Architectures for the Future Networks and the Next Generation Internet

Abstract

Ultra-low power attentive systems with always-on operation and signal monitoring with a disproportionately higher peak performance are now being in high demand, due to the convergence of AI and IoT. In this talk, circuits and architectures to enable exceptionally low power consumption in the common case while achieving high peak performance are discussed for next-generation intelligent systems.
Several silicon demonstrations are presented for accelerators, processors and SRAMs with enhanced peak performance above traditionally allowed at nominal voltage, yet at reduced minimum energy. Energy-quality scaling is explored as additional dimension to break the conventional performance- energy tradeoff in error-resilient applications such as AI and vision, from networks on chip to memories and accelerators. Further performance and energy improvements are discussed through uncommonly flexible in-memory broad-purpose computing frameworks for true data locality, from buffering to signal conditioning and neural net workloads. Finally,
challenges and opportunities for the decade ahead are discussed to enable next-generation attentive & intelligent systems withdivergently high peak performance, common-case performance and low minimum power.