Characterizing the Machine Learning Capabilities of General-Purpose CPUs

Abstract

Machine learning is a ever-growing field, with many people interested in trying it out on their own. While there is specialized hardware to run machine learning algorithms, it costs a considerable amount to invest into even the low-end devices. Those who are merely curious will find themselves using what hardware they have - their laptops, desktops, and tablets - to run machine learning applications. There exist test benches and published results for most processors, but machine learning operations are not part of the widespread tests. There also exist machine learning test benches, but they are specialized for more powerful hardware. This thesis finds the intersection of the two, characterizing the machine learning capabilities of general-purpose CPU's using DeepBench, Baidu's machine learning test bench, and VTune Amplifier, Intel's hardware analysis tool, in order to inform people what they should expect when running machine learning algorithms on their general-purpose devices. We find that of the two main machine learning operations, matrix multiply and convolution, matrix multiplication performs well on most machines, while convolution has mixed results.