A Fully-Connected Boltzmann Machine with Virtual FPGAs
Behraz Vatankhahghadim · TSpace (University of Toronto) · 2018
This thesis introduces a 1024-node hardware for a Boltzmann machine, implemented in an Arria 10 field-programmable gate array (FPGA). This system consists of a network of connected nodes, the states of which undergo changes during a procedure termed annealing. Large memory resources required to save connectivity data are the main factor preventing the increase of the design's size and complexity. To alleviate this, an extension approach, whereby a problem is divided into sub-problems (virtual engines) which successively undergo annealing, is examined. Several parameters of this scheme are evaluated based on two metrics, leading to the conclusion that there is an optimal number of engine swaps for good performance, and that cyclically selecting engines improves the outcome. The extension method is demonstrated through a hardware prototype (with a personal computer as a controller) that can be a starting point for interface design and timing characterization.