Understanding Large Language Models in Your Pockets: Performance Study on COTS Mobile Devices

TL;DR

Privacy concerns motivate on-device deployment of LLMs, but performance on real mobile hardware remains underexplored. The authors conduct a comprehensive multi-vendor measurement study across seven devices and six models, evaluating user-centric metrics (throughput, latency, response quality) and developer-centric factors (resource utilization, DVFS, thermal throttling, battery) using CPU and GPU inference frameworks. They reveal substantial hardware- and software-driven bottlenecks, show how quantization and architecture interact with performance, and demonstrate the potential of hybrid CPU-GPU strategies to improve efficiency. The findings provide actionable guidance for optimizing on-device LLMs and inform the design of future mobile system architectures toward higher performance and energy efficiency.

Abstract

As large language models (LLMs) increasingly integrate into every aspect of our work and daily lives, there are growing concerns about user privacy, which push the trend toward local deployment of these models. There are a number of lightweight LLMs (e.g., Gemini Nano, LLAMA2 7B) that can run locally on smartphones, providing users with greater control over their personal data. As a rapidly emerging application, we are concerned about their performance on commercial-off-the-shelf mobile devices. To fully understand the current landscape of LLM deployment on mobile platforms, we conduct a comprehensive measurement study on mobile devices. While user experience is the primary concern for end-users, developers focus more on the underlying implementations. Therefore, we evaluate both user-centric metrics-such as token throughput, latency, and response quality-and developer-critical factors, including resource utilization, OS strategies, battery consumption, and launch time. We also provide comprehensive comparisons across the mobile system-on-chips (SoCs) from major vendors, highlighting their performance differences in handling LLM workloads, which may help developers identify and address bottlenecks for mobile LLM applications. We hope that this study can provide insights for both the development of on-device LLMs and the design for future mobile system architecture.

Figures (18)

• Figure 1: Measurement workflow.
• Figure 2: Inference performance on CPU (llama.cpp)
• Figure 3: Inference performance (MLC LLM) on GPU with 64-token prompt
• Figure 4: Performance of Llama-2-7B with Background Task (Music App) on Xiaomi14 Pro.
• Figure 5: Performance of Llama-2-7B with AI Task (YOLOv11n) on Xiaomi14 Pro.