Engineering Commons^™

Memristive Architectures And Algorithms For Approximate Graph-Based Inference, Mohammad M.A. Taha

Dissertations and Theses

The goal of this thesis is to design fast, low-power, robust graph-based inference systems. Our approach to this is by using (1) in-memory computing, (2) approximate computing principles, and (3) memristive devices.

This work is motivated by the fact that conventional von Neumann architectures are not efficient for inference applications, mainly due to the data transfer bottleneck. Adding cache memories and using GPUs is a remedy, however, does not eliminate this bottleneck. In-memory computing is an alternative approach that performs all computations inside the memory, thus eliminating the data transfer bottleneck.

The memristor, which is a passive two-terminal device, is …