Energy- and size-efficient ultra-fast plasmonic circuits for neuromorphic computing architectures
Energy- and size-efficient ultra-fast plasmonic circuits for neuromorphic computing architectures
Energy- and size-efficient ultra-fast plasmonic circuits for neuromorphic computing architectures

Motivation

Computing industry is rapidly moving from a programming to a learning era, where the 10 GMAC/s/W (MAC – multiply-accumulate operation) digital energy efficiency wall of current CMOS electronics and von-Neumann architectures simply cannot keep the pace with the new computational power metrics required. The breakdown of Moore’s and Koomey’s laws referring to CPU and energy efficiency scaling is only validating the industrial consensus: by 2020, the reign of the von-Neumann architecture will begin fading away after 75 years of dominance, with non-von-Neumann layouts expected to be the key-enablers for optimizing Deep Learning (DL) and particularly inferencing, reducing the energy requirements for many computational classes. This new computing paradigm has already begun to unfold, leading to the development of large neuromorphic machines that already exceed the energy and size-efficiency walls of classical platforms.

However, the size and energy advantages of electronic processors are naturally counteracted by the speed and power limits of the electronic interconnects in the circuits due to RC parasitic effects, preventing further improvement in energy efficiency and computational power required for unleashing the huge potential of Deep Neural Networks and Artificial Intelligence (AI). This reality has already triggered a new research field – neuromorphic photonics – which aims to transfer the well-known high-bandwidth and low-energy interconnect credentials of photonic circuitry in the area of neuromorphic platforms. However, prevailing the interconnect industry does certainly not imply that photonics is the suitable technology when computations are required: having the critical dimensions of few mm’s and/or consuming > 25 mW even for weighing functionality.

Being currently at the dawn of neuromorphic computing, a future-proof solution that could dominate this landscape for many years should obviously rely on the best-performing and top-efficient technology mixture that can support the complete DL learning model portfolio. This is exactly where plasmonics comes to bridge what electronics and photonics lack: the size of electronics and the speed of photonics.

© 2020 WinPhoS, AUTH. All Rights Reserved.