In the era of ubiquitous connectivity, edge computing scenarios such as autonomous driving and smart cities increasingly demand diverse computational capabilities, stringent real-time processing, and multidisciplinary integration. These requirements pose significant challenges to the adaptability and real-time performance of existing systems, which often fall short in efficiency.
Led by Professor Ye Xiaochun, a collaborative team at the Institute of Computing Technology, Chinese Academy of Sciences, has been advancing research on key technologies for reconfigurable edge computing systems, focusing on synergistic optimization of architecture, operators, and algorithms. To address the complex processing needs of edge scenarios, the team has developed a heterogeneous fusion dataflow architecture. This innovative model employs a decoupled dataflow method as a universal abstract execution model, facilitating efficient, data-driven computing across various domains. This architecture supports efficient integrated execution of algorithms in scenarios such as digital signal processing, artificial intelligence, and video processing, meeting high concurrency, strong real-time requirements, and reconfigurable computing needs.
Furthermore, for the optimization of core operators, the researchers proposed a two-level sparse computing filtering mechanism that filters out ineffective computations at multiple levels, such as instruction blocks and individual instructions, significantly enhancing neural network edge inference performance and overall energy efficiency. Lastly, the team has also optimized typical algorithms, including designing a new type of graph neural network to reduce computational demands on the edge, and proposing dynamic adjustment learning strategies to address the issue of uneven coverage in multisource data collected at the edge, thus balancing data processing and enhancing overall performance.
This research has been applied to edge intelligence processing scenarios like autonomous driving and smart cities, significantly enhancing processing efficiency and meeting the high demands for energy-efficient, real-time computing in edge environments.
- (a)Heterogeneous fusion dataflow architecture
- (b)Autonomous driving edge scenario
The work has been published in prestigious international journals and conferences, including ACM TACO 2024, IEEE TCSVT 2023, IEEE TPDS 2023, and AAAI 2023. The research has received support from the National Key R&D Program of China, the National Natural Science Foundation of China, and the International Partnership Program of the Chinese Academy of Sciences.
Papers are available at the following links:
https://dl.acm.org/doi/10.1145/3637906
https://ieeexplore.ieee.org/document/10364740
downloadFile