GENN: Enable Flexible and Efficient AI for Resource-Constrained Platforms
Published in 56th IEEE/ACM International Symposium on Microarchitecture, Student Research Competition (MICRO SRC), 2023
Research Statement: Deep learning (DL) is experiencing increased interest in resource-constrained devices. However, modern DL frameworks, e.g., TensorFlow and PyTorch, are designed and optimized for high-performance platforms. For usability, most frameworks use interpreted languages and require extensive libraries like Nvidia CUDA; executing them on resource-constrained devices remains challenging. Therefore, it is crucial to enable more memory and environment-friendly AI for platforms such as low-end Internet of Things (IoTs), simulators, and high-level synthesis (HLS). To this end, we develop GENN, an automatic PyTorch-to-C model conversion pipeline with a high degree of generality, flexibility, and usability.
Authors: Yan Zhu, Kaija Mikes, Karthik Ganesan, Natalie Enright Jerger
Download Paper | Download Slides