Computing is at a critical juncture. Applications such as AI/ML demand much larger memory, higher bandwidth, and lower-energy compute compared to business as usual. New hardware systems, powered by heterogenous memory, logic, and 3D integration, are required for large energy efficiency, throughput, and scaleup benefits. I will present my contributions to three such heterogeneous systems:

1. The first edge AI/ML chips with full on-chip inference and training of CNNs and Transformers using foundry Resistive RAM (RRAM). The heterogeneous combination of RRAM and SRAM enables new circuit-architecture-algorithm interplay, resulting in 9× or higher end-to-end Energy-Delay-Product (EDP) benefits versus traditional duty-cycled systems using SRAM, DRAM, or Flash.
2. New multi-chip Illusion systems for scaleup to 16× larger AI/ML models with 11× or higher EDP benefits versus traditional systems. Illusion minimizes inter-chip traffic by co-optimizing per-chip memory size, heterogeneous inter- and intra-chip interconnects, and idle power via fine-grained power management, thus creating the illusion (within 10% EDP) of a Dream Chip with all resources on-chip.
3. The first foundry heterogeneous monolithic 3D hardware integrating silicon CMOS, carbon nanotube field-effect transistors and RRAM, demonstrating 4× memory bandwidth versus iso-footprint, apples-to-apples conventional designs. Such benefits are only possible through new 3D architectures instead of 3D physical design alone.

Bigger benefits can be obtained for a wide set of applications by such hardware systems powered by a wider set of heterogeneous technologies. These also create exciting avenues for new courses.