Analyzing the Hardware-Software Implications of Multi-modal DNN Workloads using MMBench

The explosive growth of various types of big data and advances in AI technologies have catalyzed a new type of applications called multi-modal DNNs. Multi-modal DNNs are capable of interpreting and reasoning about information from multiple modalities, making them more applicable to real-world AI scenarios. In recent research, multi-modal DNNs have outperformed the best uni-modal DNN in a wide range of applications from traditional multimedia to emerging autonomous systems. However, despite their importance and superiority, very limited research attention has been devoted to understand the characteristics of multi-modal DNNs and their implications on current computing software/hardware platforms. To facilitate research and advance the understanding of these multi-modal DNN workloads, we first present MMbench, an open-source benchmark suite consisting of a set of real-world multi-modal DNN workloads with relevant performance metrics for evaluation. Then we use MMbench to conduct an in-depth analysis on the characteristics of multi-modal DNNs. We study their implications on application and programming framework, operating and scheduling system, as well as execution hardware. Finally, we conduct a case study and extend our benchmark to edge devices. We hope that our work can provide guidance for future software/hardware design and optimization to underpin multi-modal DNNs on both cloud and edge computing platforms.