핵심 개념
Optimizing pipeline cooperation for efficient CNN inference on diverse mobile devices.
초록
The article introduces PICO, a framework for accelerating CNN inference on mobile devices through pipeline cooperation. It addresses challenges in mapping CNN to devices and optimizing inference workload distribution. The framework features graph partition and many-to-many mapping algorithms, improving throughput by 1.8 ∼6.8× in experiments with Raspberry-Pi devices. Challenges include redundant calculations due to CNN properties and complex structures like DAGs.
- Introduction to the growth of mobile devices and the need for intelligent applications.
- Challenges of resource limitations and data transmission in mobile environments.
- Benefits of cooperative CNN inference on multiple devices.
- Issues with redundant calculations and complex CNN structures.
- Presentation of PICO framework with graph partition and mapping algorithms.
- Experiment results showing throughput improvement with Raspberry-Pi devices.
통계
1.8 ∼6.8× throughput improvement observed in experiments with Raspberry-Pi devices.