost performance per watt for future AI technologies.

Beyond memory, integration must evolve as well. TSV-based 3DIC offers significant gains, but Monolithic 3D (M3D) promises a true leap by stacking logic and cache with ultra-short, dense vertical links, improving bandwidth, latency, and energy efficiency. Overcoming thermal and low-temperature processing challenges will make M3D the foundation for compact, high-performance AI systems beyond TSV-based 3DIC during the next decade.