FPGA-Based QBoost with Large-Scale Annealing Processor and Accelerated Hyperparameter Search
Takashi Takemoto, Normann Mertig, Masato Hayashi, Saki Susa-Tanaka, Hiroshi Teramoto, Atsuyoshi Nakamura, Ichigaku Takigawa, Shin-ichi Minato, Tamiki Komatsuzaki, Masanao Yamaoka · 2018
QBoost is a recently proposed machine learning algorithm, designed to exploit the benefits of emerging annealing processors which solve NP hard problems in combinatorial optimization a hundred times faster than conventional CPUs. In this paper, we present the first FPGA-based implementation of QBoost, incorporating a large-scale annealing processor with 2704 spins. In contrast to previous implementations, based on quantum annealers, we utilize the flexibility of FPGAs for implementing a fast, integrated QBoost engine which combines the annealing processor and the modules of the hyperparameter search on a single FPGA. As opposed to quantum annealers, this accelerates the time required for scanning the hyperparameter space from the order of hours to a single second.