A Design Methodology for Post-Moore's Law Accelerators: The Case of a\n Photonic Neuromorphic Processor

Over the past decade alternative technologies have gained momentum as\nconventional digital electronics continue to approach their limitations, due to\nthe end of Moore's Law and Dennard Scaling. At the same time, we are facing new\napplication challenges such as those due to the enormous increase in data. The\nattention, has therefore, shifted from homogeneous computing to specialized\nheterogeneous solutions. As an example, brain-inspired computing has re-emerged\nas a viable solution for many applications. Such new processors, however, have\nwidened the abstraction gamut from device level to applications. Therefore,\nefficient abstractions that can provide vertical design-flow tools for such\ntechnologies became critical. Photonics in general, and neuromorphic photonics\nin particular, are among the promising alternatives to electronics. While the\narsenal of device level toolbox for photonics, and high-level neural network\nplatforms are rapidly expanding, there has not been much work to bridge this\ngap. Here, we present a design methodology to mitigate this problem by\nextending high-level hardware-agnostic neural network design tools with\nfunctional and performance models of photonic components. In this paper we\ndetail this tool and methodology by using design examples and associated\nresults. We show that adopting this approach enables designers to efficiently\nnavigate the design space and devise hardware-aware systems with alternative\ntechnologies.\n