Benchmarking Machine-Learning Performance
Machine-learning attracts a lot of interest in recent years, more than ever before, as it utilizes the increasing amounts of data for many smart applications. This creates a flood of innovative ideas in this domain, constantly introducing new algorithmic and modeling techniques. The rapid pace of innovation poses a challenge in designing hardware that would best fit the needs of this domain several years from now. The goal of this work was to help coping with this challenge, by constructing a benchmark that represents the variety of fundamental compute requirements relevant to this field. The talk describes the analysis method used for building this benchmark. While no one can predict the future of this domain, interesting insights were deduced from its past, and helped hardware and software experts assess several possibilities for the future.