Scalable Synthesis of 3-D Crossbars for Flow-based Computing

Faced with the demise of Moore's law, the field of computer architecture is currently undergoing rapid evolution. Emerging devices, like memristors, play a crucial role in this development. Computer design is a computationally challenging task, but the tools that currently exist for designing memristor-based computing units produce designs that are slow, or power-hungry, or both. In this paper, we propose a design automation techniques which are uniquely geared towards a low-latency, low-power variant of memristor-based computation. We develop logical operations between crossbars which exploit sneak paths as a computational mechanism, and leverage the compositional power of and-inverter graphs for scalable synthesis of crossbar circuits which utilize these fundamental operations as building blocks. In order to take advantage of the latest developments in fabrication, we devise a method to build 3-D crossbars while maintaining functionality. Designs produced by our method outperform state-of-the-art designs in the average case, and by several orders of magnitude in the best cases.