OpBench: an operator-level GPU benchmark for deep learning

Abstract

Operators (such as Conv and ReLU) play an important role in deep neural networks. Every neural network is composed of a series of differentiable operators. However, existing AI benchmarks mainly focus on accessing model training and inference performance of deep learning systems on specific models. To help GPU hardware find computing bottlenecks and intuitively evaluate GPU performance on specific deep learning tasks, this paper focuses on evaluating GPU performance at the operator level. We statistically analyze the information of operators on 12 representative deep learning models from six prominent AI tasks and provide an operator dataset to show the different importance of various types of operators in different networks. An operator-level benchmark, OpBench, is proposed on the basis of this dataset, allowing users to choose from a given range of models and set the input sizes according to their demands. This benchmark offers a detailed operator-level performance report for AI and hardware developers. We also evaluate four GPU models on OpBench and find that their performances differ on various types of operators and are not fully consistent with the performance metric FLOPS (floating point operations per second).