Comparative Analysis of Edge AIFrameworks on Resource-Limited Devices

Thesis Abstract

The rapid proliferation of edge computing and on-device intelligence has intensified the need to evaluate how diverse AI frameworks perform within resource-constrained environments, where compute power, memory, and energy budgets are severely limited. This study addresses the problem of inconsistent performance claims across edge AI frameworks, which hampers informed selection for real-time inference, model deployment, and privacy-preserving analytics on mobile and embedded devices. The aim is to conduct a rigorous comparative analysis of leading edge AI frameworks—TensorFlow Lite, PyTorch Mobile, ONNX Runtime, and MediaPipe—across a standardized set of resource-limited hardware platforms, to establish a comprehensive performance and usability benchmark. Specific objectives include (1) quantifying latency, throughput, energy consumption, and memory footprint for representative computer vision and sensor fusion workloads; (2) assessing model accuracy and robustness to quantization, pruning, and fixed-point arithmetic in constrained environments; (3) evaluating framework-level overheads, tooling maturity, and deployment practicality; (4) examining cross-framework portability and reproducibility of results; and (5) deriving guidelines for framework selection aligned with application requirements (e.g., real-time inference vs. battery life, offline operation vs. streaming data). The methodology adopts a mixed-methods, cross-sectional research design. The population comprises AI edge frameworks deployed on resource-limited devices, including Raspberry Pi 4B with 4GB RAM, NVIDIA Jetson Nano, and ARM Cortex-A53-based smartphones. A stratified sample of 40 devices is assembled, ensuring representation of CPU- and memory-constrained profiles. Data collection instruments include (i) standardized benchmark suites for computer vision and sensor data tasks (image classification, object detection, and pose estimation) implemented in each framework; (ii) a hardware energy measurement setup using high-precision power monitors; (iii) a reproducibility checklist and developer interviews to capture tooling maturity and deployment challenges. Quantitative data will be analyzed using ANOVA and multivariate regression to compare latency, frames per second, energy per inference, and memory usage across frameworks and devices, with post hoc tests to identify significant pairwise differences. Model accuracy will be evaluated with statistical significance tests under quantization and pruning scenarios. Qualitative data from developer interviews will be analyzed using thematic analysis to identify common friction points and usability factors, anchored by the Diffusion of Innovations theory and the Technology Acceptance Model to interpret adoption dynamics. The study will also apply a cross-framework portability metric to assess reproducibility of results under equivalent configurations, aligning with standards in reproducible research. Expected findings include (a) framework-specific trade-offs, with TensorFlow Lite and ONNX Runtime likely delivering lower latency on CPU-bound devices, while PyTorch Mobile may exhibit higher development simplicity but greater energy variance under heavy quantization; (b) energy efficiency patterns distinguishing static-graph vs. dynamic-graph execution models; (c) quantization and pruning effects on accuracy with diminishing returns beyond aggressive compression, influencing deployment decisions; (d) reproducibility challenges arising from divergent default settings, preprocessing pipelines, and operator support; and (e) a synthesis of practical guidelines informing framework selection by application domain, hardware profile, and desired lifecycle considerations. The study contributes to knowledge by providing a rigorous, empirical, cross-framework comparison applicable across diverse edge scenarios, clarifying how architectural choices influence performance under tight resource constraints, and offering a decision framework for practitioners and researchers. It advances understanding of how quantization, model optimization, and framework internals interact to shape accuracy-energy-performance envelopes, and it highlights gaps in operator coverage and tooling maturity that hinder portability. The conclusion will articulate concrete recommendations for selecting and optimizing edge AI frameworks for resource-limited devices, including a decision matrix and a checklist for reproducible benchmarking, alongside suggestions for future research into adaptive framework configurations and ecosystem enhancements to improve cross-platform interoperability.