The Truth About Edge AI Sustainability: A Three-Way Game Between Performance, Energy Consumption, and Privacy

This article, based on a real-device study of the Samsung Galaxy S25 Ultra, reveals key truths about edge AI in the three-way game between performance, energy consumption, and privacy: quantization techniques have negligible energy-saving effects; MoE architectures with 7B parameters achieve energy consumption levels comparable to 1-2B models; and 3B parameter models strike the optimal balance between quality and energy efficiency. It also discusses the practical constraints and future directions of edge AI.

Edge AI promises three major benefits: privacy protection (data stays local), offline availability, and low latency. However, it faces physical constraints of mobile devices: limited battery capacity, restricted heat dissipation, and tight memory (flagship phones only have 12-16GB RAM, which is shared). The core challenge is how to run AI models on resource-constrained devices.

The research team used a reproducible experimental pipeline to measure three key metrics on the Samsung Galaxy S25 Ultra (non-rooted, reflecting ordinary user scenarios): energy consumption (affects battery life), latency (affects user experience), and generation quality (output usefulness). It covers 8 mainstream edge models with parameters ranging from 0.5B to 9B. Methodological innovations include fine-grained measurements without rooting, a reproducible pipeline, and multi-model comparisons.

1. Quantization Paradox: While modern quantization techniques reduce memory usage, they offer almost no additional energy-saving benefits (since energy consumption on mobile devices mainly comes from memory access rather than computation); 2. MoE Architecture Miracle: A model with a total of 7B parameters only activates 1-2B parameters during inference, resulting in energy consumption close to small models but with the advantages of large capacity; 3. Medium-Sized Model Advantage: 3B parameter models (e.g., Qwen2.5-3B) achieve the optimal balance between quality, energy consumption, latency, and memory. Small models lack quality, while large models have high energy consumption and diminishing marginal returns.

Edge processing keeps data on the device, reduces leakage risks, and gives users control over their data. Privacy and energy consumption are synergistic in some scenarios: avoiding network transmission saves energy, and local caching reduces repeated computations. However, edge computing may increase processor energy consumption. For medium-complexity tasks, the total energy consumption of edge computing may be lower than that of cloud computing, and privacy is better.

• Model developers: Emphasize architectural innovation (e.g., MoE), optimize energy consumption rather than just speed, and focus on medium-sized models (2B-4B parameters);
• Device manufacturers: Optimize hardware-software collaboration, prioritize improving memory bandwidth, and promote energy efficiency ratios;
• Application developers: Choose appropriate model sizes (3B is sufficient for most scenarios), prioritize MoE architectures, and balance quality and battery life.

Limitations: Tested only on the Samsung Galaxy S25 Ultra (a top flagship), with no exploration of mid-to-low-end device characteristics; focused on text generation, with multi-modal tasks yet to be studied; used fixed test sets, with no coverage of dynamic workloads. Future directions: Cross-device validation, multi-modal expansion, adaptive strategies (dynamic model adjustment), and exploration of more efficient architectures.