Exploring Smartphone-Based Edge AI Inferences Using Real Testbeds
I tiakina i:
| I whakaputaina i: | Sensors vol. 25, no. 9 (2025), p. 2875 |
|---|---|
| Kaituhi matua: | |
| Ētahi atu kaituhi: | , |
| I whakaputaina: |
MDPI AG
|
| Ngā marau: | |
| Urunga tuihono: | Citation/Abstract Full Text + Graphics Full Text - PDF |
| Ngā Tūtohu: |
Kāore He Tūtohu, Me noho koe te mea tuatahi ki te tūtohu i tēnei pūkete!
|
| Whakarāpopotonga: | The increasing availability of lightweight pre-trained models and AI execution frameworks is causing edge AI to become ubiquitous. Particularly, deep learning (DL) models are being used in computer vision (CV) for performing object recognition and image classification tasks in various application domains requiring prompt inferences. Regarding edge AI task execution platforms, some approaches show a strong dependency on cloud resources to complement the computing power offered by local nodes. Other approaches distribute workload horizontally, i.e., by harnessing the power of nearby edge nodes. Many of these efforts experiment with real settings comprising SBC (Single-Board Computer)-like edge nodes only, but few of these consider nomadic hardware such as smartphones. Given the huge popularity of smartphones worldwide and the unlimited scenarios where smartphone clusters could be exploited for providing computing power, this paper sheds some light in answering the following question: Is smartphone-based edge AI a competitive approach for real-time CV inferences? To empirically answer this, we use three pre-trained DL models and eight heterogeneous edge nodes including five low/mid-end smartphones and three SBCs, and compare the performance achieved using workloads from three image stream processing scenarios. Experiments were run with the help of a toolset designed for reproducing battery-driven edge computing tests. We compared latency and energy efficiency achieved by using either several smartphone clusters testbeds or SBCs only. Additionally, for battery-driven settings, we include metrics to measure how workload execution impacts smartphone battery levels. As per the computing capability shown in our experiments, we conclude that edge AI based on smartphone clusters can help in providing valuable resources to contribute to the expansion of edge AI in application scenarios requiring real-time performance. |
|---|---|
| ISSN: | 1424-8220 |
| DOI: | 10.3390/s25092875 |
| Puna: | Health & Medical Collection |